JP5542732B2 - Data extraction apparatus, data extraction method, and program thereof - Google Patents

Description

  The present invention relates to a technique for extracting data from a set of text data, and more particularly to a technique for extracting a character string having a similar relationship from a set of text data using a character string having a relationship with a specific content as an input.

  Currently, various natural language processing technologies are being researched and developed, and there are many methods for extracting necessary information from a vast knowledge source such as WEB. One of such methods is to input a character string (for example, <Hiroshima>, <Hanshin>, etc.) that has some relationship with specific contents, and have a similar relationship from a large amount of text data (for example, document data). Some collect character strings (eg <Yakult>). Such a method is called “set expansion”. A character string handled in set expansion is called an “entity”, an extraction target entity is called a “positive example entity”, and an entity that is not extracted (not an extraction target) is called a “negative example entity”. Furthermore, the entity that is input first in the set expansion is called “seed entity”, the positive seed entity is called “positive seed entity”, and the negative seed entity is called “negative seed entity”.

Hereinafter, an example of a conventional set expansion will be outlined.
Step I: Using text data containing positive example entities (eg <Hiroshima>, <Hanshin>, etc.) The feature of the negative example entity is extracted (featured). The initial value of the positive example entity is a positive example seed entity, and the initial value of the negative example entity is a negative example seed entity.
Step II: Using the identity of the positive example entity and the identity of the negative example entity obtained in Step I as learning data, an identification model for identifying whether any entity is a positive example entity or a negative example entity Generate.
Step III: An entity before identification (for example, <Yakult>) and its features are extracted from text data, and the unknown entity is identified using the identification model obtained in Step II.
Step IV: Add the features of the entities that are identified as positive example entities with high reliability and the features of the entities that are estimated to be negative example entities with low reliability to the learning data.
Step V: Determine whether or not the convergence condition is satisfied. If not, return to Step I and repeat the process. If the convergence condition is satisfied, the process is terminated. A framework for iterative learning in which identification is performed based on a once learned model and used as new learning data is called a bootstrap method.

ここで|e, p|はエンティティeとそれと共起するパターンpとの組のクエリログ中での出現頻度を表す。また、*はp又はeのワイルドカードを表す。すなわち、|e, *|はエンティティeと何れかのパターン*との組のクエリログ中での出現頻度を表し、|*, p|は何れかのエンティティ*とパターンpとの組のクエリログ中での出現頻度を表す。
また、このエンティティeに対し、エンティティ信頼度r_Eとパターン信頼度r_Pを計算する。r_E, r_Pは以下で定義される。

ここで|E|及び|P|はそれぞれエンティティe及びパターンpの総数を表す。また、max_e pmiはエンティティをeに固定した場合のPMIの最大値を表し、max_ppmiはパターンをpに固定した場合のPMIの最大値を表す。
ステップC：エンティティ信頼度r_Eに基づきエンティティｅを新たな正例エンティティとするか否かを判定する。
ステップD：必要な数の正例エンティティが得られていない場合にはステップBに戻って処理を繰り返す。必要な数の正例エンティティが得られた場合には処理を終了する。 Next, TChai which is another example of set expansion (for example, see Non-Patent Document 1) will be outlined. In TChai, a search query log (hereinafter referred to as query log) is used as a resource. A query log is a set of queries (keywords consisting of several words) from a user used for keyword search.
Step A: The pattern p, which is a word that co-occurs with the positive seed entity, is extracted from the query log, and is used as a feature of the positive seed entity. This step is only performed once.
Step B: PMI (Pointwise Mutual Information) in two terms of the entity e unknown to be a positive example entity and the pattern p co-occurring with it is calculated.

Here, | e, p | represents the appearance frequency in the query log of a set of the entity e and the pattern p co-occurring with it. * Represents a p or e wildcard. That is, | e, * | represents the frequency of occurrence in the query log of the pair of entity e and any pattern *, and | *, p | is the query log of the pair of any entity * and pattern p. Represents the appearance frequency of
In addition, an entity reliability r _E and a pattern reliability r _P are calculated for this entity e. r _E and r _P are defined below.

Here, | E | and | P | represent the total number of entities e and patterns p, respectively. Max _e pmi represents the maximum value of PMI when the entity is fixed to e, and max _p pmi represents the maximum value of PMI when the pattern is fixed to p.
Step C: It is determined whether or not the entity e is a new positive entity based on the entity reliability r _E.
Step D: If the required number of positive entity is not obtained, return to Step B and repeat the process. If the required number of positive entity is obtained, the process is terminated.

Mamoru Komachi, Kumi Suzuki, “Improvement of Semi-Supervised Semantic Knowledge Acquisition from Search Logs”, Transactions of the Japanese Society for Artificial Intelligence, Vol. 23, No. 3, 2008, p. 217-225

The conventional set expansion has a problem of semantic drift.
For example, suppose that the positive entity <Yakult> can be acquired by the conventional set expansion for the positive seed entity <Hiroshima><Hanshin> representing the team name. <Yakult> is also a beverage name, so by adding <Yakult> to a new example entity, a beverage-type entity such as <Cola> will be acquired as a normal entity in the next iteration. There is a possibility that the topic of positive example entities will shift. The phenomenon that the topic of positive entity acquired in this way shifts is called semantic drift.
In TChai, in order to suppress the influence of semantic drift, the above-mentioned reliability is used, and an algorithm that does not select general entities and patterns that tend to appear in any query in common is used. However, semantic drift can still occur in TChai, so it is desirable to reduce semantic drift from another perspective.

  The present invention has been made in view of such a point, and an object thereof is to provide a technique capable of reducing the semantic drift.

In the first aspect of the present invention, unsupervised learning data obtained from text data is a topic model that describes the relationship between topic information representing the appropriateness of a plurality of topic candidates for text data as an index value and the text data. The example topic information extracted from the topic model corresponding to the topic of the text data including the example entity that is the character string to be extracted is used as at least part of the features of the example entity, The negative example topic information extracted from the topic model corresponding to the topic of the text data containing the negative example entity that is a non-character string is set as at least part of the negative example entity feature, and the positive example entity feature and the negative example entity feature By using the learning process with supervised learning data, An entity that is a character string included in text data selected from a set of text data by generating an identification model that is a function that outputs information for identifying whether the entity is a positive example entity or negative example entity Is the target entity, the topic information extracted from the topic model corresponding to the topic of the selected text data is used as at least a part of the feature of the target entity, and the feature of the target entity is input to the identification model. Identify positive entity or negative entity, identify target entity as positive entity when target entity is identified as positive entity, and negative target entity when target entity is identified as negative entity Example entity To.

  In the second aspect of the present invention, the first positive example entity selected from the set of positive example entities that are the character strings to be extracted and the first selected from the set of positive example attributes that are character strings representing the attributes of the positive example entities. Characters representing the attributes of the first negative example entity and the negative example entity selected from the set of the first positive example entity-positive example attribute pair that is a pair with the positive example attribute and the negative example entity that is a character string that is not to be extracted. A first negative example entity-negative example attribute pair that is a set with a first negative example attribute selected from a set of negative example attributes that are columns is generated, and the first positive example entity and the first are generated from the set of text data. A character string including a pair with a positive example attribute is selected, and information indicating the characteristics of the first positive example entity-positive example attribute pair for the selected character string is used as the feature of the first positive example entity-positive example attribute pair. Less From the set of text data, a character string including a pair of the first negative example entity and the first negative example attribute is selected, and the first negative example entity-negative example attribute pair for the selected character string is selected. Information representing the characteristics is at least part of the features of the first negative example entity-negative example attribute pair, and the features of the first positive example entity-positive example attribute pair and the features of the first negative example entity-negative example attribute pair Through the learning process using supervised learning data as an input, the entity-attribute pair is a positive example entity-positive example attribute by inputting the identity of the entity-attribute pair that is a set of an entity that is an arbitrary character string and the attribute of the entity. A first identification model, which is a function for outputting information for identifying a pair or a negative example entity-negative example attribute pair, is generated, and is selected from a set of text data. Select text data, select a character string included in the selected text data as a first target entity, select a character string different from the first target entity from the selected text data as a first target attribute, A set of the target entity and the first target attribute is defined as a first target entity-target attribute pair, and information representing the characteristics of the first target entity-target attribute pair in the selected text data is the first target entity-target. At least part of the feature of the attribute pair, the feature of the first target entity-target attribute pair is input to the first identification model, and the first target entity-target attribute pair is a positive entity-positive attribute pair or negative Identify whether it is an example entity-negative example attribute pair, and the first target entity-target attribute pair is a positive example entity -When it is identified as a positive example attribute pair, the first target attribute is added to the set of positive example attributes, and the first target entity-target attribute pair is identified as a negative example entity-negative example attribute pair. The first target attribute is added to the set of negative example attributes, and the second positive example entity selected from the set of positive example entities and the second positive example attribute selected from the set of positive example attributes A second negative example entity-negative that is a set of two positive example entity-positive example attribute pairs, a second negative example entity selected from the set of negative example entities, and a second negative example attribute selected from the set of negative example attributes An example attribute pair is generated, a character string including a pair of a second positive example entity and a second positive example attribute is selected from a set of text data, and a second positive example entity-positive example for the selected character string is selected. Information representing the characteristics of attribute pairs Is selected as a character string that includes a pair of the second negative example entity and the second negative example attribute from the set of text data, and at least part of the feature of the second positive example entity-positive example attribute pair. Information representing the characteristics of the second negative example entity-negative example attribute pair for the character string is at least part of the features of the second negative example entity-negative example attribute pair, and the second positive example entity-positive example attribute pair The feature of the entity-attribute pair that is a set of an entity that is an arbitrary character string and the attribute of the entity is obtained by learning processing using the feature and the feature of the second negative example entity-negative example attribute pair as supervised learning data. A function that outputs information for identifying whether the entity-attribute pair is a positive entity-positive example attribute pair or a negative example entity-negative example attribute pair. A second identification model is generated, any text data is selected from a set of text data, a character string included in the selected text data is selected as a second target entity, and the second target is selected from the selected text data. A character string different from the entity is selected as a second target attribute, and a set of the second target entity and the second target attribute is set as a second target entity-target attribute pair, and the second target entity in the selected text data- Information representing the characteristics of the target attribute pair is set as at least a part of the feature of the second target entity-target attribute pair, and the feature of the second target entity-target attribute pair is input to the second identification model, and the second target Identifies whether the entity-target attribute pair is a positive entity-positive example attribute pair or a negative example entity-negative example attribute pair; When the second target entity-target attribute pair is identified as a positive entity-positive attribute pair, the second entity is added to the set of positive entities, and the second target entity-target attribute pair is negative. If it is identified as an example entity-negative example attribute pair, the second target entity is added to the set of negative example entities.

  As described above, in the present invention, since at least one of topic information and attributes is reflected in entity identification, semantic drift can be reduced.

FIG. 1 is a block diagram for illustrating a functional configuration of the data extraction device according to the first embodiment. 2A and 2B are block diagrams for illustrating the functional configuration of the automatic generation unit. FIG. 3 is a diagram for illustrating a data extraction process of the data extraction apparatus according to the first embodiment. FIG. 4 is a diagram illustrating a set D of text data stored in the storage unit. FIG. 5A is a diagram illustrating a set D ′ of text data with topic information. FIG. 5B is a diagram illustrating a pair (fP _e ^j , <+1>) and a pair (fN _e ^j , <-1>) output by the topic information extraction unit. FIG. 6 is a block diagram for illustrating a functional configuration of the data extraction apparatus according to the second embodiment. FIG. 7 is a diagram for illustrating the data extraction processing of the data extraction device of the second embodiment. FIG. 8A is a diagram illustrating a pair (fP _a ^j , <+1>) and a pair (fN _a ^j , <-1>) output by the attribute identifying feature extraction unit. FIG. 8B is a diagram illustrating a pair (fP _e ^j , <+1>) and a pair (fN _e ^j , <−1>) output by the entity identifying feature extraction unit. FIG. 9 is a block diagram for illustrating a functional configuration of the data extraction device 3 of the third embodiment. FIG. 10 is a diagram for illustrating data extraction processing of the data extraction device 3 according to the third embodiment.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
[First Embodiment]
<Configuration>
FIG. 1 is a block diagram for illustrating a functional configuration of a data extraction apparatus 1 according to the first embodiment.
As illustrated in FIG. 1, the data extraction device 1 includes a storage unit 11 a-11 e, a topic assignment unit 12, a feature extraction unit 13, a topic information extraction unit 14, an identification learning unit 15, an entity identification unit 16, and a convergence determination unit 17. , And an output unit 18 and a control unit 19, and execute each process under the control of the control unit 19. The data extraction apparatus 1 is configured by reading a special program into a known or dedicated computer including, for example, a CPU (central processing unit), a RAM (random-access memory), a ROM (read-only memory), and the like. It is a special device. For example, the storage unit 11a-11e is a hard disk, a semiconductor memory, or the like, and includes a topic assignment unit 12, a feature extraction unit 13, a topic information extraction unit 14, an identification learning unit 15, an entity identification unit 16, a convergence determination unit 17, and an output unit. 18 and the control unit 19 are a CPU in which a special program is read. Further, at least a part of these may be configured by an integrated circuit or the like. In addition, although the arrows shown in FIG. 1 indicate the flow of information, some arrows are omitted for the sake of description (the same applies to other block diagrams described below).

<Pre-processing>
As pre-processing, a set D of text data stored in the storage unit 11a, topic models TM ⁰ is stored in the storage unit 11b.
Text data means data including character text. Examples of text data are document data, queries, chart data including phrases, phrase data, word string data, and the like. In this embodiment, an example is shown in which document data after preprocessing such as morphological analysis, specific expression extraction, dependency analysis, sentence boundary identification, and the like is used as text data.

“Topic model TM ⁰ ” means a model (function, formula) describing the relationship between topic information corresponding to a topic corresponding to text data and a character string included in the text data. Specific examples of the “character string” include a word, a word string, a phrase, a sentence, a character, and a symbol. The topic corresponding to the text data means a topic (text, topic, matter, event, topic, classification, etc.) of the text data. The text data does not always include the word representing the topic itself. The topic information may be any information as long as it corresponds to the topic corresponding to the text data. For example, for each topic candidate corresponding to the text data (for example, <Team name> or <Company name>), an index (for example, probability, weight coefficient, probability, A score which is a function value of the weighting coefficient) is given, and at least a part of the indices may be the topic information of the text data.

ここでdはテキストデータの集合Dに属するテキストデータd∈Dを表し、p(d)はテキストデータの集合Dにおけるテキストデータdの出現確率を表す。z∈Zは隠れ変数であり、各zが１つのトピックの候補に対応する。Zは隠れ変数zの集合を表す。以下ではｚを1以上Z以下の自然数とし、Zを隠れ変数の総数（トピックの候補の総数）とする。p(z)は隠れ変数zに対する確率であり、

を満たす。vは文字列を表し、Vは文字列vの集合を表す。p(v|z)は隠れ変数zにおける文字列vの生成確率（隠れ変数zが与えられたときの文字列vの事後確率）であり、

を満たす。n_dvはテキストデータd中に文字列vが出現した回数である。 The topic model is acquired in advance from unsupervised learning data (learning data obtained from text data whose relationship with topic information is not specified). For example, when it is desired to acquire a desired entity from 1 million document data on the web, a topic model is learned using learning data obtained from these 1 million document data.
Specific examples of Topic Model TM ⁰ are UM (Unigram Mixtures) (Andrew K. McCallum, Kamal Nigam, "Employing EM and Pool-Based Active Learning for Text Classification", ICML'98, 1998, etc.), LDA (Latent Dirichlet Allocation) and DM (Dirichlet Mixtures). An example of using the UM as topic models TM ⁰ below.
Topic model TM ⁰ in this case is defined by the following form.

Here, d represents text data dεD belonging to the text data set D, and p (d) represents the appearance probability of the text data d in the text data set D. z∈Z is a hidden variable, and each z corresponds to one candidate topic. Z represents a set of hidden variables z. In the following, z is a natural number between 1 and Z, and Z is the total number of hidden variables (total number of topic candidates). p (z) is the probability for the hidden variable z

Meet. v represents a character string, and V represents a set of character strings v. p (v | z) is the generation probability of the character string v in the hidden variable z (the posterior probability of the character string v when the hidden variable z is given),

Meet. n _dv is the number of times the character string v appears in the text data d.

Learning topic model TM ⁰ is performed using the EM algorithm is a type of iterative optimization techniques, parameters by learning p (z), p (v | z) is obtained. Each resulting parameters p (z), p (v | z) is stored in the storage unit 11b as the information for identifying the topic model TM ^0. This is equivalent to the topic model TM ⁰ is stored in the storage unit 11b.

  In this embodiment, a character string included in text data is referred to as an “entity”, an extraction target entity is referred to as a “positive example entity”, and an entity that is not extracted (not extracted) is referred to as a “negative example entity”. Also, the first input entity is called “seed entity”, the positive seed entity is called “positive seed entity”, and the negative seed entity is called “negative seed entity”.

<Data extraction process>
FIG. 3 is a diagram for illustrating data extraction processing of the data extraction device 1 of the first embodiment. Hereinafter, the data extraction process of the data extraction apparatus 1 will be exemplified with reference to FIG.
<< Initialization: Step S11 >>
The control unit 19 initializes the value of j to j = 1.
<< Topic Assignment: Step S12 >>
Topics imparting unit 12, using the topic models TM ⁰ stored in the storage unit 11b, and the topic information corresponding to the topic of the text data set D of text data stored in the storage unit 11a includes generating respectively. The topic assigning unit 12 associates each generated topic information with each corresponding text data, and generates text data with topic information including text data and topic information. The generated set D ′ of text data with topic information is stored in the storage unit 11c. Any information may be used as topic information as long as it corresponds to the topic of each text data. The following illustrates the topic information generated a UM as topic models TM ^0.

さらに、確率の乗法定理より、トピック付与部１２は、p(z,d)及びp(z)を用い、隠れ変数zが与えられたときのテキストデータdの事後確率p(d|z)を、以下のように求めることができる。
p(d|z)=p(z,d)/p(z) …(7)
またさらに、ベイズの定理より、トピック付与部１２は、得られたp(d), p(d|z)及びp(z)を用い、テキストデータdが与えられたときの隠れ変数zの事後確率p(z|d)を以下のように求めることができる。
p(z|d)=p(d|z)p(z)/p(d) …(8) [Example of topic information]
The topic assignment unit 12 uses the parameters p (z) and p (v | z) of the topic model TM ⁰ stored in the storage unit 11b, n _dv obtained from the text data d and the character string v, and uses the equation (4). Accordingly, p (d) corresponding to the text data d belonging to the text data set D stored in the storage unit 11a can be calculated. Also, from the probability multiplication theorem, the topic assigning unit 12 can obtain the joint probability p (z, v) for z and v using p (z) and p (v | z) as follows. .
p (z, v) = p (z) p (v | z)… (5)
Also, the topic assigning unit 12 can obtain the joint probability p (z, d) for z and d using p (z, v) and n _dv as follows.

Further, according to the probability multiplication theorem, the topic assignment unit 12 uses p (z, d) and p (z), and calculates the posterior probability p (d | z) of the text data d when the hidden variable z is given. The following can be obtained.
p (d | z) = p (z, d) / p (z) (7)
Furthermore, according to Bayes' theorem, the topic assigning unit 12 uses the obtained p (d), p (d | z), and p (z), and the posterior of the hidden variable z when the text data d is given. The probability p (z | d) can be obtained as follows.
p (z | d) = p (d | z) p (z) / p (d)… (8)

That is, the topic assigning unit 12 uses the parameters p (z) and p (v | z) of the topic model TM ⁰ stored in the storage unit 11b, and uses the posterior probability p (z of the hidden variable z for any text data d. | d) can be calculated. Note that the procedure for calculating the posterior probability p (z | d) is not limited to the above. As long as p (z | d) is finally obtained, the posterior probability p (z | d) may be calculated by any calculation procedure.

The posterior probability p (z | d) is an index representing the appropriateness of the topic candidates, and these can be used as topic information. Hereinafter, such topic information will be exemplified.
Topic information example 1: Among the posterior probabilities p (z _n | d) (n = 1,..., Z) corresponding to the hidden variables z _n corresponding to the candidates for each topic, the top having the largest posterior probability The topic information of the text data d is a set of N hidden variables z _{n ′} and corresponding posterior probabilities p (z _{n ′} | d) or mappings of the posterior probabilities p (z _{n ′} | d). . N is a natural number constant between 1 and Z. For example, if N = 1 and p (z ₅ | d) = 0.95 is the maximum posterior probability, the set of hidden variable z ₅ and posterior probability p (z ₅ | d) = 0.95 Use topic information.

Topic information example 2: Among the posterior probabilities p (z _n | d) (n = 1,..., Z) corresponding to the hidden variables z _n corresponding to the candidates for each topic, the top with the largest posterior probability Let N hidden variables z _{n ′} or a map of the hidden variables z _{n ′ be} topic information of the text data d. For example, if N = 1 and p (z ₅ | d) = 0.95 is the maximum posterior probability, the hidden variable z ₅ is set as the topic information of the text data d.

Topic information example 3: Top N posteriors of each posterior probability p (z _n | d) (n = 1, ..., Z) corresponding to each hidden variable z _n corresponding to each topic candidate A mapping of probability p (z _{n '} | d) or posterior probability p (z _n' | d) to each element of the n'th dimension and a Z-dimensional vector with the other ZN elements to 0 as text data d Topic information. For example, if N = 1 and p (z ₂ | d) = 0.95 is the maximum posterior probability, the Z-dimensional vector (0, 0.95, 0, ..., 0) is used as the topic information of the text data d. .

Topic information example 4: The top N posteriors of the posterior probabilities p (z _n | d) (n = 1,..., Z) corresponding to the hidden variables z _n corresponding to the candidate candidates. A Z-dimensional vector with the n'th element corresponding to each probability p (z _{n '} | d) as the first constant (eg 1) and the other ZN elements as the second constant (eg 0) as text The topic information of data d. For example, when N = 1 and p (z ₂ | d) = 0.95 is the maximum posterior probability, the Z-dimensional vector (0, 1, 0, ..., 0) is used as topic information of the text data d. .

Topic information example 5: In the topic information example 1 or 2, “the top N hidden variables z _{n ′} having a large posterior probability” are replaced with “hidden variables z _{n ′} having a posterior probability equal to or greater than a threshold”. Define topic information for text data d.

Examples of topical information 6: In Example 3 or 4 topic information, "top N posterior probability p (z _n was replaced with _'| | d)" and "(z _n threshold or more posterior probability _p' d)" The topic information of the text data d is determined by the method.

FIG. 4 is a diagram illustrating a set D of text data stored in the storage unit 11a, and FIG. 5A is a diagram illustrating a set D ′ of text data with topic information. The text data set D illustrated in FIG. 4 is data in which text data, which is document data after preprocessing, is associated with an ID of the text data. Further, in the set D ′ of text data with topic information illustrated in FIG. 5A, the text data, the ID of the text data, and the topic information generated by the topic information example 1 are associated with the text data. Data. Thus, topic information is assigned to each text data, and the same topic information corresponds to entities appearing in the same text data.
If topic information is previously assigned to text data, the topic information may be used. Further, when the set D ′ of text data with topic information is generated in advance, the process of step S12 may not be executed (end of description of [example of topic information]).

<< Feature Extraction: Step S13 >>
An example of an entity that the user desires is input to the feature extraction unit 13 as a positive example seed entity RP _e ⁰ . For example, <Hiroshima> is entered as a positive seed entity. Further, the negative example seed entity RN _e ⁰ is input to the feature extraction unit 13. For example, <Japan> etc. is entered as a negative example seed entity. The positive example seed entity RP _e ⁰ is a positive example entity in the first process (j = 1), and the negative example seed entity RN _e ⁰ is a negative example entity in the first process (j = 1).

The positive seed entity RP _e ⁰ has been selected by the user. The negative example seed entity RN _e ⁰ may be selected by the user, or may be generated semi-automatically or fully automatically from the text data set D. A method for generating the negative example seed entity RN _e ⁰ semi-automatically or fully automatically will be exemplified below.

[Example of semi-automatic generation of negative example seed entity RN _e ⁰ ]
A negative example seed entity generation unit (not shown) extracts a predetermined number of text data not including any positive example seed entity RP _e ⁰ from the set D of text data, and randomly selects one from each extracted text data Select nouns and output them as negative entity candidates. A display unit (not shown) displays these negative example entity candidates, and performs a display prompting the user to select a negative example seed entity from them. The selection content by the user is input to the negative example seed entity generation unit, and the negative example seed entity generation unit outputs the selected negative example entity candidate as the negative example seed entity RN _e ⁰ ([negative example seed entity RN _e ⁰ Example of semi-automatic generation method]

[Example of automatic generation method of negative example seed entity RN _e ⁰ ]
<Method 1>
The data extraction apparatus 1 of this example includes an automatic generation unit 110 that automatically generates a negative example seed entity RN _e ⁰ (FIG. 1). The automatic generation unit 110 includes a positive example distribution processing unit 111, a negative example topic determination unit 112, and a negative example seed entity generation unit 113 (FIG. 2A).

ここでdはPDに含まれるある１テキストデータを示し、P(PD)はテキストデータの集合PDの出現確率を表す。vは単語などの文字列を表し、Vは文字列vの集合を表す。p(v)は文字列vの生成確率であり、以下の関係を満たす。

ここでn_dvはテキストデータd中に文字列vが出現した回数である。この生成確率p(v)（正例確率分布を表すパラメータ）は最尤推定法を用いて容易に求めることができる。具体的には、生成確率p(v)は以下の式によって計算され得る。

ここでn_PDvはテキストデータの集合PD中に文字列vが出現した回数を表す。N_PDはテキストデータの集合PDに含まれる文字列の総数（例えば、総単語数）を表し、N_PD＝Σ_v n_PDvである。これらの値は、記憶部１１ａに格納されたテキストデータの集合Dから得ることができる。
この例のデータ抽出装置１は、各文字列v=v₁,...,v_V∈Vに対応する各生成確率p(v₁),..., p(v_V)を正例確率分布のパラメータβ_p={p(v₁),..., p(v_V)}として出力する。 First, the positive example distribution processing unit 111 is the appearance probability distribution of all entities included in the text data set PD including the positive example seed entity RP _e ⁰ among the text data set D stored in the storage unit 11a. Information (parameters) representing the positive probability distribution is obtained. A typical example of a positive probability distribution is a multinomial distribution of all entities according to the Bag-of-Words assumption. The following shows an example in which a character string such as a word is an entity, and the multinomial distribution of all entities according to the Bag-of-Words assumption is a positive probability distribution.
The parameter representing the positive example probability distribution in this example is the generation probability p (v) of the character string v that is an entity. The generation probability p (v) satisfies the following relationship.

Here, d indicates one text data included in the PD, and P (PD) indicates the appearance probability of the text data set PD. v represents a character string such as a word, and V represents a set of character strings v. p (v) is the generation probability of the character string v and satisfies the following relationship.

Here, n _dv is the number of times the character string v appears in the text data d. The generation probability p (v) (a parameter representing a positive example probability distribution) can be easily obtained using the maximum likelihood estimation method. Specifically, the generation probability p (v) can be calculated by the following equation.

Here, n _PDv represents the number of occurrences of the character string v in the text data set PD. N _PD represents the total number of character strings (for example, the total number of words) included in the text data set PD, and N _PD = Σ _v n _PDv . These values can be obtained from a set D of text data stored in the storage unit 11a.
The data extraction apparatus 1 in this example uses the generation probabilities p (v ₁ ), ..., p (v _V ) corresponding to the character strings v = v ₁ , ..., v _V ∈V as positive example probabilities. Output as distribution parameters β _p = {p (v ₁ ), ..., p (v _V )}.

  Next, the negative example topic determination unit 112 obtains, for each topic information, information representing a topic probability distribution that is an appearance probability distribution of entities included in a set of text data corresponding to the same topic information. At least a part of topic information is selected as negative example topic information based on the distance between the positive example probability distribution and the topic probability distribution obtained using the information and the topic probability distribution. That is, the negative example topic determination unit 112 obtains an information amount distance between the positive example probability distribution and the topic probability distribution, and selects negative example topic information from the topic information corresponding to the topic probability distribution having a large information amount distance. .

ただし、0≦λ≦1であり、D_KLはKL-divergenceを表す。確率分布p,qの間のKL-divergenceは以下のように定義される。

ただし、xは確率変数を表す。
この場合、負例トピック決定部１１２は、p=β_P、q∈{β_1, β_2, … , β_T}とし、β_Pと各β_t(t=1,2,…,T)との間のJS-divergenceを計算する。負例トピック決定部１１２は、例えば、(1)正例確率分布とのJS-divergenceがある一定の閾値以上のパラメータβ_tに対応するトピック情報、或いは(2)正例確率分布とのJS-divergenceの大きな方から順にN個のパラメータβ_tに対応するトピック情報を負例トピック情報とする。負例トピック決定部１１２は、負例トピック情報を特定する情報（例えばt）を出力する。 The parameter of the positive probability distribution is β _p = {p (v ₁ ), ..., p (v _V )} described above, and information (parameter) representing the topic probability distribution is β _t = {p ( v ₁ | z _t ), ..., p (v _V | z _t )} (t = 1,..., T and T are positive integers). KL-divergence and JS-divergence are used as the distance measure between probability distributions, but here JS-divergence with symmetric distance is used. JS-divergence D _JS (q || p) between two probability distributions p and q is defined as follows.

However, 0 ≦ λ ≦ 1, and D _KL represents KL-divergence. The KL-divergence between the probability distributions p and q is defined as follows.

Where x represents a random variable.
In this case, the negative example topic determination unit 112 sets p = β _P , q∈ {β _1, β _2, ..., Β _T }, and β _P and each β _t (t = 1, 2,..., T) Calculate JS-divergence between The negative example topic determination unit 112 may, for example, (1) Topic information corresponding to a parameter β _t having a certain threshold value or more with a JS-divergence with a positive example probability distribution, or (2) JS- Topic information corresponding to N parameters β _t in order from the largest divergence is set as negative example topic information. The negative example topic determination unit 112 outputs information for specifying negative example topic information (for example, t).

Information specifying negative example topic information is input to the negative example seed entity generation unit 113. The negative example seed entity generation unit 113 selects an entity corresponding to the negative example topic information selected by the negative example topic determination unit 112 as the negative example seed entity RN _e ⁰ . An example of a method for selecting such a negative example seed entity RN _e ⁰ is as follows.
(Selection Method 1) A character string such as a word is selected as a negative example seed entity RN _e ⁰ .
(Selection Method 2) Text data such as a document is selected as a negative example seed entity.
In any of the selection methods, an entity that is strongly related to the negative example topic information (a negative example topic has a high degree of contribution) is set as a negative example seed entity RN _e ⁰ . Specific examples of the selection methods 1 and 2 are shown below.
(Selection Method 1) When a character string v such as a word is selected as the negative example seed entity RN _e ⁰ , the negative example seed entity generation unit 113 uses the parameters β _t to p (z _t | corresponding to the negative example topic information. v) is calculated as follows.
p (z _t | v) = p (v _t | z) p (z) / Σ _z p (v | z) p (z)
The negative example seed entity generation unit 113 selects a character string v having a large value p (z _t | v) as the negative example seed entity RN _e ⁰ . For example, the negative examples seed entity generating unit 113, p | predetermined number in descending order of (z _t v) of p (z _t | v) is selected, the selected p (z _t | v) string corresponding to v Is a negative seed entity RN _e ⁰ . Alternatively, the negative example seed entity generation unit 113 selects, for example, p (z _t | v) larger than the threshold, and sets the character string v corresponding to the selected p (z _t | v) as the negative example seed entity RN _e. ^Set to ⁰ .
(Selection method 2) When text data such as a document is selected as a negative example seed entity RN _e ⁰ , for example, a posteriori probability p (z | d) that is topic information corresponding to all text data d is calculated in advance. The text data d having a large value of p (z | d) is stored in the storage unit 11a for each topic (for each hidden variable z). For example, a predetermined number of text data d is selected for each topic in descending order of p (z | d) and stored in the storage unit 11a, or p (z larger than the threshold for each topic z The text data d corresponding to | d) is selected and stored in the storage unit 11a. The negative example seed entity generation unit 113 thus stores p (z _t | d) corresponding to the negative example topic information obtained by the negative example topic determination unit 112 from the text data d stored in the storage unit 11a. Is selected as a negative example seed entity RN _e ⁰ (end of description of <Method 1>).

<Method 2>
The data extraction apparatus 1 of this example includes an automatic generation unit 120 that automatically generates a negative example seed entity RN _e ⁰ (FIG. 1). The automatic generation unit 120 includes a seed positive example topic score generation unit 121, a negative example topic determination unit 122, and a negative example seed entity generation unit 113 (FIG. 2B).

First, the seed positive example topic score creation unit 121 aggregates seed positive example topic information indicating the appropriateness of each topic z with respect to the text data d including the positive example seed entity RP _e ⁰ for each topic (for each hidden variable z). Then, the tabulation result for each topic (for each hidden variable z) obtained thereby is obtained as a seed positive example topic score for the topic. For example, the seed positive example topic score creating unit 121 sums the posterior probabilities p (z | d) obtained by the topic assigning unit 12 corresponding to the positive example document PD, that is, _{Σ_dεPD} p (z | d ) Is calculated for each topic (for each hidden variable z), and is used as a seed positive example topic score for the topic (hidden variable z). Alternatively, for example, a monotonically increasing function value of _{Σ_dεPD} p (z | d) may be used as a seed positive example topic score for the topic (hidden variable z).

Next, the negative example topic determination unit 122 sets topic information corresponding to the topic selected based on the magnitude of the topic seed positive example topic score as negative example topic information. For example, the negative example topic determination unit 122 selects a predetermined number of topics (hidden variable z) in ascending order of the seed positive example topic score, and sets topic information corresponding to the selected topic as negative example topic information. Alternatively, the negative example topic determination unit 122 selects a topic (hidden variable z) whose seed positive example topic score is equal to or less than a predetermined threshold, and sets topic information corresponding to the selected topic as negative example topic information. The negative example topic determination unit 122 outputs information specifying negative example topic information (for example, _t corresponding to the hidden variable z t).

Thereafter, as in <Method 1>, the negative example seed entity generation unit 113 selects the entity corresponding to the negative example topic information selected by the negative example topic determination unit 122 as the negative example seed entity RN _e ⁰ (< End of description of Method 2>).

According to the method for automatically generating the negative example seed entity RN _e ⁰ described above, an initial set of negative examples (negative example seed entity) is automatically generated from topic information having low relevance to the positive example entity. The possibility of drifting can be reduced and the accuracy of the resulting entity set can be increased.

The feature extraction unit 13 extracts any positive example entity RP _e ^j-1 (the initial positive example entity RP _e ⁰ is the positive example seed entity RP _e ⁰ ) from the text data set D stored in the storage unit 11a. The “original text” that is the character string to be included is extracted. Examples of the positive example text are sentences, phrases, word strings, and the like included in the text data. One or more positive example texts are extracted for a set of positive example entities RP _e ^j-1 and text data. The feature extraction unit 13 extracts a feature fP ′ _e ^j representing the feature of the positive example entity RP _e ^j−1 determined by the relationship with the extracted positive example text. In this example, the positive example entity RP _e ^j-1 of a feature fP _'e ^j is extracted for each positive example text containing positive example entity RP _e ^j-1. In the following, the feature fP ′ _e ^j of the positive example entity RP _e ^j−1 is exemplified.

[Example of feature fP ' _e ^j of positive entity RP _e ^j-1 ]
Positive example entity RP _e ^j-1 of a feature fP _'e ^j corresponds to the positive examples text (positive example entity RP _e ^j-1 that contains the text data to a string containing a) positive example text and the Contains information representing the relationship with the positive entity RP _e ^j-1 . Any information may be used as a feature as long as it is such information.
For example, the position in the entity (matching entities) from within the front and rear predetermined number of words (positive examples in the text) that matches the positive examples entity RP _e ^j-1 in the text data including either positive examples entity RP _e ^j-1 A pair of information indicating the relative position of the surrounding word with respect to the matching entity (surface feature), part of speech information (part of speech feature) and proper noun information (proprietary noun feature) of the matching entity or surrounding word ) or syntax information (syntax feature), positive cases entity RP _e positive example entity RP _e ^j-1 of the number of occurrences of at ^j-1 number of occurrences and the set D of the text data in the text data (number of occurrences feature) Among them, information corresponding to at least one of them can be set as a feature fP ′ _e ^j .

Examples of surface features are "ex + 1 =" is "", "ex-1 =" in "", etc., and these are relative words ("wa" in the former example) and relative positions of the surrounding words relative to the matching entity ( In the former example, it is information that identifies a pair with information representing “ex + 1”). “Ex” represents a matching entity, “ex + β” represents a word after β words of the matching entity ex, and “ex-β” represents a word before β words of the matching entity ex. Examples of part-of-speech features are “ex + 1 = POS: particle” and “ex = POS: noun”, which are relative positions of surrounding words relative to the matching entity (“ex + 1” in the former example and “ex + 1” in the latter example) “Ex”) and information that identifies a set of matching entities or parts of speech of surrounding words. Examples of proper noun features are “ex = ORG”, “ex-1 = ORG”, etc., which are information that identifies the relative position of the surrounding word relative to the matching entity and the matching entity or the proper noun of the surrounding word. . An example of the syntactic feature is information indicating the “dependency hierarchy” of the matching entity in the positive example text. Examples of Occurrences feature is described in a positive example number of entities RP _e ^j-1, including a set D of the text data and the text data (Example of positive example entity RP _e ^j-1 of a feature fP _'e ^j] the end).

Similarly, the feature extraction unit 13 extracts any negative example entity RN _e ^j-1 (the initial negative example entity RN _e ⁰ is a negative example seed entity RN _e) from the text data set D stored in the storage unit 11a. “Negative text” that is a character string including “ ^0” ) is extracted. Examples of negative example text are sentences, phrases, word strings, etc. included in the text data. One or more negative example texts are extracted for a set of negative example entity RN _e ^j-1 and text data. The feature extraction unit 13 extracts a feature fN ′ _e ^j representing the characteristics of the negative example entity RN _e ^j−1 determined by the relationship with the extracted negative example text. Negative examples entity RN _e ^j-1 of a feature fN _'e ^j corresponds to the negative sample text (a negative example entity RN string containing _e ^j-1 that contains the text data), a negative sample text and the Contains information representing the relationship with the negative entity RN _e ^j-1 . In this example, the negative examples entity RN _e ^j-1 of a feature fN _'e ^j is extracted for each negative example text containing a negative example entity RN _e ^j-1. A specific example of the feature fN ′ _e ^j of the negative example entity RN _e ^j−1 is the same as the case of the feature fP ′ _e ^j of the positive example entity RP _e ^j−1 described above. For example, "fP to" positive cases "to" negative examples "" RP _e ^j-1 "is" RN _e ^j-1 "in the specific example of the positive sample entity RP _e ^j-1 of a feature fP _'e ^j described above ' _e ^j-1 ' is replaced by 'fN' _e ^j-1 '.
The feature extraction unit 13 includes a pair (fP ′ _e ^j , <+1>) of a feature fP ′ _e ^j of the positive example entity RP _e ^j−1 and a label <+1> representing the positive example, and a negative example entity A pair (fN ' _e ^j , <-1>) of a feature fN _e ^j of RN _e ^j-1 and a label <-1> representing a negative example is output.

<< Topic Information Extraction: Step S14 >>
A pair of a positive example entity RP _e ^j−1 , a negative example entity RN _e ^j−1 , a feature fP ′ _e ^j of the positive example entity RP _e ^j−1 and a label <+1> representing the positive example (fP ′ _e ^j , <+1>), and a pair (fN ' _e ^j , <-1>) of the feature fN _e ^{j of} the negative example entity RN _e ^{j-1 and} the label <-1> representing the negative example Input to the unit 14.
The topic information extraction unit 14 uses the topic information (correct information) included in the text data with topic information including the text data including the positive entity RP _e ^j-1 from the set D ′ of text data with topic information stored in the storage unit 11c. Example Topic information corresponding to text data including entity RP _e ^j-1 is selected. The topic information selected in this way will be referred to as “normal topic information” corresponding to a set of the positive entity RP _e ^j-1 and text data. Since topic information is given for each text data, the same example topic information corresponds to each example text included in the same text data. The topic information extraction unit 14 sets positive example topic information corresponding to a set of the positive example entity RP _e ^j-1 and text data, and each positive example entity RP _e ^j− corresponding to each positive example text included in the text data. ^In addition to the feature fP ′ _e ^j of ^1, the feature of each positive example entity RP _e ^j−1 corresponding to each positive example text is updated to fP _e ^j . That is, positive example entity RP _e ^j-1 of a feature fP _e ^j corresponding to positive cases text is a positive example entity RP _e ^j-1 of a feature fP _'e ^j and positive sample topic information corresponding to the positive example text including. Thus, the positive example topic information is made part of the feature fP _e ^j .

Similarly, the topic information extraction unit 14 includes topics included in the text data with topic information including text data including the negative entity RN _e ^j-1 from the set D ′ of text data with topic information stored in the storage unit 11c. Select information (topic information corresponding to text data including negative example entity RN _e ^j-1 ). The topic information selected in this way will be referred to as “negative example topic information” corresponding to the set of negative example entity RN _e ^j−1 and text data. Since topic information is provided for each text data, the same negative example topic information corresponds to each negative example text included in the same text data. The topic information extraction unit 14 sets negative example topic information corresponding to a set of negative example entity RN _e ^j-1 and text data, and each negative example entity RN _e ^j− corresponding to each negative example text included in the text data. ^In addition to the feature fN ′ _e ^j of ^1, the feature of each negative example entity RN _e ^j−1 corresponding to each negative example text is updated to fN _e ^j . That is, the negative examples entity RN _e ^j-1 of a feature fN _e ^j corresponding to the negative example text, a negative example entity RN _e ^j-1 of a feature fN _'e ^j and negative cases topic information corresponding to the negative sample text Including. Thus, the negative example topic information is part of the feature fN _e ^j .

Note that features fP _e ^j and fN _e ^j corresponding to all positive example texts and negative example texts included in the set D ′ of text data with topic information may be generated, or some positive example texts and negative examples Only the features fP _e ^j and fN _e ^j corresponding to the text may be generated. Hereinafter, an example in which only the features fP _e ^j and fN _e ^j corresponding to some positive example texts and negative example texts are generated will be described.
[Example of generating only features corresponding to some positive texts and negative texts]
The ambiguous positive example entity RP _e ^j-1 may be input to the feature extraction unit 13. For example, a positive entity RP _e ^j-1 such as <Hanshin> may be used as a team name or a company name. In this example, the positive examples entity in the set D 'topic information with text data RP _e ^j-1 is estimated what kind of are often used in the sense, positive cases entities in the estimated mean RP _e Select positive example topic information and negative example topic information only for character strings (positive example text and negative example text) that are assumed to be used by ^j-1, and identify the positive example entity RP _e ^j-1 Generate features fN _e ^j of fP _e ^j and negative entity RN _e ^j-1 . Thereby, the improvement of the learning precision in the identification learning part 15 mentioned later and the identification precision in the entity identification part 16 is anticipated.

First, in order to estimate what the positive example entity RP _e ^j-1 is used in the set D ′ of text data with topic information, the feature extraction unit 13 sets the set D of text data with topic information. The character strings (positive example text and negative example text) included in each text data included in 'are given a topic candidate of the character string and a topic candidate score indicating the appropriateness of the topic candidate for the character string. . Topic Candidate scores, for example, either be calculated using the topic models TM ⁰ described above, is calculated from the information obtained in the process in step S12 described above are stored in the storage unit 11c. Hereinafter, topic candidate scores s (z _n , v) corresponding to z _n (n = 1,..., Z) corresponding to each topic candidate and a character string v included in the text data will be exemplified.
s (z _n , v) = p (z _n | v) = p (v | z _n ) p (z _n ) / p (v)… (9)
Note that p (v | z _n ) and p (z _n ) are obtained as parameters of the topic model TM ⁰ at z = z _n , and p (v) is obtained from equation (5) at z = z _n . It is obtained from the joint probability p (z _n , v) and the parameter p (z _n ) = Σ _z p (v | z) p (z _n ).

その他、式(9)(10)の写像をトピック候補スコアs(z_n,v)としてもよい。 Further, the following topic candidate scores s (z _n , v) may be used.

In addition, the mapping of equations (9) and (10) may be used as the topic candidate score s (z _n , v).

その他、トピック候補スコアs(z_n,v)(v∈V_p)の単調増加関数値を各正例トピックスコアS(z_n)(n=1,...,Z)とするなど、その他の集計方法で正例トピックスコアS(z_n)が計算されてもよい。 Next, the feature extraction unit 13 totals the topic candidate scores corresponding to the same topic candidates, and sets the total result for each topic candidate as a positive example topic score for each of the topic candidates. For example, the feature extraction unit 13 calculates each positive example topic score S (z _n ) (n = 1,..., Z) corresponding to each of the topic candidates according to the equation (11), and stores it in the storage unit 11c. Store. V _p represents a set of positive example texts.

In addition, the monotonically increasing function value of the topic candidate score s (z _n , v) (v∈V _p ) is set as each positive example topic score S (z _n ) (n = 1, ..., Z), etc. The positive example topic score S (z _n ) may be calculated by the aggregation method.

Next, the feature extraction unit 13 selects a candidate for a topic in which each positive example topic score S (z _n ) (n = 1,..., Z) satisfies a specific criterion, and selects it as a positive criterion reference topic S _e. Is stored in the storage unit 11c. The number of positive cases reference topic S _e to be selected may be two or more may be one. For example, most to a large positive example topic score S (z _n) in the candidate corresponding topic values may be a positive example reference topic S _e, a predetermined number of positive cases topic score S for a large order chosen value A candidate for a topic corresponding to each of (z _n ) may be set as a positive example reference topic S _e , or a candidate for a topic corresponding to a positive example topic score S (z _n ) greater than or equal to a reference value is a positive example reference topic S _It may be _e .

Feature extraction unit 13, for example, a positive sample text corresponding to the candidate of the same topic and one of positive cases reference topic S _e (string containing the text data), yet, corresponding to the candidate of the same topic Generates the feature fP _e ^j of the positive example entity RP _e ^j-1 corresponding to the positive example text whose topic candidate score satisfies a certain criterion, but does not generate the features of the positive example entity corresponding to other positive example texts . Further, feature extraction unit 13 is, for example, a negative example text corresponding to the candidate of the same topic and one of positive cases reference topic S _e (string containing the text data), yet, the candidate of the same topic Generates the feature fN _e ^j of the negative example entity RN _e ^j-1 corresponding to the negative example text whose corresponding topic candidate score satisfies a certain criterion, but the negative example entity corresponding to other negative example texts Do not generate.

A specific feature generation example is shown below.
Feature Generation Example 1: feature extraction unit 13 is a positive example text corresponding to the same topic candidates either positive cases reference topic S _e, yet, topic candidate score corresponding to the candidate of the same topic the Generates the feature fP _e ^j of the positive example entity RP _e ^j-1 corresponding to the maximum positive example text among the topic candidate scores corresponding to the positive example text, but positive examples corresponding to other positive example texts Do not generate entity features. Further, e.g., feature extraction unit 13 is a negative example text corresponding to the candidate of the same topic and one of positive cases reference topic S _e, yet, topic candidate score corresponding to the candidate of the same topic the negative negative sample entity but generates a negative example entity RN _e ^j-1 of a feature fN _e ^j corresponding to the negative example text with the maximum in the topic candidate score corresponding to the example text corresponding to the other negative sample text Do not generate features of.

Feature Generation Example 2: feature extraction unit 13 is a positive example text corresponding to the same topic candidates either positive cases reference topic S _e, yet, topic candidate score corresponding to the candidate of the same topic reference The feature fP _e ^j of the positive example entity RP _e ^j-1 corresponding to the positive example text that is greater than or equal to the value is generated, but the feature of the positive example entity corresponding to the other positive example text is not generated. Further, for example, feature extraction unit 13, positive example reference topic is negative example text corresponding to the candidate of the same topic and one of S _e, yet, topic candidate score reference value corresponding to the candidate of the same topic The feature fN _e ^j of the negative example entity RN _e ^j-1 corresponding to the negative example text as described above is generated, but the feature of the negative example entity corresponding to other negative example texts is not generated.

Feature Generation Example 3: feature extraction unit 13 is a positive example text corresponding to the same topic candidates either positive cases reference topic S _e, yet, topic candidate score corresponding to the candidate of the same topic the Generates the feature fP _e ^j of the positive example entity RP _e ^j-1 corresponding to the maximum positive example text among the topic candidate scores corresponding to the positive example text, but positive examples corresponding to other positive example texts Do not generate entity features. On the other hand, the negative examples entity RN _e ^j-1 of a feature fN _e ^j, all negative examples entity RN corresponding to negative sample text _e ^j-1 of a feature fN _e ^j is generated.

Feature Generation Example 4: feature extraction unit 13 is a positive example text corresponding to the same topic candidates either positive cases reference topic S _e, yet, topic candidate score corresponding to the candidate of the same topic reference The feature fP _e ^j of the positive example entity RP _e ^j-1 corresponding to the positive example text that is greater than or equal to the value is generated, but the feature of the positive example entity corresponding to the other positive example text is not generated. On the other hand, the negative examples entity RN _e ^j-1 of a feature fN _e ^j, negative examples entity RN _e ^j-1 of a feature fN _e ^j are generated for all of the negative sample text ([part of the positive examples End of description of example] [Only features corresponding to text and negative example text are generated].

The topic information extraction unit 14 includes a pair (fP _e ^j , <+1>) of the feature fP _e ^j of the positive example entity RP _e ^{j-1 and} the label <+1> representing the positive example, and the negative example entity RN A pair (fN _e ^j , <-1>) of a feature fN _e ^j of _e ^j-1 and a label <-1> representing a negative example is output.

FIG. 5B is a diagram illustrating a pair (fP _e ^j , <+1>) and a pair (fN _e ^j , <-1>) output from the topic information extraction unit 14. “POS” represents a part-of-speech feature, and “BOS” represents that no word exists at the corresponding position. For example, the identity of the example entity ex = <Hiroshima> corresponding to the example text included in the text data <T1> is fP _e ^j = (ex-2 = "Yakult", ex-2 = POS: noun, ex-1 = "VS", ex-1 = POS: noun, ex + 1 = "no", ex + 1 = POS: particle, ex + 2 = "Yakult", ex + 2 = POS: particle, topic information = (z ₂ , 08)).

<< Identification Learning: Step S15 >>
A pair (fP _e ^j , <+1>) of the feature fP _e ^{j of the} positive example entity RP _e ^{j-1 and} the label <+1> representing the positive example, and the feature fN of the negative example entity RN _e ^j-1 A set (fN _e ^j , <-1>) of _e ^j and a label <−1> representing a negative example is input to the identification learning unit 15. The feature fP _e ^j of the positive example entity RP _e ^j-1 can be used as supervised learning data for the positive example, and the feature fN _e ^j of the negative example entity RN _e ^j-1 can be used as supervised learning data for the negative example. The identification learning unit 15 performs the identification model ME _e ^j by learning processing using the feature fP _e ^j of the positive example entity RP _e ^j-1 and the feature fN _e ^{j of the} negative example entity RN _e ^j-1 as supervised learning data. Is generated. This identification model ME _e ^j is a function that outputs the information for identifying whether the entity is a positive example entity or a negative example entity with the feature of an arbitrary entity as an input. Any model may be used as long as it is such an identification model ME _e ^j .
An example of the discriminant model ME _e ^j is the maximum entropy model with regularization term (Reference 1 “Berger, AL, Pietra, VJD and Pietra,“ A maximum entropy approach to natural language processing ”, SAD 1996.”), regularization Conditional random fields with terms (CRFs, Ref. 2, "Lafferty, J. and McCallum, A. and Pereira, F." Conditional random fields: Probabilistic models for segmenting and labeling sequence data ", MACHINE LEARNING, pp. 282- 289, 2001. ", support vector machines (SVMs, reference 3" Vapnik, VN "the nature of statistical learning theory", Springer Verlag, 1995. "), and the like. in the learning of the identification model ME _e ^j of each example , For the feature fP _e ^{j of} the positive example entity RP _e ^j-1 and the feature fN _e ^{j of the} negative example entity RN _e ^j-1 used as supervised learning data, the degree of influence on the identification model ME _e ^j attached indicator representing the magnitude (weight for feature) is, they identify the identification model ME _e ^j In particular, the model as exemplified in References 1-3 is not a model in which all features are weighted (for example, a maximum entropy model without a regularization term), and is determined to be effective for identification. Only specific features are weighted, and a specific example of a maximum entropy model with a regularization term is shown below.

に対するエントロピー

を最大化する各重み（パラメータ）λ_qに対応するP_λ(y|x)であるP(y|x)を識別モデルME_e ^jとする。ただし、

であり、qは各学習データ(x,y)の組にそれぞれ対応するラベルであり、p'(x)は学習データ(x,y)におけるxの出現頻度であり、f_q(x,y)はqに対応する素性関数(feature function)である。 When the maximum entropy model with a regularization term is used, the discriminative learning unit 15 uses (x, y) ∈ {(fP _e ^j , <+1>), (fN _e ^j , <-1>)} as learning data. As a conditional probability

Entropy for

P (y | x) that is P _λ (y | x) corresponding to each weight (parameter) λ _q that maximizes is set as an identification model ME _e ^j . However,

Q is a label corresponding to each set of learning data (x, y), p ′ (x) is the frequency of occurrence of x in the learning data (x, y), and f _q (x, y ) Is a feature function corresponding to q.

Here, each weight λ _q serves as an index representing the magnitude of the influence of the feature fP _e ^j or fN _e ^j of the learning data (x, y) corresponding to q on the identification model ME _e ^j . In the example of the maximum entropy model with regularization terms, the weight λ _q is not assigned to the feature fP _e ^j or fN _e ^{j of} all the learning data (x, y), and the feature with low importance The weight λ _q corresponding to is not attached. That is, the weight λ _q corresponding to the feature of low importance is set to 0.

If only the features fP _e ^j and fN _e ^j corresponding to some positive example texts and negative example texts are generated in step S14, the features corresponding to some positive example texts and negative example texts are generated. Only (x, y) ε {(fP _e ^j , <+1>), (fN _e ^j , <-1>)} corresponding to fP _e ^j and fN _e ^j is used as learning data. For example, when the features fP _e ^j and fN _e ^j are generated as in the above-mentioned “example in which only features corresponding to some positive example texts and negative example texts are generated”, the above-described positive example criteria Positive example text corresponding to the same topic candidate as any of the topics (a character string included in the text data) and corresponding to the positive example text in which the topic candidate score of the same topic candidate satisfies a specific criterion Only the features of entities and / or negative example entities are taken as supervised learning data.
The identification model ME _e ^j generated by the learning process is stored in the storage unit 11d. For example, the parameter identification model ME _e ^j generated by the learning processing is stored in the storage unit 11d.

<< Entity Identification: Step S16 >>
The entity identifying unit 16 selects any text data with topic information from the set D ′ of topic data with topic information stored in the storage unit 11c, and a character string included in the text data included in the selected text data with topic information. Is an entity RD _e ^j .

  Note that all the text data with topic information may be selected from the set D ′ of text data with topic information, but it is not preferable in terms of calculation efficiency to make all the text data to be identified. For this reason, it is desirable to select an identification target by a specific method. Specific examples are shown below.

[Example of selection method]
Selection method example 1:
In example 1 of the selection method, the entity identification unit 16 uses the feature fP _e ^{j of} the positive example entity RP _e ^{j-1 and} the feature of the negative example entity RN _e ^j-1 used as the supervised learning data by the identification learning unit 15. Among the fN _e ^j , an index (for example, the above-mentioned weight λ _q ) indicating the degree of influence on the identification model ME _e ^j generated from the features satisfies a specific criterion, that is, the identification model ME _e ^j A feature fP _e ^j and / or fN _e ^j that has a large influence on is selected. For example, the entity identification unit 16 selects a feature fP _e ^j and / or fN _e ^{j in} which the absolute value of the weight λ _q is larger than a threshold value.
In addition, the entity identification unit 16 selects text data with topic information including text data including a character string corresponding to the selected feature fP _e ^j and / or fN _e ^j , and the selected text data with topic information includes. An entity that is a character string included in the text data is set as a target entity RD _e ^j . For example, the entity identification unit 16 extracts a surface feature word from the selected feature fP _e ^j and / or fN _e ^j, and selects text data with topic information including text data including the surface feature word. For example, the feature FNC (x−2 = “POS: noun”, x−1 = “in”) where the selected features consist of a combination of surface features and part-of-speech features before the entity ex (FNC is If it is a function), the entity identification unit 16 extracts the surface feature word “de” from the selected feature FNC (x−2 = “POS: noun”, x−1 = “de”), and the word “ Select text data with topic information that includes text data containing "

Selection method example 2:
Example 2 selection method, entity identification unit 16 is a positive example text corresponding to the same topic candidates either positive cases reference topic S _e previously described (string containing the text data), yet, of the same An entity included in a positive example text whose topic candidate score satisfies a specific criterion is set as a target entity RD _e ^j .
For example, the entity identification unit 16 is a positive example text corresponding to the same topic candidates either positive cases reference topic S _e, yet, topic candidate score is the positive example text corresponding to the candidate of the same topic The entity included in the positive example text that is the largest among the topic candidate scores corresponding to is the target entity RD _e ^j .
Alternatively, for example, the entity identification unit 16 is a positive example text corresponding to the same topic candidates either positive cases reference topic S _e, yet, topic candidate score reference value corresponding to the candidate of the same topic The entity included in the above-described positive example text is set as a target entity RD _e ^j (end of the description of [example of selection method]).

Feature extraction unit 13, from the set D of text data stored in the storage unit 11a, extracts the "target text" is a string containing the target entity RD _e ^j. Examples of the target text are sentences, phrases, word strings, and the like included in the text data. One or more target texts are extracted for a set of the target entity RD _e ^j and text data.
The feature extraction unit 13 extracts a feature fD ′ _e ^j representing the characteristics of the target entity RD _e ^j determined by the relationship with the extracted target text. Feature fD _'e ^j of the target entity RD _e ^j corresponds to the (those contained in the text data to a string containing the target entity RD _e ^j) target text, the target text and the relevant target entity RD _e ^j Contains information representing the relationship. The specific process is the same as the case of extracting the feature fP ′ _e ^j of the positive entity RP _e ^j−1 described above. For example, “positive entity RP _e ^j ” is replaced with “target entity RD _e ^j ”, “feature fP ' _e ^j ” is replaced with “feature fD' _e ^j ”, and “positive text” is replaced with “target text”. Is the same as the process of extracting the feature fP ′ _e ^j of the positive entity RP _e ^j−1 described above.

The feature fD ′ _e ^j of the target entity RD _e ^j corresponding to the target text is input to the topic information extraction unit 14. The topic information extraction unit 14 selects topic information (topic information corresponding to the target text) included in the text data with topic information including the target text from the storage unit 11c. Thus the selected topic information, corresponding to the set of target entities RD _e ^j and text data is referred to as a "target topic information". Since topic information is given for each text data, the same target topic information corresponds to each target text included in the same text data. The topic information extraction unit 14 converts the target topic information corresponding to the set of the target entity RD _e ^j and the text data into the feature fD ′ _e ^j of each target entity RD _e ^j corresponding to each target text included in the text data. In addition, the feature of each target entity RD _e ^j corresponding to each target text is updated to fD _e ^j . That is, feature fD _e ^j of the target entity RD _e ^j corresponding to the target text includes a feature fD _'e ^j and the target topic information of the target entity RD _e ^j corresponding to the target text. In this way, the target topic information is part of the feature fD _e ^j .

Feature fD _e ^j of the target entity RD _e ^j is input to the entity identification unit 16. The entity identification unit 16 inputs the feature fD _e ^j of the target entity RD _e ^j to the identification model ME _e ^j read from the storage unit 11d, and identifies whether the target entity RD _e ^j is a positive example entity or a negative example entity. For example, the identification model ME _e ^{when j} maximum entropy model with regularization term is used as a, x = fD _e ^j identification model ME _e is ^j P | is substituted into (y x) P (1 | x ) And P (−1 | x), and by comparing an index (such as reliability) corresponding to them with a threshold value, the target entity RD _e ^j is identified as a positive example entity or a negative example entity.

If the target entity is identified as a positive entity, the entity identifying unit 16 stores the target entity RD _e ^j as a new positive entity RP _e ^j in the storage unit 11e. On the other hand, when the target entity is identified as a negative example entity, the entity identification unit 16 stores the target entity RD _e ^j as a new negative example entity RN _e ^j in the storage unit 11e.

<< Convergence determination: steps S17 to S19 >>
The convergence determination unit 17 determines whether the convergence condition is satisfied. Examples of convergence conditions are given below.
[Example of convergence condition]
Example of convergence condition 1: The convergence determination unit 17 of this example determines that the convergence condition is satisfied when there is no target entity RD _e ^j newly assigned to the positive example entity RP _e ^j .
Convergence condition example 2: The convergence determination unit 17 in this example satisfies the convergence condition when there is no newly allocated target entity RD _e ^j-1 even if the iterations of steps S13 to S17 are repeated more than the reference number of times. to decide.
Example 3 of convergence condition: The convergence determination unit 17 in this example determines that the convergence condition is satisfied when the value of j is equal to or greater than the reference value (end of explanation of [example of convergence condition] / step S17).
When the convergence determination unit 17 determines that the convergence condition is satisfied, the iterations from step S13 to S17 are completed, and the output unit 18 outputs all the positive example entities RP ^j _e stored in the storage unit 11e for processing. The process ends (step S19). Otherwise, the control unit 19 sets j + 1 as a new value of j (step S18), and the positive example entity RP ^j _e and the negative example entity RN ^j _e stored in the storage unit 11e are used as a feature extraction unit. 13 and the iterations of steps S13 to S16 are executed.

<Identification examples>
The specific identification example in 1st Embodiment is illustrated.
In this example using the topic models TM ⁰ corresponding to the candidate of the two topics. Specifically, a topic model TM ^{0 in} which “Team name” and “Company name” are topic candidates is used. Also assume that the positive seed entity is <Hiroshima>, the negative seed entity is <Mainichi Newspaper>, and the user is requesting a set of entities for the team name.
Further, the posterior probability of the topic “team name” for the text data d ₁ including the positive seed entity <Hiroshima>, calculated using the topic model TM ⁰ , is p (team name | d ₁ ) = 0.9, and Assume that the posterior probability of the topic “company name” for the text data d ₁ is p (company name | d ₁ ) = 0.1. On the other hand, the posterior probability of the topic “Team Name” for the text data d ₂ including the negative seed entity <Mainichi Shimbun> is p (Team name | d ₂ ) = 0.1, and the topic “Company Name” for the text data d ₂ Is the posterior probability of p (company name | d ₂ ) = 0.9. Here, when the topic information as described above, "Example topic information 1 (N = 2)" was established, topic information for the text data d ₁ is ((team name, 0.9), (company name, 0.1) ), and the topic information for the text data d ₂ ((team name, 0.1), and (company name, 0.9)) (step S12).

In this example, it is assumed that the feature extraction in step S13 is not performed and only topic information is used as a feature. In this case, the identity of the positive seed entity <Hiroshima> is ((Team name, 0.9), (Company name, 0.1)), and the identity of the negative seed entity <Mainichi Shimbun> is ((Team name, 0.1), ( Company name, 0.9)). Therefore, the learning data is
(((Team name, 0.9), (Company name, 0.1)), <+1>)
(((Team name, 0.1), (Company name, 0.9)), <-1>)
(Step S14).
An identification model is learned based on such learning data (step S15). As a result of the learning, an identification model having a large weight on the positive side with respect to “Team name” and a small weight on the positive side with respect to “Company name” will be obtained.

Next, it is assumed that the target entity <Hanshin> not included in the seed entity is input. Here, the posterior probability of the topic “team name” for the text data d ₃ including the target entity <Hanshin>, calculated in the same manner as described above, is p (ball team name | d ₁ ) = 0.8, and the text data d ₃ The posterior probability of the topic “company name” for p is (company name | d ₁ ) = 0.2. In this case, the identity of the target entity <Hanshin> is ((Team name, 0.8), (Company name, 0.2)). When this feature ((Team name, 0.8), (Company name, 0.2)) is identified by the above identification model, the result ((Team name, 0.8), (Company name, 0.2)) is correct. It can be determined that it corresponds to the example entity (step S16).

  By using topic information as at least part of the feature in this way, identification along the positive seed entity and negative seed entity representing the user's request is performed appropriately, and the extraction of the entity with suppressed semantic drift can be performed. It becomes possible. This is an advantage of using topic information. In particular, in the case of set expansion, the number of positive seed entities represented is often small, and the information that can be used is very limited. Therefore, if only the surrounding context is used as the feature, the data becomes sparse and the identification accuracy often decreases. . The topic information acts as an effective feature in identification in a scene where there is little data available in this way.

<Features of First Embodiment>
As described above, since the topic information is used as at least part of the feature in the method of the present embodiment, the semantic drift can be suppressed. The method of this embodiment can be used regardless of the type of text data that is a resource, and has a wide range of applications.

[Second Embodiment]
The second embodiment is a modification of the first embodiment, and suppresses the semantic drift using the attribute of the entity. An “attribute” is a character string in text data that represents the characteristics of an entity. Examples of such character strings are nouns, words, word strings, phrases, sentences, and the like. A specific example of the attribute is a noun existing within W words before and after the entity. W is an integer greater than or equal to 1 representing the window size. For example, if <Hanshin> in the text data "Hanshin game breaking news ..." is an entity and the window size is set to W = 3, <match> and <breaking news> are attribute candidates for the entity <Hanshin>. Is done.

The same attribute corresponds to a plurality of entities related to each other. For example, the entity <Yakult> having the same team name as the entity <Hiroshima> having the team name corresponds to the same attributes such as <match> and <pitcher>. For this reason, the attribute is a constraint condition to be satisfied by the entity to be searched. Using this fact, in the second embodiment, an entity is identified using a set of an entity and its attribute. For example, suppose that the positive seed entity is <Hiroshima>, which is a team name, and that it aims to acquire an entity, <Yakult>, which is also a team name. In this case, attributes common to these entities are <match>, <pitcher>, and the like. For example, the entity is identified using a pair of a positive seed entity <Hiroshima> and its attribute <match>. Here, <Yakult> which is a team name is a positive entity, but <Yakult> also has a meaning as a beverage name (for example, T5 in FIG. 5A). Therefore, <Yakult> is an entity that is prone to semantic drift. However, the attribute of the entity <Yakult>, which is the beverage name, is not <Game> or <Pitcher> but <Lactic acid bacteria> or <Beverage>. In this embodiment, by using a set of an entity and its attribute, it is possible to specify what the entity means and to reduce semantic drift.
Below, it demonstrates centering around the difference of 1st Embodiment, and abbreviate | omits description about the matter which is common in 1st Embodiment. Further, the same reference numerals as those in the first embodiment are used for portions common to the first embodiment.

<Configuration>
FIG. 6 is a block diagram for illustrating a functional configuration of the data extraction device 2 of the second embodiment.
As illustrated in FIG. 6, the data extraction device 2 includes storage units 11a, 11d, 11e, 21d, and 21e, an initial attribute set generation unit 22, an attribute identification feature extraction unit 23a, an entity identification feature extraction unit 23b, an attribute It has an identification learning unit 25a, an entity identification learning unit 25b, an attribute identification unit 26a, an entity identification unit 26b, a convergence determination unit 17, an output unit 18, and a control unit 19, and executes each process under the control of the control unit 19 To do.
Note that the data extraction device 2 is a special device configured by, for example, reading a special program into a known or dedicated computer. For example, the storage units 11a, 11d, 11e, 21d, and 21e are hard disks, semiconductor memories, and the like, and include an initial attribute set generation unit 22, an attribute identification feature extraction unit 23a, an entity identification feature extraction unit 23b, and an attribute identification learning unit. 25a, the entity identification learning unit 25b, the attribute identification unit 26a, the entity identification unit 26b, the convergence determination unit 17, the output unit 18, and the control unit 19 are a CPU or the like in which a special program is read. Further, at least a part of these may be configured by an integrated circuit or the like.

<Pre-processing>
As pre-processing, a set D of text data is stored in the storage unit 11a. The text data set D is the same as in the first embodiment.

<Data extraction process>
FIG. 7 is a diagram for illustrating a data extraction process of the data extraction device 2 of the second embodiment. In this embodiment, a co-training method that alternately updates entities and attributes is used. That is, in steps S22 to S24, the positive example and negative example entities are not updated, and only the positive example and negative example attributes are updated. On the other hand, in steps S25-S27, the positive and negative example attributes are not updated, and only the positive and negative example entities are updated. Hereinafter, the data extraction processing of the data extraction device 2 will be illustrated with reference to FIG.

<< Initialization: Step S11 >>
The control unit 19 initializes the value of j to j = 1.
<< Initial attribute set generation: Step S21 >>
The positive example seed entity RP _e ⁰ and the negative example seed entity RN _e ⁰ are input to the initial attribute set generation unit 22. For example, RP _e ⁰ = <Hiroshima> is input as a positive example seed entity, and RN _e ⁰ = <Japan> is input as a negative example seed entity. The positive seed entity RP _e ⁰ has been selected by the user. The negative example seed entity RN _e ⁰ may be selected by the user, or may be generated semi-automatically from the text data set D. A method for generating the negative seed entity RN _e ⁰ semi-automatically will be illustrated below.

[Example of semi-automatic generation of negative example seed entity RN _e ⁰ ]
A negative example seed entity generation unit (not shown) extracts and extracts a predetermined number of text data that does not contain any positive example seed entity RP _e ⁰ or a positive example attribute RP _a ⁰ described later from the text data set D. The nouns are selected at random from each of the text data and are output as negative example entity candidates. A display unit (not shown) displays these negative example entity candidates, and performs a display prompting the user to select a negative example seed entity from them. The selection by the user is input to the negative example seed entity generation unit, and the negative example seed entity generation unit outputs the selected negative example entity candidate as the positive example seed entity RP _e ⁰ ([negative example seed entity RN _e ⁰ End of description of semi-automatic generation method example].
The initial attribute set generation unit 22 uses the input positive example seed entity RP _e ⁰ , the negative example seed entity RN _e ^0, and the text data set D stored in the storage unit 11a, and uses the positive example seed entity RP _e ^0. A set of positive example attributes RP _a ⁰ that are character strings representing the attributes of the negative example attribute and a set of negative example attributes RN _a ⁰ that are character strings representing the attributes of the negative example seed entity RN _e ⁰ are generated.

(A) First, the initial attribute set generation unit 22 selects any character string other than the positive example entity RP _e ⁰ as a positive example attribute candidate from a set of text data including the positive example seed entity RP _e ⁰ . For example, the initial attribute set generation unit 22 obtains a predetermined number of positive example texts including the positive example seed entity RP _e ⁰ from the storage unit 11a, and directly or one phrase with the positive example seed entity RP _e ⁰ in each positive example text. Only words that are in the relationship of being held are extracted as positive example attribute candidates.

(B) Next, the initial attribute set generation unit 22 sets the correct attribute in the set D including all the text data and the frequency that the correct attribute candidate is included in the set of character strings including the correct example seed entity RP _e ^0. An index (statistic) indicating the magnitude of the difference from the frequency of including example attribute candidates is obtained, and a predetermined number of positive example attribute candidates from the one with the large index, that is, the positive example attribute having a large difference in frequency. _{Let the} candidate be the positive example attribute RP _a ⁰ (initial value of the positive example attribute). As there is a large difference between the positive example attribute candidates of these frequencies positive examples seed entity RP _e ⁰ and related strongly, deserves its positive example attribute RP _a ⁰ of the positive sample seed entity RP _e ^0. Examples of such indices are given below, but other statistics may be used.

χ²値が高い正例属性候補αほど、正例シードエンティティRP_e ⁰と関係の深い、即ち属性としてふさわしいといえる。よって、この例の初期属性集合生成部２２は、χ²値が高い正例属性候補αを正例属性RP_a ⁰として抽出する。例えば、χ²値が基準値以上となる正例属性候補αを正例属性RP_a ⁰とする。 [Example of metrics]
Indicator example 1:
In index example 1, the following χ ² values are used as indices.

It can be said that the positive example attribute candidate α having a higher χ ² value is more closely related to the positive example seed entity RP _e ⁰ , that is, suitable as an attribute. Therefore, the initial attribute set generation unit 22 of this example extracts the positive example attribute candidate α having a high χ ² value as the positive example attribute RP _a ⁰ . For example, a positive example attribute candidate α whose χ ² value is greater than or equal to a reference value is set as a positive example attribute RP _a ⁰ .

ここで|RP_e ⁰, α|は正例シードエンティティRP_e ⁰の集合と正例属性候補αとの組の出現頻度を表す。また、*はRP_e ⁰又はαのワイルドカードを表す。
PMI値が大きい正例属性候補αほど、正例シードエンティティRP_e ⁰と関係の深い、即ち属性としてふさわしいといえる。よって、この例の初期属性集合生成部２２は、PMI値が大きな正例属性候補αを正例属性RP_a ⁰として抽出する。例えば、PMI値が基準値以上となる正例属性候補αを正例属性RP_a ⁰とする（[指標の例]の説明終わり）。 Indicator example 2:
In the index example 2, the following PMIs in the two terms of the positive example seed entity RP _e ⁰ and the positive example attribute candidate α are used as the index.

Here, | RP _e ⁰ , α | represents the appearance frequency of a set of a set of positive example seed entities RP _e ⁰ and a positive example attribute candidate α. * Represents an RP _e ⁰ or α wild card.
It can be said that the positive example attribute candidate α having a larger PMI value is more closely related to the positive example seed entity RP _e ⁰ , that is, suitable as an attribute. Therefore, the initial attribute set generation unit 22 in this example extracts the positive example attribute candidate α having a large PMI value as the positive example attribute RP _a ⁰ . For example, a positive example attribute candidate α having a PMI value equal to or higher than a reference value is set as a positive example attribute RP _a ⁰ (end of description of [index example]).

In this method, first, the correct attribute candidate is roughly narrowed down using the syntax information in (A), so that the calculation time in (B) can be greatly reduced. Further, after extracting the positive example attribute RP _a ⁰ (initial value of the positive example attribute) by the above (A) and (B), it is manually checked whether or not an appropriate attribute is selected, and the final positive attribute is selected. The example attribute RP _a ⁰ may be determined.

Initial attribute set generation unit 22 performs the same processing for the negative sample seed entity RN _e ^0, extracting a negative sample attribute RN _a ^0. That is, the initial attribute set generation unit 22 selects one of the strings from the set of text data other than the negative examples seed entity RN _e ⁰ containing a negative example seed entity RN _e ⁰ as a negative example attribute candidate, a negative sample This represents the magnitude of the difference between the frequency that the negative example attribute candidate is included in the set of character strings including the seed entity RN _e ⁰ and the frequency that the negative example attribute candidate is included in the set D consisting of all text data. The negative example attribute candidate that satisfies the index, that is, the negative example attribute candidate having a large difference in frequency is set as a negative example attribute RN _a ⁰ (initial value of the negative example attribute).

Instead of the method described above, the initial attribute set generation unit 22 may select the negative example seed entity RN _e ⁰ and the negative example attribute RN _a ⁰ corresponding thereto semi-automatically. For example, the initial attribute set generation unit 22 extracts a predetermined number of text data that does not include any positive example seed entity RP _e ⁰ or positive example attribute RP _a ⁰ from the text data set D, and extracts each text data from the extracted text data Two nouns are selected at random, and one is output as a negative example entity candidate and the other as a negative example attribute candidate. A display unit (not shown) displays these and displays to prompt the user to select a negative example seed entity RN _e ⁰ and a corresponding negative example attribute RN _a ⁰ from these. The content selected by the user is input to the initial attribute set generation unit 22, and the initial attribute set generation unit 22 outputs a set of the selected negative example seed entity RN _e ⁰ and negative example attribute RN _a ⁰ .

The initial attribute set generation unit 22 outputs a set of positive example seed entities RP _e ^0, a set of negative example seed entities RN _e ^0, a set of extracted positive example attributes RP _a ⁰ , and _a set of negative example attributes RN _a ⁰ To do. For example, the initial attribute set generation unit 22 acquires text corresponding to T1, T2, and T10 as text including the positive seed entity RP _e ⁰ from the text data of FIG. 4, and T1, T2 by the above processing. A set {<VS>, <First game>, <Pitcher>} of positive example attributes RP _a ⁰ included in the text corresponding to is extracted and output. Similarly, the initial attribute set generation unit 22 acquires, for example, a text corresponding to T7 as a text including the negative example seed entity RN _e ⁰ , and extracts a set R {<population>} of the negative example attribute RN _a ^0. Output.

<< Attribute Identification Feature Extraction: Step S22 >>
A set of positive entity RP _e ^j-1, a set of negative example entity RN _e ^j-1, a set of positive example attributes RP _a ^j-1 , and a set of negative example attributes RN _a ^j-1 are attribute identification features. The data is input to the extraction unit 23a.
The attribute identifying feature extraction unit 23a is a set of a first positive example entity selected from the set of positive example entities RP _e ^j-1 and a first positive example attribute selected from the set of positive example attributes RP _a ^j-1. A first negative example entity and a positive example attribute selected from a set of a first positive example entity-positive example attribute pair PP ₁ (RP _e ^j-1 , RP _a ^j-1 ) and a negative example entity RN _e ^j-1 first negative example entity is a set of a first negative example selected attributes from the set of RN _a ^j-1 - to produce a negative example attribute pair ^{_{PN 1 (RN e j-1}} , RN a j-1). PP ₁ (RP _e ^j-1 , RP _a ^j-1 ) and PN ₁ (RN _e ^j-1 , RN _a ^j-1 ) are RP _e ^j-1 , RP _a ^j-1 , RN _e ^j-1 And RN _a ^j-1 may be generated for all possible combinations, or only some of them may be generated.
Next, the attribute identifying feature extracting unit 23a determines the positive entity RP _e ^{j-1 of} PP ₁ (RP _e ^j-1 , RP _a ^j-1 ) from the text data set D stored in the storage unit 11a. The “first positive example text”, which is a character string including a pair with the positive example attribute RP _a ^j−1 , is selected. Examples of the first positive example text are sentences, phrases, word strings, etc. included in the text data. One or more first example texts are extracted for a set of first example entity-example example attribute pair PP ₁ (RP _e ^j−1 , RP _a ^j−1 ) and text data.
The attribute identifying feature extraction unit 23a uses the first positive example entity-positive example attribute pair PP ₁ (RP _e ^j-1 , RP _a ^j-1 ) for the first positive example text as the first positive example information. The feature fP _a ^j of the example entity-positive example attribute pair PP ₁ (RP _e ^j−1 , RP _a ^j−1 ) is assumed. In this example, the feature fP _a ^j of PP ₁ (RP _e ^j−1 , RP _a ^j−1 ) is extracted for each first positive example text. An example of the feature fP _a ^j of PP ₁ (RP _e ^j-1 , RP _a ^j-1 ) is a character including the first example text (the example entity RP _e ^j-1 and the example attribute RP _a ^j-1) Information included in the text data) and the first positive example entity RP _e ^j-1 and the first positive example attribute RP _a ^j-1 .

For example, _a predetermined word before and after a character string (matching attribute) that matches the positive example attribute RP _a ^j-1 in text data including any positive example entity RP _e ^j-1 and the positive example attribute RP _a ^j-1 A pair (surface layer feature) of notation of words (neighboring words) located within a few (in the first example text) and information indicating the relative position of the neighboring words with respect to the matching attribute, part of speech information of matching attributes or surrounding words (Part of speech feature), proper noun information (proprietary noun feature), syntax information (syntactic feature), the number of appearances of matching attributes in text data, and the number of appearances of matching attributes in text data set D (appearance frequency feature) Among them, information corresponding to at least one of them can be set as _a feature fP _a ^j . This specific example is the same as [Example of feature fP ′ _e ^j of positive example entity RP _e ^j−1 ] of the first embodiment, except that the positive example attribute is used as a reference. For example, the positive example entity RP _e ^j-1 is ex = <Hanshin>, the positive example attribute RP _a ^j-1 is ey = <Pitcher>, and the first positive example text is “Hanshin / Ha / Pitcher / Team”. Assuming that “/ is good”, an example of the extracted feature fP _a ^j is as follows. Here, it is assumed that the feature extraction range is within two words before and after the entity and attribute.
Surface features: "ex + 1 =" is """ex + 2 = ey""ey-2 = ex", "ey-1 =" is "", "ey + 1 =" camp "", "ey + 2 = "is""
Part-of-speech features: “ex + 1 = particle”, “ey−1 = particle”, “ey + 1 = noun”, “ey + 1 = particle”
Proper noun features: “ex = ORG (organization name)” “ey-2 = ORG (organization name)”
Syntactic feature: “ex hierarchy = ey hierarchy” (both related to “good”)

Similarly, the attribute identifying feature extraction unit 23a extracts the negative example entity RN _e ^{j-1 of} PN ₁ (RN _e ^j-1 , RN _a ^j-1 ) from the text data set D stored in the storage unit 11a. And “first negative example text” which is a character string including a pair of the negative example attribute RN _a ^j−1 . Examples of the first negative example text are sentences, phrases, word strings, and the like included in the text data. One or more first negative example texts are extracted for a set of first negative example entity-negative example attribute pair PN ₁ (RN _e ^j−1 , RN _a ^j−1 ) and text data.
The attribute identifying feature extracting unit 23a uses the first negative example entity-negative example attribute pair PN ₁ (RN _e ^j−1 , RN _a ^j−1 ) for the first negative example text as the first negative example. The feature fN _a ^j of the example entity-negative example attribute pair PN ₁ (RN _e ^j−1 , RN _a ^j−1 ) is assumed. In this example, the feature fN _a ^j of PN ₁ (RN _e ^j−1 , RN _a ^j−1 ) is extracted for each first negative example text. An example of _a feature fN _a ^j of PN ₁ (RN _e ^j-1 , RN _a ^j-1 ) is a character including the first negative example text (negative example entity RN _e ^j-1 and negative example attribute RN _a ^j-1 Information included in the text data) and the first negative example entity RN _e ^j-1 and the first negative example attribute RN _a ^j-1 . The specific example is the same as the case of the feature fP _a ^j of PP ₁ (RP _e ^j−1 , RP _a ^j−1 ) corresponding to the positive example described above.

The attribute identifying feature extraction unit 23a sets a pair (fP _a ^j , <+1) of a feature fP _a ^{j of} PP ₁ (RP _e ^j−1 , RP _a ^j−1 ) and a label <+1> representing a positive example. >) And a pair (fN _a ^j , <-1>) of the feature fN _a ^{j of} PN ₁ (RN _e ^j-1 , RN _a ^j-1 ) and the label <-1> representing a negative example To do.
FIG. 8A is a diagram illustrating a pair (fP _a ^j , <+1>) and a pair (fN _a ^j , <−1>) output by the attribute identifying feature extraction unit 23a. In this example, the notation of two words before and after the entity (ex) and the attribute (ey) is used as a feature.

<< Attribute Identification Learning: Step S23 >>
A pair (fP _a ^j , <+1>) of _a feature fP _a ^{j of} PP ₁ (RP _e ^j-1 , RP _a ^j-1 ) and a label <+1> representing a positive example, and PN ₁ (RN A pair (fN _a ^j , <-1>) of _a feature fN _a ^{j of} _e ^j−1 , RN _a ^j−1 ) and a label <−1> representing a negative example is input to the attribute identification learning unit 25a. The attribute identification learning unit 25a uses the feature fP _a ^{j of} PP ₁ (RP _e ^j−1 , RP _a ^j−1 ) and the feature fN _a ^{j of} PN ₁ (RN _e ^j−1 , RN _a ^j−1 ). A first identification model ME _a ^j is generated by learning processing using supervised learning data. The first identification model ME _a ^j is input with the feature of an entity-attribute pair that is a set of an entity that is an arbitrary character string and an attribute of the entity, and the pair is a positive entity-positive example attribute pair or a negative example. This is a function for outputting information for identifying an entity-negative example attribute pair. Any model may be used as long as it is such an identification model ME _e ^j . For example, the first identification model ME _a ^j may be generated in the same manner as the above-described identification model ME _e ^j .
The first identification model ME _a ^j generated by the learning process is stored in the storage unit 21d. For example, parameters of the first identification model ME _a ^j generated by the learning processing is stored in the storage unit 21d.

<< Attribute Identification: Step S24 >>
The attribute identification unit 26a selects any text data from the set D of text data stored in the storage unit 11a, and selects a character string included in the selected text data as the first target entity RD _e ^j . The attribute identifying unit 26a selects a character string different from the first target entity RD _e ^j from the selected text data as the first target attribute RD _a ^j . Then, the attribute identifying unit 26a sets a pair of the first target entity RD _e ^j and the first target attribute RD _a ^j as the first target entity-target attribute pair PD ₁ (RD _e ^j , RD _a ^j ).

  Note that all text data may be selected from the text data set D, but it is not preferable in terms of computational efficiency to target all text data. For this reason, it is desirable to select a target by a specific method. Specific examples are shown below.

[Example of selection method]
First condition:
The attribute identifying unit 26a includes any positive example entity RP ^j-1 _e or negative example entity RN ^j-1 _e , and within an arbitrary window size from the entity RP ^j-1 _e or RN ^j-1 _e ( Here, text data including nouns is selected, and nouns within the window size are set as attribute candidates.

Second condition:
Since the number of objects may become enormous under the first condition alone, the attribute identification unit 26a uses the PP ₁ (RP _e ^j−1 , RP _a ^j−) used as supervised learning data in the attribute identification learning unit 25a. of feature fP _a ^j and ^{_{PN 1 (RN e j-1}} , RN a j-1) of a feature fN _a ^{j 1),} first identification model ME magnitude of the degree of influence on _a ^j generated from them feature that index representing (e.g. aforementioned weight lambda _q) satisfies certain criteria, i.e., influence of the the first identification model ME _a ^j selects a large feature fP _a ^j and / or fN _a ^j. For example, the attribute identifying unit 26a selects _a feature fP _a ^j and / or fN _a ^{j in} which the absolute value of the weight λ _q is larger than a threshold value.

The attribute identifying unit 26a selects text data including a character string corresponding to the selected feature fP _a ^j and / or fN _a ^j from the set of text data selected under the first condition. The attribute identifying unit 26a sets the character string included in the selected text data as the first target entity RD _e ^j and the first target attribute RD _a ^j . For example, the attribute identifying unit 26a extracts a surface feature word from the selected feature fP _a ^j and / or fN _a ^j , and sets the text data including the surface feature word as a set of text data selected under the first condition. The character string included in the selected text data is set as the first target entity RD _e ^j and the first target attribute RD _a ^j .

  For example, the selected feature was a feature FNC (x-2 = “POS: noun”, x−1 = “VS”) in which the two words before entity ex consisted of a combination of surface features and part-of-speech features. In this case, the attribute identification unit 26a extracts the surface feature word “VS” from the selected feature FNC (x−2 = “POS: noun”, x−1 = “VS”), and is selected under the first condition. Select text data including the word “VS” from the set of text data (end of description of [example of selection method]).

The attribute identifying feature extraction unit 23a is a character string that includes a set of the first target entity RD _e ^j and the first target attribute RD _a ^j from the text data set D stored in the storage unit 11a. Select Target text. Examples of the first target text are sentences, phrases, word strings, and the like included in the text data. One or more first target texts are extracted for a set of first target entity-target attribute pair PD ₁ (RD _e ^j , RD _a ^j ) and text data.

Attribute identification feature extracting unit 23a, the first target entity for a first target text - target attribute pair ^{_{PD 1 (RD e j, RD}} a j) the first target entity information indicating features of the - target attribute pair PD ₁ The feature fD _a ^{j of} (RD _e ^j , RD _a ^j ) is assumed. In this example, the feature fD _a ^j of PD ₁ (RD _e ^j , RD _a ^j ) is extracted for each first target text. An example of the feature fD _a ^j of PD ₁ (RD _e ^j , RD _a ^j ) is a text including a first target text (a first target entity RD _e ^j and a first target attribute RD _a ^j−1). Information included in the data), the first target entity RD _e ^j, and the first target attribute RD _a ^j-1 . The specific example is the same as the case of the feature fP _a ^j of PP ₁ (RP _e ^j−1 , RP _a ^j−1 ) corresponding to the positive example described above.

The feature fD _a ^j of PD ₁ (RD _e ^j , RD _a ^j ) corresponding to the first target text is input to the attribute identifying unit 26a. The attribute identification unit 26a inputs the feature fD _a ^j of PD ₁ (RD _e ^j , RD _a ^j ) to the first identification model ME _a ^j read from the storage unit 21d, and PD ₁ (RD _e ^j , RD _a ^j ) Identifies a positive entity-positive example attribute pair or a negative example entity-negative example attribute pair.
Here, when the attribute identifying unit 26a identifies PD ₁ (RD _e ^j , RD _a ^j ) as a positive entity-positive example attribute pair, the attribute identifying unit 26a determines the _first PD ₁ (RD _e ^j , RD _a ^j ). 1 target attribute RD _a ^j stored in the storage unit 21e as positive examples attribute RP _a ^j, is added to the set of positive examples attribute RP _a ^j. Further, when the attribute identifying unit 26a identifies that PD ₁ (RD _e ^j , RD _a ^j ) is a negative entity-negative example attribute pair, the attribute identifying unit 26a determines the _first PD ₁ (RD _e ^j , RD _a ^j ). stored in the storage unit 21e target attribute RD _a ^j as a negative example attribute RN _a ^j, is added to the set of negative examples attribute RN _a ^j. That is, in steps S22 to S24, the positive example and negative example entities are not updated, and only the positive example and negative example attributes are updated.

For example, the attribute identification unit 26a selects the text data of T10 from the text data set D in FIG. 4, the word <Hiroshima> included in the text data is set as the first target entity RD _e ^j , and the word <war> is the first. It is assumed that one target attribute RD _a ^j is selected. In this case, for example, the attribute identifying feature extraction unit 23a sets the text data of T10 including the set of <Hiroshima> and <war> as the first target text, and PD ₁ (RD _e ^j , RD for the text data of T10 _a ^j ) The feature fD _a ^j of “<Hiroshima> − <war>” is extracted. The attribute identification unit 26a inputs the feature fD _a ^j of PD ₁ (RD _e ^j , RD _a ^j ) = "<Hiroshima>-<war >> into the first identification model ME _a ^j and PD ₁ (RD _e ^j , RD _a ^j ) is identified as a positive entity-positive example attribute pair or a negative example entity-negative example attribute pair. For example, if “<Hiroshima>-<Battle>” is identified as a positive entity-positive case attribute pair, an attribute <Battle> is added to the set of positive case attributes RP _a ^j . Of the PDs ₁ (RD _e ^j , RD _a ^j ) identified as positive examples or negative examples, only the first target attribute RD _a ^j to which the reliability exceeding the threshold is given is used as the positive example attribute RP _a. ^It may be added to the set of ^j or negative example attributes RN _a ^j . In the above example {<VS>, <first leg>, <pitcher>, <War>} are added to the set of positive examples attribute RP ^j _a.

<< Entity Identification Feature Extraction: Step S25 >>
The set of positive entity RP _e ^j-1, the set of negative example entity RN _e ^j-1 , the set of positive example attributes RP _a ^{j and} the set of negative example attributes RN _a ^j updated as described above are for entity identification. It is input to the feature extraction unit 23b.
The entity identifying feature extraction unit 23b is a set of a second positive example entity selected from the set of positive example entities RP _e ^j-1 and a second positive example attribute selected from the set of positive example attributes RP _a ^j . Two positive example entities—positive example attribute pair PP ₂ (RP _e ^j−1 , RP _a ^j ) and a second negative example entity selected from the set of negative example entities RN _e ^j−1 and negative example attribute RN _a ^j A second negative example entity-negative example attribute pair PN ₂ (RN _e ^j−1 , RN _a ^j ) that is a pair with the second negative example attribute selected from the set is generated. PP ₂ (RP _e ^j-1 , RP _a ^j ) and PN ₂ (RN _e ^j-1 , RN _a ^j ) are taken from RP _e ^j-1 , RP _a ^j , RN _e ^j-1, and RN _a ^j It may be generated for all possible combinations, or only some of those combinations.
Next, the entity identifying feature extraction unit 23b extracts the positive example entity RP _e ^j-1 and the positive example of PP ₂ (RP _e ^j-1 , RP _a ^j ) from the text data set D stored in the storage unit 11a. The “second positive example text” that is a character string including a pair with the attribute RP _a ^j is selected. Examples of the second positive example text are sentences, phrases, word strings, and the like included in the text data. One or more second positive example texts are extracted for a set of the second positive example entity-positive example attribute pair PP ₂ (RP _e ^j−1 , RP _a ^j ) and text data.
The entity identifying feature extraction unit 23b obtains information representing the characteristics of the second positive example entity-positive example attribute pair PP ₂ (RP _e ^j−1 , RP _a ^j ) with respect to the second positive example text. A feature fP _e ^j of the positive attribute pair PP ₂ (RP _e ^j−1 , RP _a ^j ). In this example, the feature fP _e ^j of PP ₂ (RP _e ^j−1 , RP _a ^j ) is extracted for each second positive example text. An example of the feature fP _e ^j of PP ₂ (RP _e ^j−1 , RP _a ^j ) is a character string that includes the second positive example text (the positive example entity RP _e ^j−1 and the positive example attribute RP _a ^j). Information included in the text data) and the second positive example entity RP _e ^j-1 and the second positive example attribute RP _a ^j . The specific example is the same as the case of the feature fP _a ^j of PP ₁ (RP _e ^j−1 , RP _a ^j−1 ) described above (step S22).

Similarly, the entity identifying feature extraction unit 23b reads from the text data set D stored in the storage unit 11a a character including a set of any negative example entity RN _e ^j-1 and negative example attribute RN _a ^j. Select the column "second negative example text". Examples of the second negative example text are sentences, phrases, word strings, and the like included in the text data. One or more second negative example texts are extracted for a pair of second negative example entity-negative example attribute pair PN ₂ (RN _e ^j−1 , RN _a ^j ) and text data.
The entity identifying feature extraction unit 23b obtains information representing the characteristics of the second negative example entity-negative example attribute pair PN ₂ (RN _e ^j−1 , RN _a ^j ) with respect to the second negative example text. A feature fN _e ^j of the negative example attribute pair PN ₂ (RN _e ^j−1 , RN _a ^j ). In this example, the feature fN _e ^j of PN ₂ (RN _e ^j−1 , RN _a ^j ) is extracted for each second negative example text. An example of a feature fN _e ^j of PN ₂ (RN _e ^j−1 , RN _a ^j ) is a character string including _a second negative example text (negative example entity RN _e ^j−1 and negative example attribute RN _a ^j) Information included in the text data) and the second negative example entity RN _e ^j-1 and the second negative example attribute RN _a ^j . The specific example is the same as the case of the feature fP _a ^j of PP ₁ (RP _e ^j−1 , RP _a ^j−1 ) described above (step S22).

The entity identifying feature extraction unit 23b sets (fP _e ^j , <+1>) a feature fP _e ^{j of} PP ₂ (RP _e ^j−1 , RP _a ^j ) and a label <+1> representing a positive example. And a pair (fN _e ^j , <-1>) of a feature fN _e ^{j of} PN ₂ (RN _e ^j−1 , RN _a ^j ) and a label <−1> representing a negative example.

FIG. 8B is a diagram illustrating a pair (fP _e ^j , <+1>) and a pair (fN _e ^j , <-1>) output by the entity identifying feature extraction unit 23b. In this example, the notation of two words before and after the entity (ex) and the attribute (ey) is used as a feature.
<< Entity Identification Learning: Step S26 >>
A pair (fP _e ^j , <+1>) of a feature fP _e ^{j of} PP ₂ (RP _e ^j−1 , RP _a ^j ) and a label <+1> representing a positive example, and PN ₂ (RN _e ^{j −1} , RN _a ^j ) of the feature fN _e ^j and a negative example label <-1> (fN _e ^j , <-1>) is input to the entity identification learning unit 25b. The entity identification learning unit 25b uses the feature fP _e ^{j of} PP ₂ (RP _e ^j−1 , RP _a ^j ) and the feature fN _e ^{j of} PN ₂ (RN _e ^j−1 , RN _a ^j ) as supervised learning data. Through the learning process described above, the second identification model ME _e ^j is generated. This two-discriminating model ME _e ^j receives the identity of an entity-attribute pair that is a set of an entity that is an arbitrary character string and an attribute of the entity, and the pair is a positive entity-positive attribute pair or a negative entity. A function that outputs information for identifying a negative example attribute pair. Any model may be used as long as it is such a second identification model ME _e ^j . For example, the second identification model ME _e ^j may be generated in the same manner as the above-described identification model ME _e ^j .
Second identification model ME _e ^j generated by the learning processing is stored in the storage unit 11d. For example, the parameters of the second identification model ME _e ^j generated by the learning processing is stored in the storage unit 11d.

<< Entity Identification: Step S27 >>
Entity identification unit 26b selects one of the text data from the set D of text data stored in the storage unit 11a, selects a character string included in the text data selected as a second target entity RD _e ^j. The entity identifying unit 26b selects a character string different from the second target entity RD _e ^j from the selected text data as the second target attribute RD _a ^j . Then, the entity identification unit 26b sets the pair of the second target entity RD _e ^j and the second target attribute RD _a ^j as the second target entity-target attribute pair PD ₂ (RD _e ^j , RD _a ^j ).

  Note that all text data may be selected from the text data set D, but it is not preferable in terms of computational efficiency to target all text data. For this reason, it is desirable to select a target by a specific method. Specific examples are shown below.

[Example of selection method]
First condition:
Entity identification unit 26b includes one positive cases attribute RP ^j _a and negative examples attribute RN ^j _a, and to the attribute RP ^j _a The RN ^j _a from within any window size (here, 3 words) Text data including nouns is extracted, and nouns within the window size are used as entity candidates.

Second condition:
Since the number of targets may become enormous under the first condition alone, the entity identification unit 26b uses PP ₂ (RP _e ^j−1 , RP _a ^j ) used as supervised learning data in the entity identification learning unit 25b. Of the features fP _e ^j and PN ₂ (RN _e ^j−1 , RN _a ^j ) of the features fN _e ^j of the, and an index indicating the degree of influence on the second identification model ME _e ^j generated from them ( for example feature the aforementioned weight lambda _q) satisfies certain criteria, i.e., influence of the to the second identification model ME _e ^j selects a large feature fP _e ^j and / or fN _e ^j. For example, the entity identifying unit 26b selects a feature fP _e ^j and / or fN _e ^{j in} which the absolute value of the weight λ _q is larger than a threshold value.

The entity identification unit 26b selects text data including a character string corresponding to the selected feature fP _e ^j and / or fN _e ^j from the set of text data selected under the first condition. The entity identifying unit 26b sets the character string included in the selected text data as the second target entity RD _e ^j and the second target attribute RD _a ^j . For example, the entity identification unit 26b extracts a surface feature word from the selected feature fP _e ^j and / or fN _e ^j , and sets text data including the surface feature word as a set of text data selected under the first condition. Select from. (End of description of [Example of selection method]).

The entity identifying feature extraction unit 23b is a character string that includes a set of the second target entity RD _e ^j and the second target attribute RD _a ^j from the text data set D stored in the storage unit 11a. Select Target text. Examples of the second target text are sentences, phrases, word strings, and the like included in the text data. One or more second target texts are extracted for a set of the second target entity-target attribute pair PD ₂ (RD _e ^j , RD _a ^j ) and text data.

Entity identification feature extracting unit 23b, the second target entity for a second target text - target attribute pair ^{_{PD 2 (RD e j, RD}} a j) said second target entity information indicating features of the - target attribute pair PD ₂ A feature fD _e ^{j of} (RD _e ^j , RD _a ^j ) is assumed. In this example, the feature fD _e ^j of PD ₂ (RD _e ^j , RD _a ^j ) is extracted for each second target text. An example of the feature fD _e ^j of PD ₂ (RD _e ^j , RD _a ^j ) is a text that includes the second target text (second target entity RD _e ^j and second target attribute RD _a ^j-1). Information included in the data), the second target entity RD _e ^j, and the second target attribute RD _a ^j-1 . The specific example is the same as the case of the feature fP _a ^j of PP ₁ (RP _e ^j−1 , RP _a ^j−1 ) described above (step S22).

The feature fD _e ^j of PD ₂ (RD _e ^j , RD _a ^j ) corresponding to the second target text is input to the entity identification unit 26b. The entity identification unit 26b inputs the feature fD _e ^j of PD ₂ (RD _e ^j , RD _a ^j ) into the second identification model ME _e ^j read from the storage unit 11d, and PD ₂ (RD _e ^j , RD _a ^j ) Identifies a positive entity-positive example attribute pair or a negative example entity-negative example attribute pair.

Here, when the entity identifying unit 26b identifies PD ₂ (RD _e ^j , RD _a ^j ) as a positive example entity-positive example attribute pair, the entity identifying unit 26b determines the _second PD ₂ (RD _e ^j , RD _a ^j ). 2 target entities RD _e ^j stored in the storage unit 11e as positive examples entity RP _e ^j, to add to the set of positive examples entity RP _e ^j. Further, when the entity identifying unit 26b identifies that PD ₂ (RD _e ^j , RD _a ^j ) is a negative example entity-negative example attribute pair, the entity identifying unit 26b sets the _second PD ₂ (RD _e ^j , RD _a ^j ). stored in the storage unit 11e target entity RD _e ^j as a negative example entity RN _e ^j, to add to the set of negative examples entity RN _e ^j. That is, in steps S25-S27, the positive example and negative example attributes are not updated, and only the positive example and negative example entities are updated.

<< Convergence determination: steps S17 to S19 >>
The convergence determination unit 17 determines whether the convergence condition is satisfied as in the first embodiment (step S17).
When the convergence determination unit 17 determines that the convergence condition is satisfied, the iterations from step S22 to S27 are finished, and the output unit 18 outputs all the positive example entities RP ^j _e stored in the storage unit 11e for processing. The process ends (step S19). In other cases, the control unit 19 sets j + 1 as a new value of j (step S18), the positive example entity RP ^j _e and the negative example entity RN ^j _e stored in the storage unit 11e, and the storage unit 21e. enter a positive example attribute RP ^j _a and negative cases attribute RN ^j _a is stored in the attribute identifying feature extraction unit 23a, iterations of steps S22 S27 is executed.

<Features of Second Embodiment>
As described above, in the method according to the present embodiment, identification is performed using a pair of an entity and its attribute, so that semantic drift can be suppressed. For example, the entity <Hanshin> has ambiguity, and the entity <Hanshin> alone cannot identify whether the entity <Hanshin> indicates a railroad name or a team name. However, if <Hanshin>-<Game> or <Hanshin>-<Crew> with attributes of <Game> or <Crew> are used as constraints, each <Hanshin> is used in a different meaning. Can be identified.

  In this embodiment, since the co-training method is used, highly accurate identification is possible. In the above description, the example in which the positive example and the negative example entity are updated (S25 to S27) after the positive example and negative example attributes are updated (steps S22 to S24) is shown. However, the positive example and negative example attributes may be updated after the positive example and negative example entities are updated.

  Note that espresso is known as a relation extraction technique that handles entity-attribute pairs (Reference 4 “Patrick Pantel and Marco Pennacchiotti.,“ Espresso: Leveraging Generic Patterns for Automatically Harvesting Semantic Relations. ”, COLING-ACL, 2006. "). In the case of espresso, since the purpose is to acquire an entity-attribute pair, it is necessary to give entity-attribute pairs as positive examples and negative examples in advance. On the other hand, since this embodiment uses an attribute for acquiring an entity, only an entity need be given as an initial value.

  In addition, espresso consists of an iteration that calculates the reliability of an entity / attribute pair and an iteration of a reliability calculation for the feature, whereas in this form it is an iteration that calculates the reliability of the entity and the reliability of the attribute. It consists of iterations that calculate Now all we want is an entity, and attribute information is only acquired additionally. In other words, it is not necessary to have high completeness of attributes, and it is sufficient to use only a sufficient amount of attributes that are sufficiently reliable and can suppress the semantic drift. For the purpose of this embodiment, it can be said that the method of this embodiment that can directly evaluate the reliability of each entity / attribute is more appropriate than the reliability obtained in pairs as in espresso.

In addition, if you try to acquire 100 new entity-attribute pairs with espresso, you cannot control how many new entities and new attributes are included in each. For example, an inappropriate situation such as 1 entity × 100 attributes may occur. In the method of this embodiment, the iteration for calculating the reliability of the entity and the iteration for calculating the reliability of the attribute are executed separately, so that the number of entities and the number of attributes can be freely controlled separately. In the present embodiment, it is possible to finely control, for example, 100 entities and 10 attributes.
In addition, as in the first embodiment, the method of this embodiment can be used regardless of the type of text data that is a resource, and has a wide range of applications.

[Third Embodiment]
The third embodiment is a combination of the first embodiment and the second embodiment. In other words, learning of the identification model and identification by the identification model are performed using both topic information and attributes. Below, it demonstrates centering around difference with 1st and 2nd embodiment. In addition, the same reference numerals as those in the first and second embodiments are used for portions common to the first and second embodiments.

<Configuration>
FIG. 9 is a block diagram for illustrating a functional configuration of the data extraction device 3 of the third embodiment.
As illustrated in FIG. 9, the data extraction device 3 includes storage units 11a-11e, 21d, and 21e, an initial attribute set generation unit 22, an attribute identification feature extraction unit 23a, an entity identification feature extraction unit 23b, and topic information extraction. Units 34 a and 34 b, attribute identification learning unit 35 a, entity identification learning unit 35 b, attribute identification unit 36 a, entity identification unit 36 b, convergence determination unit 17, output unit 18, and control unit 19. Originally, each process is executed. The data extraction device 3 is a special device configured by, for example, reading a special program into a known or dedicated computer.

<Pre-processing>
This is the same as in the first embodiment.
<Data extraction process>
FIG. 10 is a diagram for illustrating data extraction processing of the data extraction device 3 according to the third embodiment.
First, the same processing as steps S11, S12, S21, and S22 of the first and second embodiments is executed.

<< Topic Information Extraction: Step S321 >>
A pair (fP _a ^j , <+1>) of the feature fP _a ^{j of} PP ₁ (RP _e ^j-1 , RP _a ^j-1 ) generated in step S22 and a label <+1> representing a positive example, The topic information extraction unit is _a set (fN _a ^j , <-1>) of the feature fN _a ^{j of} PN ₁ (RN _e ^j-1 , RN _a ^j-1 ) and the label <-1> representing a negative example 34a. In order to avoid confusion, these are expressed as a pair (fP '' _a ^j , <+1>) and a pair (fN '' _a ^j , <-1>) below.

The topic information extraction unit 34a performs first processing corresponding to the text data including the set of the first positive example entity RP _e ^j-1 and the first positive example attribute RP _a ^j-1 by the same process as in step S14 described above. Extract positive topic information. The topic information extraction unit 34a applies the first positive example topic to the feature fP ″ _a ^j of PP ₁ (RP _e ^j−1 , RP _a ^j−1 ) corresponding to each first positive example text included in the text data. The information added is set as a new feature fP _a ^{j of} each PP ₁ (RP _e ^j−1 , RP _a ^j−1 ) corresponding to each first positive example text. That is, the feature fP _a ^j of the PP ₁ (RP _e ^j−1 , RP _a ^j−1 ) generated by the topic information extraction unit 34a is selected from the set D ′ of text data with topic information. It includes topic information included in text data with topic information including text data including a set of _a positive example entity RP _e ^j-1 and a first positive example attribute RP _a ^j-1 .

Similarly, the topic information extraction unit 34a performs processing similar to that in step S14 described above, and performs processing corresponding to text data including a set of the negative example entity RN _e ^j-1 and the first negative example attribute RN _a ^j-1 . 1 Negative topic information is extracted. The topic information extraction unit 34a applies the first negative example topic to the feature fN ″ _a ^j of PN ₁ (RN _e ^j−1 , RN _a ^j−1 ) corresponding to each first negative example text included in the text data. The information added is set as a new feature fN _a ^{j of} each PN ₁ (RN _e ^j−1 , RN _a ^j−1 ) corresponding to each first negative example text. That is, the feature fN _a ^j of the PN ₁ (RN _e ^j−1 , RN _a ^j−1 ) generated by the topic information extraction unit 34 _a is selected from the set D ′ of text data with topic information. It includes topic information included in text data with topic information including text data including a set of _a negative example entity RN _e ^j-1 and a first negative example attribute RN _a ^j-1 .
The topic information extraction unit 34a sets a pair (fP _a ^j , <+1) of the feature fP _a ^j of the generated PP ₁ (RP _e ^j−1 , RP _a ^j−1 ) and a label <+1> representing a positive example. >) And a pair (fN _a ^j , <-1>) of the feature fN _a ^{j of} PN ₁ (RN _e ^j-1 , RN _a ^j-1 ) and the label <-1> representing a negative example To do.

<< Attribute Identification Learning: Step S33 >>
A pair (fP _a ^j , <+1>) of _a feature fP _a ^{j of} PP ₁ (RP _e ^j-1 , RP _a ^j-1 ) and a label <+1> representing a positive example, and PN ₁ (RN A set (fN _a ^j , <-1>) of the feature fN _a ^{j of} _e ^j−1 , RN _a ^j−1 ) and a label <−1> representing a negative example is input to the attribute identification learning unit 35a. The attribute identification learning unit 35a uses these as supervised learning data, generates the first identification model ME _a ^j as in step S23 described above, and stores it in the storage unit 21d.

<< Attribute Identification: Step S34 >>
The attribute identifying unit 36a first generates _a feature fD _a ^j of PD ₁ (RD _e ^j , RD _a ^j ) corresponding to the first target text as in step S24. Hereinafter, in order to avoid confusion, the feature fD _a ^j of PD ₁ (RD _e ^j , RD _a ^j ) corresponding to the first target text created in the same manner as in step S24 is denoted as fD ″ _a ^j . Next, the attribute identifying unit 36a extracts first target topic information corresponding to text data including a set of the target entity RD _e ^j and the first target attribute RD _a ^j , as in step S15. The attribute identification unit 36a is obtained by adding the first target topic information to the feature fD ″ _a ^j of each PD ₁ (RD _e ^j , RD _a ^j ) corresponding to each first target text included in the text data. , The feature fD _e ^{j of} each PD ₁ (RD _e ^j , RD _a ^j ) corresponding to each first target text. That is, the feature fD _e ^j of PD ₁ (RD _e ^j , RD _a ^j ) generated by the attribute identifying unit 36a is the first target entity RD _e ^j selected from the set D ′ of text data with topic information. It includes topic information included in text data with topic information including text data including a set with the first target attribute RD _a ^j (see, for example, FIG. 5A).

Similarly to step S24, the attribute identification unit 36a inputs the feature fD _a ^j of PD ₁ (RD _e ^j , RD _a ^j ) to the first identification model ME _a ^j read from the storage unit 21d, and PD ₁ (RD _e ^j , RD _a ^j ) identifies whether it is a positive entity-positive example attribute pair or a negative example entity-negative example attribute pair. When the attribute identifying unit 36a identifies PD ₁ (RD _e ^j , RD _a ^j ) as a positive entity-positive example attribute pair, the first target attribute of the PD ₁ (RD _e ^j , RD _a ^j ) the RD _a ^j stored in the storage unit 21e as positive examples attribute RP _a ^j, is added to the set of positive examples attribute RP _a ^j. Further, when the attribute identifying unit 36a identifies that PD ₁ (RD _e ^j , RD _a ^j ) is a negative entity-negative example attribute pair, the attribute identifying unit 36a determines the _first PD ₁ (RD _e ^j , RD _a ^j ). stored in the storage unit 21e target attribute RD _a ^j as a negative example attribute RN _a ^j, is added to the set of negative examples attribute RN _a ^j.

<< Topic Information Extraction: Step S351 >>
Next, the process of step S25 described above is executed, and a pair of the feature fP _e ^{j of} PP ₂ (RP _e ^j−1 , RP _a ^j ) obtained thereby and a label <+1> representing a positive example ( fP _e ^j , <+1>) and a pair of features fN _e ^{j of} PN ₂ (RN _e ^j-1 , RN _a ^j ) and a label <-1> representing a negative example (fN _e ^j , <- 1>) is input to the topic information extraction unit 34b. In order to avoid confusion, these are expressed as a pair (fP '' _e ^j , <+1>) and a pair (fN '' _e ^j , <-1>) below.

The topic information extraction unit 34b performs a process similar to that in step S14 described above, and corresponds to a set of text data including a set of the second positive example entity RP _e ^j-1 and the second positive example attribute RP _a ^j . 2. Extract the example topic information. The topic information extraction unit 34b adds the second positive example topic information to the feature fP ″ _e ^j of PP ₂ (RP _e ^j−1 , RP _a ^j ) corresponding to each second positive example text included in the text data. The addition is set as a new feature fP _e ^{j of} each PP ₂ (RP _e ^j−1 , RP _a ^j ) corresponding to each second example text. That is, the feature fP _e ^j of the PP ₂ (RP _e ^j−1 , RP _a ^j ) generated by the topic information extraction unit 34b is selected from the set D ′ of text data with topic information. The topic information included in the text data with topic information including the text data including the set of the entity RP _e ^j-1 and the second positive example attribute RP _a ^j is included (see, for example, FIG. 5A).

Similarly, the topic information extraction unit 34b performs a process similar to that in step S14 described above, and performs processing corresponding to text data including a set of the second negative example entity RN _e ^j-1 and the second negative example attribute RN _a ^j . 2 Negative example topic information is extracted. The topic information extraction unit 34b adds the second negative example topic information to the feature fN ″ _e ^j of PN ₂ (RN _e ^j−1 , RN _a ^j ) corresponding to each second negative example text included in the text data. The addition is set as a new feature fN _e ^{j of} each PN ₂ (RN _e ^j−1 , RN _a ^j ) corresponding to each second negative example text. That is, the feature fN _e ^j of the PN ₂ (RN _e ^j−1 , RN _a ^j ) generated by the topic information extraction unit 34b is selected from the set D ′ of text data with topic information. It includes topic information included in text data with topic information including text data including a set of an entity RN _e ^j-1 and a second negative example attribute RN _a ^j .

The topic information extraction unit 34b generates a set (fP _e ^j , <+1>) of the feature fP _e ^j of the generated PP ₂ (RP _e ^j−1 , RP _a ^j ) and a label <+1> representing a positive example. And a pair (fN _e ^j , <-1>) of a feature fN _e ^{j of} PN ₂ (RN _e ^j−1 , RN _a ^j ) and a label <−1> representing a negative example.

<< Entity Identification Learning: Step S36 >>
A pair (fP _e ^j , <+1>) of a feature fP _e ^{j of} PP ₂ (RP _e ^j−1 , RP _a ^j ) and a label <+1> representing a positive example, and PN ₂ (RN _e ^{j −1} , RN _a ^j ) of the feature fN _e ^j and a negative example label <-1> (fN _e ^j , <-1>) is input to the entity identification learning unit 35b. Entity identification learning unit 35b is these and supervised learning data to generate second identification model ME _e ^j similarly to step S26 described above, and stores in the storage unit 11d.

<< Entity Identification: Step S37 >>
The entity identification unit 36b first generates a feature fD _e ^j of PD ₂ (RD _e ^j , RD _a ^j ) corresponding to the second target text, as in step S27. Hereinafter, in order to avoid confusion, the feature fD _e ^j of PD ₂ (RD _e ^j , RD _a ^j ) corresponding to the second target text created in the same manner as in step S27 is denoted as fD ″ _e ^j .

Next, the entity identifying unit 36b extracts second target topic information corresponding to text data including a set of the target entity RD _e ^j and the second target attribute RD _a ^j , as in step S15. Entity identification unit 36b, the ^{_{PD 2 (RD e j, RD}} a j) corresponding to each of the second target text to which the text data includes a plus the second target topic information on the identity fD '' _e ^j of And the feature fD _e ^{j of} each PD ₂ (RD _e ^j , RD _a ^j ) corresponding to each second target text. That is, the feature fD _e ^j of PD ₂ (RD _e ^j , RD _a ^j ) generated by the entity identification unit 36b is the second target entity RD _e ^j selected from the set D ′ of text data with topic information. The topic information included in the text data with topic information including the text data including the pair with the second target attribute RD _a ^j is included (see, for example, FIG. 5A).

As in step S27, the entity identification unit 36b inputs the feature fD _e ^j of PD ₂ (RD _e ^j , RD _a ^j ) to the second identification model ME _e ^j read from the storage unit 11d, and outputs PD ₂ (RD _e ^j , RD _a ^j ) identifies whether it is a positive entity-positive example attribute pair or a negative example entity-negative example attribute pair. When the entity identification unit 36b identifies PD ₂ (RD _e ^j , RD _a ^j ) as a positive entity-positive example attribute pair, the second target entity of the PD ₂ (RD _e ^j , RD _a ^j ) the RD _e ^j stored in the storage unit 11e as positive examples entity RP _e ^j, to add to the set of positive examples entity RP _e ^j. Further, when the entity identifying unit 36b identifies that PD ₂ (RD _e ^j , RD _a ^j ) is a negative example entity-negative example attribute pair, the entity identifying unit 36b determines the _second PD ₂ (RD _e ^j , RD _a ^j ). stored in the storage unit 11e target entity RD _e ^j as a negative example entity RN _e ^j, to add to the set of negative examples entity RN _e ^j.
Thereafter, the processing of steps S17 to S19 described above is executed.

[Other variations, etc.]
The present invention is not limited to the embodiment described above. For example, step S13 may not be executed in the first embodiment, and only topic information may be used as a feature. Further, the topic model and the learning model are not limited to the specific examples described above, as described above. In addition, the various processes described above are not only executed in time series according to the description, but may be executed in parallel or individually according to the processing capability of the apparatus that executes the processes or as necessary. Needless to say, other modifications are possible without departing from the spirit of the present invention.

Further, when the above-described configuration is realized by a computer, processing contents of functions that each device should have are described by a program. The processing functions are realized on the computer by executing the program on the computer.
The program describing the processing contents can be recorded on a computer-readable recording medium. The computer-readable recording medium is a non-transitory recording medium. Examples of such a recording medium are a magnetic recording device, an optical disk, a magneto-optical recording medium, a semiconductor memory, and the like.
The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.
A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, this computer reads the program stored in its own recording device and executes the process according to the read program. As another execution form of the program, the computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In this embodiment, the present apparatus is configured by executing a predetermined program on a computer. However, at least a part of these processing contents may be realized by hardware.

1-3 Data extraction device

Claims (16)

A pre-processing unit that learns topic information representing the appropriateness of a plurality of topic candidates for text data as an index value and a topic model that describes the relationship between the text data using unsupervised learning data obtained from the text data When,
The positive example topic information extracted from the topic models in response to the topic of the text data containing the extracted target positive cases entity is a character string and at least part of the identity of the positive examples entity, not subject to extraction string A topic information extraction unit that takes negative example topic information extracted from the topic model corresponding to a topic of text data including a negative example entity as at least part of the features of the negative example entity;
Information for identifying whether an entity is a positive example entity or a negative example entity by inputting a feature of an arbitrary entity by learning processing using the features of the positive example entity and the features of the negative example entity as supervised learning data An identification learning unit that generates an identification model that is a function that outputs
An entity that is a character string included in text data selected from a set of text data is set as a target entity, and topic information extracted from the topic model corresponding to a topic of the selected text data is at least one of the features of the target entity. When the identity of the target entity is input to the identification model to identify whether the target entity is a positive example entity or a negative example entity, and the target entity is identified as a positive example entity, An entity identification unit that sets the target entity as the negative example entity when the positive entity is identified and the target entity is identified as a negative entity;
A data extraction device. The data extraction device according to claim 1, comprising:
The identity of the positive example entity corresponds to a character string that includes the positive example entity and is included in text data that includes the positive example entity, and includes information indicating a relationship between the character string and the positive example entity. Including
The feature of the negative example entity corresponds to a character string that includes the negative example entity and is included in the text data that includes the negative example entity, and includes information indicating a relationship between the character string and the negative example entity. Including
The feature of the target entity corresponds to a character string that includes the target entity and is included in text data that includes the target entity, and includes information that represents a relationship between the character string and the target entity.
A data extraction apparatus characterized by that. The data extraction device according to claim 2, wherein
The character string included in the text data corresponds to the topic candidate of the character string included in the text data and the topic candidate score indicating the appropriateness of each of the candidate candidates for the character string included in the text data. The aggregation result for each candidate for the topic obtained by aggregating the topic candidate scores corresponding to the topic candidates is set as a positive example topic score for each of the topic candidates, and the positive example is selected from the topic candidates. If the topic score meets certain criteria, it is considered a positive reference topic,
The identification learning unit is a character string included in the text data corresponding to the same topic candidate as any one of the positive example reference topics, and the topic candidate score corresponding to the same topic candidate is specified. The feature of the positive example entity corresponding to the character string satisfying the criterion is the supervised learning data,
A data extraction apparatus characterized by that. The data extraction device according to claim 2 or 3,
The character string included in the text data corresponds to the topic candidate of the character string included in the text data and the topic candidate score indicating the appropriateness of each of the candidate candidates for the character string included in the text data. The aggregation result for each candidate for the topic obtained by aggregating the topic candidate scores corresponding to the topic candidates is set as a positive example topic score for each of the topic candidates, and the positive example is selected from the topic candidates. If the topic score meets certain criteria, it is considered a positive reference topic,
The entity identification unit is a character string included in the text data corresponding to the same topic candidate as any one of the positive reference topics, and the topic candidate score corresponding to the same topic candidate is specified. An entity included in a character string that satisfies the criteria is the target entity.
A data extraction apparatus characterized by that. The data extraction device according to any one of claims 1 to 4,
The entity identifier is
Among the features of the positive example entity and the negative example entity used as the supervised learning data in the identification learning unit, an index indicating the magnitude of the degree of influence on the identification model generated therefrom is specified An entity that is a character string included in the text data including the character string corresponding to the selected feature is selected as the target entity.
A data extraction apparatus characterized by that. A set of a first positive example entity selected from a set of positive example entities that are character strings to be extracted and a first positive example attribute selected from a set of positive example attributes that are character strings representing attributes of the positive example entities A first positive example entity-positive example attribute pair, a first negative example entity selected from a set of negative example entities that are not extracted, and a negative example attribute that is a character string representing an attribute of the negative example entity Generating a first negative example entity-negative example attribute pair that is a set with a first negative example attribute selected from the set of: and from the set of text data, the first positive example entity and the first positive example attribute; A character string including the set of the first positive entity entity-positive example attribute pair with respect to the selected character string, and information indicating characteristics of the first positive example entity-positive example attribute pair at least. A character string including a set of the first negative example entity and the first negative example attribute is selected from the set of text data, and the first negative example entity-negative example for the selected character string is selected. An attribute identifying feature extraction unit having information representing the characteristics of the attribute pair as at least part of the features of the first negative example entity-negative example attribute pair;
The learning process using the features of the first positive example entity-positive example attribute pair and the features of the first negative example entity-negative example attribute pair as supervised learning data, and the entity of an arbitrary character string and the entity A function that outputs the information for identifying whether the entity-attribute pair is a positive entity-positive example attribute pair or a negative example entity-negative example attribute pair, using the identity of the entity-attribute pair that is a pair with the attribute as an input. An attribute identification learning unit for generating a certain first identification model;
One of the text data is selected from the set of text data, a character string included in the selected text data is selected as a first target entity, and a character string different from the first target entity is selected from the selected text data. The first target attribute is selected as a first target entity and the first target attribute is set as a first target entity-target attribute pair, and the first target entity-target attribute pair in the selected text data is selected. The information representing the characteristics of the first target entity-target attribute pair is at least part of the feature, and the first target entity-target attribute pair feature is input to the first identification model, and the first target entity- Identify whether the target attribute pair is a positive entity-positive example attribute pair or a negative example entity-negative example attribute pair, When the first target entity-target attribute pair is identified as a positive entity-positive attribute pair, the first target attribute is added to the set of positive attribute, and the first target entity-target attribute An attribute identifying unit that adds the first target attribute to the set of negative example attributes when the pair is identified as a negative example entity-negative example attribute pair;
A second positive example entity-positive example attribute pair that is a set of a second positive example entity selected from the set of positive example entities and a second positive example attribute selected from the set of positive example attributes; and the negative example entity Generating a second negative example entity-negative example attribute pair which is a set of a second negative example entity selected from the set of the negative example attributes and a second negative example attribute selected from the set of the negative example attributes; To select a character string including a set of the second positive example entity and the second positive example attribute, and information indicating characteristics of the second positive example entity-positive example attribute pair for the selected character string A character string including a pair of the second negative example entity and the second negative example attribute is selected and selected from at least a part of the feature of the second positive example entity-positive example attribute pair. Shi And entities identifying feature extracting section for at least part of the identity of the negative examples attribute pair, - negative sample attribute pair information the second negative example entities that represent characteristics of - the second negative example entity for the character string
The learning process using the features of the second positive example entity-positive example attribute pair and the features of the second negative example entity-negative example attribute pair as supervised learning data, and the entity of an arbitrary character string and the entity A function that outputs the information for identifying whether the entity-attribute pair is a positive entity-positive example attribute pair or a negative example entity-negative example attribute pair, using the identity of the entity-attribute pair that is a pair with the attribute as an input. An entity identification learning unit for generating a second identification model;
One of the text data is selected from the set of text data, a character string included in the selected text data is selected as a second target entity, and a character string different from the second target entity is selected from the selected text data. The second target attribute is selected as a second target entity, and a pair of the second target entity and the second target attribute is set as a second target entity-target attribute pair, and the second target entity-target attribute pair in the selected text data is selected. Information representing the characteristics of the second target entity-at least part of the feature of the target attribute pair, and input the feature of the second target entity-target attribute pair to the second identification model to input the second target entity- Identify whether the target attribute pair is a positive entity-positive example attribute pair or a negative example entity-negative example attribute pair, When the second target entity-target attribute pair is identified as a positive entity-positive attribute pair, the second entity is added to the set of positive entity, and the second target entity-target attribute pair An entity identifier that adds the second target entity to the set of negative example entities when identified as a negative example entity-negative example attribute pair;
A data extraction device. The data extraction device according to claim 6, comprising:
The feature of the first positive example entity-positive example attribute pair is a character string including the first positive example entity and the first positive example attribute, and includes the first positive example entity and the first positive example attribute. Including information representing the relationship between what is included in the text data and the first example entity and the first example attribute,
The feature of the first negative example entity-negative example attribute pair is a character string including the first negative example entity and the first negative example attribute, and includes the first negative example entity and the first negative example attribute. Including information representing the relationship between the text data and the first negative example entity and the first negative example attribute,
The feature of the first target entity-target attribute pair is a character string including the first target entity and the first target attribute, and is included in text data including the first target entity and the first target attribute. And information representing the relationship between the first target entity and the first target attribute,
The feature of the second positive example entity-positive example attribute pair is a character string including the second positive example entity and the second positive example attribute, and includes the second positive example entity and the second positive example attribute. Including information representing the relationship between the text data and the second positive entity and the second positive attribute,
The feature of the second negative example entity-negative example attribute pair is a character string including the second negative example entity and the second negative example attribute, and includes the second negative example entity and the second negative example attribute. Including information representing a relationship between the text data and the second negative example entity and the second negative example attribute,
The feature of the second target entity-target attribute pair is a character string including the second target entity and the second target attribute, and is included in text data including the second target entity and the second target attribute. Including information indicating the relationship between the second target entity and the second target attribute,
A data extraction apparatus characterized by that. The data extraction device according to claim 6 or 7,
A character string other than the positive example entity is selected as a positive example attribute candidate from the set of text data including the positive example entity, and the positive example attribute candidate is included in the set of character strings including the positive example entity. A predetermined number of positive example attribute candidates as the initial value of the positive example attribute from a large index indicating the magnitude of the difference between the frequency and the frequency at which the positive example attribute candidate is included in the set of all text data ,
A character string other than the negative example entity is selected as a negative example attribute candidate from the set of text data including the negative example entity, and the negative example attribute candidate is included in the set of character strings including the negative example entity. A predetermined number of negative example attribute candidates from a large index indicating the magnitude of the difference between the frequency and the frequency at which the negative example attribute candidate is included in the set of all text data and the initial value of the negative example attribute An initial attribute set generation unit
A data extraction apparatus characterized by that. The data extraction device according to any one of claims 6 to 8,
The attribute identification unit
Of the features of the first positive example entity-positive example attribute pair and the features of the first negative example entity-negative example attribute pair used as the supervised learning data in the attribute identification learning unit, generated from them Select a feature whose index indicating the degree of influence on the first identification model is greater than a specific criterion, select the text data including a character string corresponding to the selected feature, and the selected text data is A character string including the first target entity and the first target attribute,
A data extraction apparatus characterized by that. The data extraction device according to any one of claims 6 to 9,
The entity identifier is
Of the features of the second positive example entity-positive example attribute pair and the features of the second negative example entity-negative example attribute pair used as the supervised learning data in the entity identification learning unit, generated from them Select a feature whose index indicating the degree of influence on the second identification model is greater than a specific criterion, select the text data including a character string corresponding to the selected feature, and the selected text data A character string including the second target entity and the second target attribute,
A data extraction apparatus characterized by that. The data extraction device according to any one of claims 6 to 10,
The feature of the first positive example entity-positive example attribute pair includes topic information corresponding to a topic of text data including a set of the first positive example entity and the first positive example attribute,
The feature of the first negative example entity-negative example attribute pair includes topic information corresponding to a topic of text data including a set of the first negative example entity and the first negative example attribute,
The feature of the first target entity-target attribute pair includes topic information corresponding to a topic of text data including a set of the first target entity and the first target attribute,
The feature of the second positive example entity-positive example attribute pair includes topic information corresponding to a topic of text data including a set of the second positive example entity and the second positive example attribute,
The feature of the second negative example entity-negative example attribute pair includes topic information corresponding to a topic of the text data including a set of the second negative example entity and the second negative example attribute;
The feature of the second target entity-target attribute pair includes topic information corresponding to a topic of the text data including a set of the second target entity and the second target attribute.
A data extraction apparatus characterized by that. The data extraction device according to any one of claims 1 to 5,
A positive example distribution processing unit for obtaining information representing a positive example probability distribution that is an appearance probability distribution of all entities included in a set of text data including a positive example seed entity;
Information representing a topic probability distribution, which is an appearance probability distribution of all entities included in a set of text data corresponding to the same topic information, is obtained for each topic information, and the information representing the positive example probability distribution and the topic probability distribution are obtained. A negative example topic determination unit that selects at least a part of the topic information as the negative example topic information based on a distance between the positive example probability distribution obtained using information to represent and the topic probability distribution;
A negative example seed entity generation unit that selects, as a negative example seed entity, an entity corresponding to the negative example topic information selected by the negative example topic determination unit;
The processing by the topic information extraction unit, the identification learning unit, and the entity identification unit is repeated one or more times,
The positive seed entity is the positive entity in the initial processing by the topic information extraction unit,
The negative example seed entity is the data extraction device, which is the negative example entity in the initial processing by the topic information extraction unit. The data extraction device according to any one of claims 1 to 5,
The seed positive example topic information indicating the appropriateness of each topic with respect to the text data including the positive seed entity is aggregated for each topic, and the obtained aggregation result for each topic is obtained as the seed positive example topic score of the topic. Seed positive example topic score creation part,
A negative example topic determination unit that uses the topic information corresponding to the topic selected based on the magnitude of the seed positive example topic score of the topic as the negative example topic information;
A negative example seed entity generation unit that selects, as a negative example seed entity, an entity corresponding to the negative example topic information selected by the negative example topic determination unit;
The processing by the topic information extraction unit, the identification learning unit, and the entity identification unit is repeated one or more times,
The positive seed entity is the positive entity used in the initial processing by the topic information extraction unit,
The negative example seed entity is the data extraction device, which is the negative example entity used in the initial processing by the topic information extraction unit. A data extraction method executed by a data extraction device,
A pre-processing unit uses topic data representing the appropriateness of a plurality of topic candidates for text data as an index value and a topic model that describes the relationship between the text data and unsupervised learning data obtained from the text data. Preprocessing steps to learn;
Topic information extraction unit, the positive example topic information in response to the topic of the text data extracted from the topic model that includes a target of extracting positive cases entity is a character string and at least part of the identity of the positive examples entity, A topic information extraction step in which negative example topic information extracted from the topic model corresponding to a topic of text data including a negative example entity that is a character string not to be extracted is at least part of the features of the negative example entity;
The discriminating and learning unit performs learning processing using the features of the positive entity and the negative entity as supervised learning data. An identification learning step for generating an identification model that is a function for outputting information for identification;
An entity that is a character string included in text data selected from a set of text data by the entity identification unit is a target entity, and topic information extracted from the topic model corresponding to the topic of the selected text data is the target entity. When the identity of the target entity is input to the identification model to identify whether the target entity is a positive entity or a negative entity, and the target entity is identified as a positive entity Identifying the target entity as the positive example entity, and identifying the target entity as the negative example entity when the target entity is identified as a negative example entity;
A data extraction method comprising: A data extraction method executed by a data extraction device,
The attribute identifying feature extraction unit selects the first positive example entity selected from the set of positive example entities that are character strings to be extracted and the first example attribute selected from the set of positive example attributes that are character strings representing the attributes of the positive example entities. A first positive example entity-positive example attribute pair that is a set of one positive example attribute, a first negative example entity selected from a set of negative example entities that are character strings not to be extracted, and attributes of the negative example entity Generating a first negative example entity-negative example attribute pair that is a set with a first negative example attribute selected from a set of negative example attributes that is a character string to represent the first positive example entity from the set of text data And a character string including a pair of the first positive example attribute, and information indicating characteristics of the first positive example entity-positive example attribute pair for the selected character string is used as the first positive example entity-positive example. A character string including at least a part of a feature of a sex pair and including a set of the first negative example entity and the first negative example attribute from the set of text data, and selecting the first character string for the selected character string An attribute identifying feature extraction step in which information representing the characteristics of the negative example entity-negative example attribute pair is at least part of the features of the first negative example entity-negative example attribute pair;
The attribute identification learning unit performs an arbitrary character string by learning processing using the feature of the first positive example entity-positive example attribute pair and the feature of the first negative example entity-negative example attribute pair as supervised learning data. To identify whether the entity-attribute pair is a positive entity-positive example attribute pair or a negative example entity-negative example attribute pair by inputting the identity of an entity-attribute pair that is a set of an entity and the attribute of the entity An attribute identification learning step for generating a first identification model which is a function for outputting information;
The attribute identification unit selects any one of the text data from the set of text data, selects a character string included in the selected text data as a first target entity, and selects the first target entity from the selected text data. The first target attribute is selected as a first target attribute, and a set of the first target entity and the first target attribute is set as a first target entity-target attribute pair, and the first target in the selected text data is selected. Information representing the characteristics of the entity-target attribute pair is set as at least a part of the feature of the first target entity-target attribute pair, and the feature of the first target entity-target attribute pair is input to the first identification model. The first target entity-target attribute pair is a positive entity-positive example attribute pair or a negative example entity-negative example attribute pair. When the first target entity-target attribute pair is identified as a positive entity-positive attribute pair, the first target attribute is added to the set of positive attribute, An attribute identifying step of adding the first target attribute to the set of negative example attributes when the target entity-target attribute pair is identified as a negative example entity-negative example attribute pair;
A second positive example entity-positive example , in which the entity identifying feature extraction unit is a set of a second positive example entity selected from the positive example entity set and a second positive example attribute selected from the positive example attribute set. A second negative example entity-negative example attribute pair that is a set of an attribute pair, a second negative example entity selected from the set of negative example entities, and a second negative example attribute selected from the set of negative example attributes. Generating and selecting a character string including a set of the second positive example entity and the second positive example attribute from the set of text data, and the second positive example entity-positive example attribute for the selected character string. Information representing the characteristics of the pair is at least a part of the feature of the second positive example entity-positive example attribute pair, and the second negative example entity and the second negative example attribute are obtained from the set of text data. A character string including a pair is selected, and information indicating the characteristics of the second negative example entity-negative example attribute pair for the selected character string is used as at least part of the features of the second negative example entity-negative example attribute pair. An entity identification feature extraction step,
The entity identification learning unit performs an arbitrary character string by learning processing using the feature of the second positive example entity-positive example attribute pair and the feature of the second negative example entity-negative example attribute pair as supervised learning data. To identify whether the entity-attribute pair is a positive entity-positive example attribute pair or a negative example entity-negative example attribute pair by inputting the identity of an entity-attribute pair that is a set of an entity and the attribute of the entity An entity identification learning step for generating a second identification model which is a function for outputting information;
The entity identification unit selects any one of the text data from the set of text data, selects a character string included in the selected text data as a second target entity, and selects the second target entity from the selected text data. A character string different from the second target attribute is selected as a second target attribute, and a set of the second target entity and the second target attribute is set as a second target entity-target attribute pair, and the second target in the selected text data is selected. Information representing the characteristics of the entity-target attribute pair is set as at least a part of the feature of the second target entity-target attribute pair, and the feature of the second target entity-target attribute pair is input to the second identification model Second target entity-target attribute pair is positive entity-positive attribute pair or negative entity-negative And when the second target entity-target attribute pair is identified as an example entity-example attribute pair, the second entity is added to the set of example entities, An entity identification step of adding the second target entity to the set of negative example entities when the two target entity-target attribute pairs are identified as negative example entity-negative example attribute pairs;
A data extraction method comprising:   A program for causing a computer to function as each unit of the data extraction device according to claim 1.

Images