JP5117744B2 - Word meaning tag assigning device and method, program, and recording medium - Google Patents

Description

  The present invention relates to a natural language analysis technique, and more particularly to a word meaning tag assignment technique for assigning a tag indicating meaning in context to each word constituting natural language data.

In the case of handling semantic information in natural language processing, a technique is used in which the meaning of a word defined by a dictionary or thesaurus, that is, a semantic tag is automatically assigned to each word included in a sentence.
Conventionally, as such a technique, a method of applying a target sentence to a parser and determining a semantic tag of a word pair in a dependency structure from left to right (for example, see Non-Patent Document 1), a hidden Markov model A method is proposed in which semantic tags of words are determined from left to right in the same manner as in morphological analysis (see, for example, Non-Patent Document 2).
In these technologies, meaning tags are described as being defined in a dictionary or thesaurus. However, words that are classified according to meaning groups are essentially the same. For example, they appear in text. Those classified by the existing automatic classification method based on the characteristics of the distribution are also included.

SenseLeaner, Rada Mihalcea and Ehsanul Faruque, "SenseLearner: Minimally Supervised Word Sense Disambiguation for All Words in Open Text" in Proceedings of ACL / SIGLEX Senseval-3, Barcelona, Spain, July 2004. Antonia Molina et al., "WSD system based on Specialized Hidden Markov Model (upv-shmm-eaw)" in Proceedings of ACL / SIGLEX Senseval-3, Barcelona, Spain, July 2004. Y. Kasahara, "Basic Word Semantic Database: Construction of Lexeed", Information Processing Society of Japan, NLC-159, 2004

  However, in such a conventional technique, when a semantic tag is assigned to a plurality of words in a target sentence, a semantic tag is assigned to each word along the order of appearance of each word. There was a problem that the application accuracy could not be increased.

In other words, when meaning tags are assigned to each word in the order of appearance of each word, the meaning tag assignment result of words in the order of appearance has a great influence on the meaning tagging of other words. There is a risk of selecting the series.
For example, if you add meaning tags to the words “tea” and “glass” in the sentence “a man wearing a glass of tea,” the co-occurrence of “tea” The word “glass” is dragged by the meaning of “tableware”, the meaning tag of “tea” is determined as “beverage”, and the next meaning tag of “glass” is determined as “tableware” from the co-occurrence relationship with the meaning tag. ”May be mistaken.

  The present invention has been made to solve such problems, and an object thereof is to provide a word meaning tag assigning device and method, a program, and a recording medium that can improve the accuracy of meaning tag assignment.

  In order to achieve such an object, the word meaning tag assignment device according to the present invention assigns a meaning tag indicating the meaning of the target word to each target word included in the target sentence composed of natural language data. A word meaning tag assigning device for each word meaning, a storage unit for storing a word dictionary composed of a meaning tag indicating the meaning and a meaning class indicating a concept of the meaning, and a target word For a processing target word selected from among them, a primary feature amount indicating a word co-occurrence relationship between the processing target word and another word included in the target sentence is generated, and the word has a word co-occurrence relationship with the processing target word A semantic co-occurrence feature amount indicating a semantic co-occurrence relationship between a semantic tag and a semantic class obtained by searching the word dictionary in the storage unit for another target word and the processing target word is generated, and the processing target word of A feature amount extraction unit that generates an extended feature amount of the processing target word from the next feature amount and the semantic co-occurrence feature amount, and the processing target word described in the word dictionary based on the extended feature amount of the processing target word A semantic tag determination unit that determines a semantic tag indicating the most appropriate meaning of the processing target word used in the target sentence among the semantic tags.

  At this time, the feature quantity extraction unit generates a primary feature quantity indicating the word co-occurrence relationship for each word co-occurrence relationship between the processing target word and other words included in the target sentence. And semantic co-occurrence indicating a semantic co-occurrence relationship with the processing target word for each of the semantic tag and semantic class obtained by searching the word dictionary in the storage unit for each of the other target words included in the target sentence. A feature amount expansion unit that generates feature amounts and generates an extended feature amount of the target word from the semantic co-occurrence feature amount and the primary feature amount of the processing target word may be provided.

  Further, the storage unit stores a semantic tag selection model indicating the weight of each semantic tag selected corresponding to the target word for each primary feature and semantic co-occurrence feature for the target word, and a semantic tag determination unit Thus, the semantic tag selection model in the storage unit is searched for the primary feature value and the semantic co-occurrence feature value of the processing target word included in the extended feature value, and the processing target word and each meaning are determined based on the weight obtained for each semantic tag. An evaluation value may be calculated for each combination with a tag, and a semantic tag for the processing target word may be determined based on the evaluation value.

  In addition, the storage unit stores an assignment order database indicating the order in which the semantic tags are assigned to the target words, and when the semantic tag determination unit determines the semantic tags for a plurality of processing target words, the storage unit grant order database is searched. Then, the order of giving each processing target word may be acquired, and the semantic tag may be sequentially determined for each processing target word based on the order of giving.

  In addition, the morphological analysis unit that performs morphological analysis on the target sentence, the syntactic and semantic analysis unit that performs syntactic and semantic analysis on the target sentence, and the types of analysis results in the morpheme analysis unit and the syntax and semantic analysis unit are integrated for each same analysis target. The target sentence input unit may further include an analysis result integration unit that outputs to the primary feature quantity extraction unit together with the target sentence.

  The word meaning tag assignment method according to the present invention is a word meaning tag assignment method for assigning a meaning tag indicating the meaning of the target word to each target word included in the target sentence composed of natural language data. A storage step of storing a word dictionary composed of a combination of a meaning tag indicating the meaning and a meaning class indicating the classification of the meaning, provided for each meaning of the word in the storage unit; For the processing target word selected from the above, a primary feature amount indicating a word co-occurrence relationship between the processing target word and another word included in the target sentence is generated, and the other target word included in the target sentence is generated. Generating a semantic co-occurrence feature amount indicating a semantic co-occurrence relationship between the semantic tag and semantic class obtained by searching the word dictionary in the storage unit and the processing target word, and a primary feature amount of the processing target word; The processing target described in the word dictionary based on the extended feature amount of the processing target word by the feature amount extraction step for generating the extended feature amount of the processing target word from the taste co-occurrence feature amount and the semantic tag determination unit A semantic tag determination step of determining a semantic tag indicating the most appropriate meaning of the processing target word used in the target sentence among the semantic tags of the word.

Moreover, the program concerning this invention is a program for making a computer perform each step of the said word meaning tag provision method.
A recording medium according to the present invention is a recording medium on which the program is recorded.

  According to the present invention, a primary feature amount indicating a word co-occurrence relationship between the processing target word and another word included in the target sentence for the processing target word selected from the target words by the feature amount extraction unit. And the semantic co-occurrence relationship between the semantic tag and semantic class obtained by searching the word dictionary in the storage unit for other target words having a word co-occurrence relationship with the processing target word and the processing target word Is generated from the primary feature value and the semantic co-occurrence feature amount of the processing target word, and the semantic tag determination unit causes the extended feature of the processing target word to be generated. Based on the amount, the semantic tag indicating the most appropriate meaning of the processing target word used in the target sentence is determined from the semantic tags of the processing target word described in the word dictionary.

  Thereby, when assigning a semantic tag for a target word, a semantic tag is assigned based on the relationship between the target word and a word having a semantic co-occurrence relationship as well as a word having a word co-occurrence relationship with the target word. be able to. Therefore, compared with the case where a meaning tag is given to each word along the order of appearance of the word, the meaning tag can be given based on a wider range of relationships, and the meaning tag assignment accuracy can be improved.

Next, embodiments of the present invention will be described with reference to the drawings.
[First embodiment]
First, with reference to FIG. 1, the word meaning tag provision apparatus concerning the 1st Embodiment of this invention is demonstrated. FIG. 1 is a block diagram showing a configuration of a word meaning tag assigning apparatus according to a first embodiment of the present invention.
The word meaning tag assigning device 10 includes a general information processing device such as a server or a personal computer, and for each target word included in the target sentence X made of natural language data, a semantic tag indicating the meaning of the target word. Are given and output as an output sentence Y.

  In the present embodiment, the storage unit stores a word dictionary that is provided for each word meaning and includes a meaning tag that indicates the meaning and a meaning class that indicates the concept of the meaning. For the processing target word selected from the target words, a primary feature amount indicating a word co-occurrence relationship between the processing target word and another word included in the target sentence is generated, and the processing target word and the word A semantic co-occurrence feature amount indicating a semantic co-occurrence relationship between a semantic tag and a semantic class obtained by searching the word dictionary in the storage unit for another target word having a co-occurrence relationship and the processing target word is generated. Then, an extended feature quantity of the processing target word is generated from the primary feature quantity and semantic co-occurrence feature quantity of the processing target word, and is described in the word dictionary by the semantic tag determination unit based on the extended feature quantity of the processing target word. The processing target Of meaning tag word, in which so as to determine the meaning tag indicating the most appropriate means of the processing target word used in the sentence.

Hereinafter, with reference to FIG. 1, the structure of the word meaning tag provision apparatus concerning the 1st Embodiment of this invention is demonstrated in detail.
The word meaning tag assigning device 10 includes, as main functional units, an arithmetic processing unit 1, a storage unit 2, an input / output interface unit (hereinafter referred to as an input / output I / F unit) 3 as in a general information processing device. A communication interface unit (hereinafter referred to as a communication I / F unit) 4, an operation input unit 5, and a screen display unit 6 are provided.

The arithmetic processing unit 1 is composed of a microprocessor such as a CPU and its peripheral circuits, and reads and executes the program 20 stored in the storage unit 2, thereby causing the hardware and the program 20 to cooperate with each other to perform various processes. Realize the part.
As main processing units realized by the arithmetic processing unit 1, there are a target sentence analysis unit 11, a feature amount extraction unit 12, a semantic tag determination unit 13, and a semantic tag output unit 14.

The storage unit 2 is composed of a storage device such as a hard disk or a memory, and stores a program 20 executed by the arithmetic processing unit 1 and various processing information used for a semantic tag assignment process. The program 20 is read from the recording medium M via, for example, the input / output I / F unit 3 or read from an external device (not shown) via the communication I / F unit 4 and stored in the storage unit 2 in advance. The
Main processing information stored in the storage unit 2 includes a word dictionary 21 and a semantic tag selection model 22.

The input / output I / F unit 3 is composed of a dedicated data input / output circuit, and is connected to a recording medium M such as a CD, a DVD, or a nonvolatile memory card in accordance with an instruction from the arithmetic processing unit 1. It has a function of inputting / outputting various data and programs such as sentence X, output sentence Y, dictionary, and database.
The communication I / F unit 4 includes a dedicated data communication circuit, and communicates with an external device such as a server connected via a communication line such as a LAN in accordance with an instruction from the arithmetic processing unit 1. It has a function of transmitting and receiving various data and programs such as X, output sentence Y, dictionary, and database.

The operation input unit 5 includes an operation input device such as a keyboard and a mouse, and has a function of detecting an operation of the operator and outputting the operation to the arithmetic processing unit 1.
The screen display unit 6 includes a screen display device such as an LCD or a PDP, and has a function of displaying various data such as a target sentence X and an output sentence Y and an operation screen on the screen according to instructions from the arithmetic processing unit 1. Yes.

FIG. 2 is a block diagram showing a main part of the word meaning tag assigning device according to the first embodiment of the present invention.
The target sentence analysis unit 11 receives the target sentence X composed of natural language data from the storage unit 2, the input / output I / F unit 3, the communication I / F unit 4, the operation input unit 5, and the like, and performs two different language analyzes. Here, the morphological analysis unit 11A, the syntactic and semantic analysis unit 11B, and the analysis result integration unit 11C are configured.

  The morpheme analysis unit 11A has a function of performing a known morpheme analysis process on the target sentence X. The syntax and semantic analysis unit 11B has a function of performing a known syntax and semantic analysis process on the target sentence X. The analysis result integration unit 11C receives a plurality of types of analysis results received from the morphological analysis unit 11A and the syntax / semantic analysis unit 11B, associates analysis results for the same sentence, and / or further associates analysis results for the same word, And a function of integrating a plurality of different analysis results for the same sentence.

  The feature quantity extraction unit 12 refers to the word dictionary 21 of the storage unit 2 based on the language analysis result obtained by the target sentence analysis unit 11, and thereby the feature quantity (features) for each target language included in the target sentence X. Information), which is composed of a primary feature quantity extraction unit 12A and a feature quantity expansion unit 12B. The word dictionary 21 is a database composed of a set of a meaning tag indicating the meaning and a meaning class indicating the classification of the meaning, which is provided for each word for each word. It should be noted that as the word dictionary 21, an ontology clearly indicating the concept assumed in an arbitrary knowledge base and the mutual relationship between the concepts may be used.

  The primary feature quantity extraction unit 12A has a function of searching the word dictionary 21 for each word extracted from the target sentence X based on the language analysis result, and a target language to which a word having a plurality of meanings among these words is a target for giving a semantic tag. And a primary (word co-occurrence) feature amount indicating a word co-occurrence relationship between the processing target word and other words included in the target sentence X for the processing target word selected from the target words And a function of generating (primary feature information).

  The feature amount expanding unit 12B searches the word dictionary 21 for other target words included in the target sentence X, and the semantic co-occurrence relationship between the semantic tag and semantic class obtained by the search and the processing target word. A function for generating semantic co-occurrence feature values (semantic co-occurrence feature information) indicating the above, and a function for generating an extended feature amount of the processing target word from the primary feature amount and semantic co-occurrence feature amount of the processing target word. is doing.

  The semantic tag determination unit 13 is described in the word dictionary 21 by referring to the semantic tag selection model 22 in the storage unit 2 based on the extended feature amount of each target language extracted by the feature amount extraction unit 12. Among the meaning tags of the processing target word, it has a function of determining a semantic tag indicating the most appropriate meaning of the processing target word used in the target sentence. Here, the semantic tag combination search unit 13A is used. Has been. The semantic tag selection model 22 is a database indicating the weight of each semantic tag selected corresponding to the target word for each primary feature value and semantic co-occurrence feature value for the target word.

  The semantic tag combination search unit 13A is configured to search the semantic tag selection model 22 for the primary feature value and the semantic co-occurrence feature value of the processing target word included in the extended feature value, and based on the weight obtained for each semantic tag. A function of calculating an evaluation value for each combination of a processing target word and each semantic tag, and a function of determining a semantic tag of the processing target word by obtaining a combination having the highest evaluation value by a search algorithm such as a beam search have.

  The semantic tag output unit 14 has a function of generating an output sentence Y to which an optimal semantic tag is assigned to each target language of the target sentence X based on the semantic tag of each target word determined by the semantic tag determination unit 13. The output sentence Y has a function of outputting to the storage unit 2, the input / output I / F unit 3, the communication I / F unit 4, the screen display unit 6, and the like.

[Operation of the first embodiment]
Next, with reference to FIG. 2, the operation of the word meaning tag assigning device according to the first exemplary embodiment of the present invention will be described.
When the operation input unit 5 detects a process start operation by the operator, the arithmetic processing unit 1 of the word meaning tag adding device 10 first receives a target sentence X to be given a semantic tag by the target sentence analyzing unit 11. Perform language analysis processing. If the target sentence X has already been subjected to language analysis, the target sentence analysis unit 11 does not need language analysis processing, and the input target sentence X and its language analysis result are passed to the feature quantity extraction unit 12.

  When the target sentence X is natural language data (text data) of “a man wearing a brown glass”, the morpheme analysis unit 11A receives the target sentence X and performs morpheme analysis processing, and the obtained word and its part of speech. Is output as an analysis result. FIG. 3 is an example of a morphological analysis result. In this example, the target sentence X is divided into seven words, and parts of speech are given to the respective words. The character position represents the start and end positions (character digits) of the character in the target sentence X of each word. For example, “brown” indicates that the target sentence X is located from the first character to the first character.

On the other hand, the syntax and semantic analysis unit 11B receives the target sentence X, performs syntax and semantic analysis, and outputs a semantic relationship between the words. FIG. 4 is an example of a syntax and semantic analysis result. In this example, (1) there is a noun phrase “tea glass”, and the two words of the main word “glass” and the modifier “tea” are connected by the particle “no”, and (2) There is a noun phrase “a man wearing a glass”, which indicates two relationships, “man” is the main character and “glass” is the target character for the predicate “it was”. The parenthesis indicates the character position in the original sentence as in FIG.
The analysis result integration unit 11C receives different language analysis results of the morpheme analysis unit 11A and the syntax / semantic analysis unit 11B, combines them into one result, and sends the result to the feature amount extraction unit 12.

Next, the arithmetic processing unit 1 causes the feature amount extraction unit 12 to extract the feature amount of each target word based on the language analysis result of the target sentence X obtained by the target sentence analysis unit 11.
First, the feature quantity extraction unit 12 uses the primary feature quantity extraction unit 12 </ b> A to identify a target word to be given a semantic tag based on the language analysis result of the target sentence X and the word dictionary 21. FIG. 5 is an example of a word dictionary.

In this example, the word “tea” has two meanings corresponding to the meaning tags “tea ₁ ” and “tea ₂ ”, and the word “glass” has two meanings corresponding to the meaning tags “glass ₁ ” and “glass ₂ ”. one of meaning is present, only the meaning of the one corresponding to the meaning tag "man _1" in the word "man" is present, the word "put" there is only the meaning of the one corresponding to the meaning tag "multiplied by _1" It shows that you are doing. Accordingly, here, two words “tea” and “glass” having a plurality of meanings are determined as meaning tagging target words.

Subsequently, the primary feature quantity extraction unit 12A sequentially selects processing target words one by one from these target words, and generates a primary feature quantity for each processing target word. FIG. 6 is a definition example of the primary feature amount. FIG. 7 is an example of primary feature values extracted according to the definition of FIG. For example, CoWord (glass) in the “tea (1, 1)” column indicates that the word “glass” co-occurs in the sentence in which the target word “tea” appears. .
At this time, the primary feature quantity extraction unit 12A first extracts a word co-occurrence relationship between the processing target word and other words included in the target sentence X based on the definition of FIG. A primary feature is generated. And the primary feature-value which consists of the enumeration from these primary feature-values is produced | generated.

  Next, the feature quantity extraction unit 12 refers to the word dictionary 21 by the feature quantity expansion unit 12B, and indicates meaning tags of other target words having a word co-occurrence relationship with the processing target word and the meaning classification. By searching for a class, a new feature quantity indicating the co-occurrence relationship between each of the obtained semantic tags and semantic classes and the processing target word is generated, and this is added to the primary feature quantity to obtain a primary feature quantity. An extended feature amount is generated by extending the feature amount to a broader concept.

FIG. 8 is an example of the extended feature amount. For example, when the word “glass” included in the primary feature value CoWord (glass) of the target word “tea” is “glass ₁ : glass cup”, the word dictionary 21 is searched and the meaning tag of the word “glass” is searched. Acquire “Glass ₁ ” and semantic class [tableware], and replace it with a new feature quantity CoWord (glass ₁ ) that replaces “Glass” of the primary feature quantity with the semantic tag “glass ₁ ”, and semantic class [tableware]. A new feature value CoWord ([tableware]) is generated.

The primary feature value is a feature value indicating a co-occurrence relationship at the word level between the target word “tea” and another word, whereas these new feature values are the same for the target word “tea”. It is a feature quantity indicating a co-occurrence relation at a semantic level with another target word having an origin relation, that is, a semantic co-occurrence relation. These new feature quantities can be said to be semantic co-occurrence feature quantities.
Therefore, by adding the semantic co-occurrence feature CoWord ([tableware]) replaced with the semantic class [tableware], other words such as “dish” and “mug” belonging to the semantic class of [tableware] appeared. It is possible to obtain a feature value indicating that the sentence has a similar context.

Next, the arithmetic processing unit 1 determines the optimum semantic tag for each target word of the target sentence X based on the extended feature amount of each target language extracted by the feature amount extraction unit 12 by the semantic tag determination unit 13. To do.
The semantic tag determination unit 13 calculates the evaluation value of each semantic tag combination from the semantic tag selection model 22 in the storage unit 2 by the semantic tag combination search unit 13A, and searches for a semantic tag combination that maximizes the evaluation value. Ask for.

  FIG. 9 is an example of a semantic tag selection model based on the maximum entropy model. The numbers in FIG. 9 represent the weights at which each semantic tag is selected when each feature amount is obtained for the target word. This weight is calculated using the maximum entropy model (A. Berger et al .: "A maximum entropy approach to natural language processing.", Computational Linguistics, 22 (1), 39- 71).

 When the semantic tag selection model is the maximum entropy model, the probability p (t | c) that the semantic tag t is selected in the context c of the target sentence is obtained by the following expressions (1) and (2): This is an evaluation value. Here, fi (t, c) is a feature function, which is a binary function that is 1 when each feature quantity i is obtained, and 0 otherwise. Further, λi is a parameter of the model and corresponds to “weight” in FIG.

In the initial state, no semantic tag is determined, and only “CoWord (glass)” has a non-zero weight among the feature values for the word “brown”.
Z (c) p (Brown ₁ | c) = 0.6 × 1 = 0.6
Z (c) p (Brown ₂ | c) = 0.2
Therefore, the evaluation value (probability) is
p (Brown ₁ | c) = 0.6 / (0.6 + 0.2) = 0.75
p (Brown ₂ | c) = 0.2 / (0.6 + 0.2) = 0.25
It is.

Similarly, for the word “glass”, from the feature quantities “CoWord (brown)” and “ARG2 / REL (predicate_kake)”,
Z (c) p (Glass ₁ | c) = 0.2 × 1 + 0.02 × 1 = 0.22
Z (c) p (glass ₂ | c) = 0.3 × 1 + 1.3 × 1 = 1.6
Therefore, the evaluation value (probability) is
p (glass ₁ | c) = 0.22 / (0.22 + 1.6) = 0.121
p (glass ₂ | c) = 1.6 / (0.22 + 1.6) = 0.879
It is.

Thereafter, the semantic tag combination search unit 13A searches for a semantic tag having the highest evaluation value by a beam search algorithm.
FIG. 10 is an example of a beam search algorithm. Here, the set of target words is W = {w ₁ ,..., W _n }, the k-th semantic tag of the word wi is t _wik , the list of determined semantic tags is T, and the beam width is b. . The search is performed using N = [W, T] as a search node. Further, the node evaluation value s (N) is obtained when s (N) = p (t _wik ), assuming that the node N determines the meaning tag of the word wi as t _wik from the immediately preceding node N_1 = [W_1, T_1]. | C) · s (N_1). As described above, p (t _wik | c) can be obtained from the extended feature quantity of FIG. 8, the morphological analysis result of FIG. 3, and the semantic tag selection model 22.

In this example, in step S1 of FIG. 10, the initial node N ₀ = [W ₀ = {brown, glass}, T0 = {}] is inserted into the initial queue Q ₀ . In the subsequent step S2, “brown” and “grass” are taken out in order from W ₀ , and a node with a semantic tag determined for each is created. Insert these nodes into the queue Q ′, and Q ′ = <[{Glass}, {Brown ₁ }], [{Glass}, {Brown ₂ }], [{Brown}, {Glass ₁ }], [{ Tea}, {glass ₂ }]>.

Here, the evaluation value s ([{glass}, {tea ₁ }]) is equal to p (tea ₁ | c) and is 0.75, and the evaluation value s ([{glass}, {tea ₂ }]) Is equal to p (brown ₂ | c) and is 0.25. Similarly, s ([{Brown}, {Glass ₁ }]) = 0.121 and s ([{Brown}, {Glass ₂ }]) = 0.879.
In step S3, when the queue Q ′ is sorted in descending order of evaluation value, Q ′ = <[{Brown}, {Glass ₂ }], [{Glass}, {Brown ₁ }], [{Glass}, { Tea ₂ }], [{Brown}, {Glass ₁ }]>. If the beam width is b = 2, the new cue Q becomes Q = <[{Brown}, {Glass ₂ }], {Glass}, {Brown ₁ }]>.

Next, from the first node [{brown}, {glass ₂ }] of the queue Q, new feature quantities CoWord (glass ₂ ), CoWord ([clothes]), ARG1 / REL-ARG2 (noun phrase_ [Clothes]). When the calculation proceeds in the same manner, in step S4, [{}, {Brown ₂ , Glass ₂ }] finally becomes the node with the highest evaluation value, and this combination of semantic tags is passed to the semantic tag output unit 14. .

FIG. 11 is an example of a semantic tag assignment result. Meaning tags “tea ₂ ” and “glass ₂ ” are assigned to the target words “tea” and “glass”, respectively. Since “male” and “kake” have only one semantic tag on the word dictionary 21, the only semantic tags “male ₁ ” and “kake ₁ ” are automatically assigned.

  As described above, in the present embodiment, the storage unit 2 stores the word dictionary 21 that is provided for each word meaning and includes a meaning tag that indicates the meaning and a meaning class that indicates the concept of the meaning. The feature amount extraction unit 12 generates a primary feature amount indicating a word co-occurrence relationship between the processing target word and another word included in the target sentence for the processing target word selected from the target words. Meaning indicating a semantic co-occurrence relationship between a semantic tag and a semantic class obtained by searching the word dictionary in the storage unit for another target word having a word co-occurrence relationship with the processing target word and the processing target word A co-occurrence feature amount is generated, an extended feature amount of the processing target word is generated from the primary feature amount and the semantic co-occurrence feature amount of the processing target word, and the semantic tag determination unit 13 converts the processing target word into an extended feature amount. Based word dictionary Of meaning tag to be processed words are written, and determines the semantic tags indicating the most appropriate means of the processing target word used in the sentence.

  Thereby, when assigning a semantic tag for a target word, a semantic tag is assigned based on the relationship between the target word and a word having a semantic co-occurrence relationship as well as a word having a word co-occurrence relationship with the target word. be able to. Therefore, compared with the case where a meaning tag is given to each word along the order of appearance of the word, the meaning tag can be given based on a wider range of relationships, and the meaning tag assignment accuracy can be improved.

  In this embodiment, language analysis is performed by a plurality of different analysis processes for the target sentence, and word co-occurrence relations and semantic co-occurrence relations are extracted based on analysis information obtained by integrating these analysis results. As a result, feature quantities can be generated accurately, and more accurate semantic tags can be assigned.

[Second Embodiment]
Next, with reference to FIG. 12, the word meaning tag provision apparatus concerning the 2nd Embodiment of this invention is demonstrated. FIG. 12 is a block diagram showing the main part of the word meaning tag assigning apparatus according to the second embodiment of the present invention, and the same or equivalent parts as those in FIG.
In the first embodiment, a case has been described as an example where the target sentence analysis unit 11 performs two language analyzes on the target sentence X, which are different in morphological analysis and syntax-semantic analysis. In the present embodiment, a case where only the morphological analysis is performed in the target sentence analysis unit 11 will be described.

  Compared to the first embodiment, the word meaning tagging device 10 according to the present embodiment is different in that the target sentence analysis unit 11 is configured only from the morpheme analysis unit 11A. Other configurations are the same as those of the first embodiment described above, and a detailed description thereof is omitted here.

[Operation of Second Embodiment]
Next, the operation of the word meaning tag assigning device according to the second exemplary embodiment of the present invention will be described with reference to FIG.
When the operation input unit 5 detects a process start operation by the operator, the arithmetic processing unit 1 of the word meaning tag adding device 10 first receives a target sentence X to be given a semantic tag by the target sentence analyzing unit 11. Perform language analysis processing.
When the target sentence X is natural language data (text data) that “a man made a tea with a glass”, the target sentence analysis unit 11 receives the target sentence X by the morpheme analysis unit 11A and performs morpheme analysis processing. The obtained word and its part of speech are output as an analysis result. FIG. 13 is an example of the analysis result.

Next, the arithmetic processing unit 1 causes the feature amount extraction unit 12 to extract the feature amount of each target word based on the language analysis result of the target sentence X obtained by the target sentence analysis unit 11.
First, the feature quantity extraction unit 12 uses the primary feature quantity extraction unit 12 </ b> A to identify a target word to be given a semantic tag based on the language analysis result of the target sentence X and the word dictionary 21.
In this example, there are two meanings for each of the words “glass”, “tea”, and “take out”, so these three words “tea”, “glass”, and “take out” are subject to semantic tagging. Determined as a word.

Subsequently, the primary feature quantity extraction unit 12A sequentially selects processing target words one by one from these target words, and generates a primary feature quantity for each processing target word.
At this time, the primary feature quantity extraction unit 12A first extracts the word co-occurrence relationship between the processing target word and other words included in the target sentence X based on the definition of FIG. A primary feature amount indicating a co-occurrence relationship is generated. And the primary feature-value which consists of the enumeration from these primary feature-values is produced | generated. FIG. 14 is an example of the primary feature amount.

  Next, the feature quantity extraction unit 12 refers to the word dictionary 21 by the feature quantity expansion unit 12B, and indicates meaning tags of other target words having a word co-occurrence relationship with the processing target word and the meaning classification. By searching for a class, a new feature quantity indicating the co-occurrence relationship between each of the obtained semantic tags and semantic classes and the processing target word is generated, and this is added to the primary feature quantity to obtain a primary feature quantity. An extended feature amount is generated by extending the feature amount to a broader concept. FIG. 15 is an example of the extended feature amount.

Next, the arithmetic processing unit 1 determines the optimum semantic tag for each target word of the target sentence X based on the extended feature amount of each target language extracted by the feature amount extraction unit 12 by the semantic tag determination unit 13. To do.
The semantic tag determination unit 13 calculates the evaluation value of each semantic tag combination from the semantic tag selection model 22 in the storage unit 2 by the semantic tag combination search unit 13A, and searches for a semantic tag combination that maximizes the evaluation value. Ask for.

FIG. 16 is an example of a semantic tag selection model based on the maximum entropy model. In the initial state, no semantic tag is determined, and “CoWord (male)”, “CoWord (brown)”, and “CoWord (out)” have a non-zero weight among the feature quantities for the word “glass”. Because
Z (c) p (Glass ₁ | c) = 0.1 × 1 + 0.1 × 1 + 0.2 × 1 = 0.4
Z (c) p (glass ₂ | c) = 0.4 + 1 + 0.1 × 1 = 0.5
Therefore, the evaluation value (probability) is
p (glass ₁ | c) = 0.4 / (0.4 + 0.5) = 0.44
p (glass ₂ | c) = 0.5 / (0.4 + 0.5) = 0.56
It is.

Similarly, with respect to the word “tea”, the feature quantities “CoWord (glass)” and “CoWord (deposit)” are used. Z (c) p (tea ₁ | c) = 0.8 × 1 + 0.9 × 1 = 1.7
Z (c) p (Brown ₂ | c) = 0.2 × 1 + 0.1 × 1 = 0.3
Therefore, the evaluation value (probability) is
p (Brown ₁ | c) = 1.7 / (1.7 + 0.3) = 0.85
p (Brown ₂ | c) = 0.3 / (1.7 + 0.3) = 0.15
It is.

Similarly, for the word “out”, Z (c) p (out ₁ | c) = 0.2 × 1 + 0 = 0.2 from the feature quantities “CoWord (glass)” and “CoWord (brown)”.
Z (c) p ( ₂ < ₂ > c) = 0.3 * 1 + 0.9 * 1 = 1.2
Therefore, the evaluation value (probability) is
p ( _{1 to be} put out | c) = 0.2 / (0.2 + 1.2) = 0.14
p ( ₂ out | c) = 0.2 / (0.2 + 1.2) = 0.86
It is.

Thereafter, the semantic tag combination search unit 13A searches for a semantic tag having the highest evaluation value by a beam search algorithm.
In this example, in step S1 of FIG. 10, the initial node N ₀ = [W ₀ = {glass, brown, out}, T0 = {}] is inserted into the initial queue Q ₀ . In the subsequent step S2, “glass”, “brown”, and “take out” are extracted in order from W ₀ , and a node for which a semantic tag is determined is created for each. Insert these nodes into the queue Q ', Q' = <[{Brown, out}, {Glass ₁ }], [{Brown, out}, {Glass ₂ }], [{Glass, out}, {Brown ₁ }], [{Glass, out}, {Brown ₂ }], [{Glass, tea}, {Out ₁ }], [{Glass, tea}, {Out ₂ }]>.

In the step S3, when the queue Q ′ is sorted in descending order of evaluation value, Q ′ = <[{glass, tea}, {out ₂ }], [{glass, out}, {tea ₁ }], [{ Tea, out}, {glass ₂ }], [{tea, out}, {glass ₁ }], [{glass, out}, {tea ₂ }], [{glass, tea}, {out ₁ }]> It becomes.
If the beam width b = 2, the new cue Q becomes Q = <[{glass, brown}, {out ₂ }], [{glass, out}, {brown ₁ }]>.

When the calculation proceeds in the same manner, in step S4, [{}, {Glass ₁ , Brown ₁ , Out ₂ }] finally becomes the node with the highest evaluation value, and this semantic tag combination is used as the semantic tag output unit 14. To pass. FIG. 17 is an example of a semantic tag assignment result. Meaning tags “Glass ₁ ”, “Brown ₁ ” and “Draw ₂ ” are assigned to the target words “Glass”, “Brown” and “Det”, respectively. Since “male” has only one semantic tag on the word dictionary 21, the only semantic tag “male ₁ ” is automatically assigned.

  As described above, in the present embodiment, since only the morphological analysis is performed in the target sentence analysis unit 11, the processing load on the arithmetic processing unit 1 can be reduced. Note that, since less information is used than in the first embodiment, the accuracy of extracting word co-occurrence relationships and semantic co-occurrence relationships may be lowered, but word co-occurrence relationships and semantic co-occurrence relationships are used. , Meaning tags can be assigned based on a wider range of relationships, and the accuracy of meaning tag assignment can be increased.

[Third embodiment]
Next, with reference to FIG. 18, the word meaning tag provision apparatus concerning the 3rd Embodiment of this invention is demonstrated. FIG. 18 is a block diagram showing a main part of a word meaning tag assigning device according to the third exemplary embodiment of the present invention. The same or equivalent parts as those in FIG.

  In the first embodiment, the semantic tag determination unit 13 refers to the semantic tag selection model 22 in the storage unit 2 based on the extended feature amount of each target language extracted by the feature amount extraction unit 12. The case where the semantic tag indicating the most appropriate meaning of the target word is determined has been described. In the present embodiment, when the assignment priority order database 23 is provided in the storage unit 2 in advance and the meaning tag determination unit 13 assigns the meaning tag of the target word, each priority word is assigned to each target word in the order based on the priority order of the assignment priority order database 23. A case where a semantic tag is added will be described.

  Compared with the first embodiment, in the word meaning tag assignment device 10 according to the present embodiment, a meaning tag assignment order determination unit 13B is added to the meaning tag determination unit 13, and the storage unit 2 is given priority. The difference is that a sequential database 23 is added. Other configurations are the same as those of the first embodiment described above, and a detailed description thereof is omitted here.

  FIG. 19 is an example of a priority database. Here, the entropy of the semantic tag for each word measured by a specific corpus is used as the priority order data indicating the priority order of each target word. In general, the lower the entropy, the smaller the variation in the use of word semantic tags, and the easier it is to determine semantic tags. Therefore, according to this example, the meaning tags are assigned to the target words in the order of (1) glass and (2) brown.

  The semantic tag determination unit 13 determines an optimal semantic tag for each target word of the target sentence X based on the extended feature amount of each target language extracted by the feature amount extraction unit 12 by the semantic tag combination search unit 13A. At this time, the meaning tag assignment order determination unit 13B determines the meaning tag assignment order to each target word in advance. The meaning tag assignment order determination unit 13B searches the assignment priority order database 23 of the storage unit 2 for each target word, obtains priority order data of each target word, here, entropy of the meaning tag, and uses this priority order data. Based on this, the order in which semantic tags are assigned to each target word is determined.

  The semantic tag combination search unit 13A selects one processing target word from each target word based on the order of assignment determined by the semantic tag assignment order determination unit 13B, and the processing target word is the same as that of the first embodiment. A semantic tag having the highest evaluation value is searched for by a similar beam search algorithm.

  As described above, in the present embodiment, the storage unit 2 stores the assignment order database 23 indicating the order in which the semantic tags are assigned to the target words, and the semantic tag determination unit 13 stores the semantic tags for a plurality of processing target words. Since the acquisition order database of the storage unit is searched to obtain the assignment order of each processing target word, and the meaning tag is determined in order for each processing target word based on these assignment orders, Semantic tags can be assigned with higher accuracy.

  In the present embodiment, the case where entropy is used as the priority order data in the meaning tag assignment order of words has been described as an example. However, the present invention is not limited to this. For example, “word” described in cited document 3 The order of meaning tag assignment may be determined based on other criteria, such as using “familiarity” and assigning semantic tags in order of increasing word familiarity.

It is a block diagram which shows the structure of the word meaning tag provision apparatus concerning the 1st Embodiment of this invention. It is a block diagram which shows the principal part of the word meaning tag provision apparatus concerning the 1st Embodiment of this invention. It is an example of a morphological analysis result. It is an example of a syntactic and semantic analysis result. It is an example of a word dictionary. It is a definition example of a primary feature-value. It is an example of the primary feature-value extracted according to the definition of FIG. It is an example of an extended feature-value. It is an example of the semantic tag selection model based on the maximum entropy model. It is an example of a beam search algorithm. It is an example of a meaning tag provision result. It is a block diagram which shows the principal part of the word meaning tag provision apparatus concerning the 2nd Embodiment of this invention. It is an example of another analysis result. It is an example of another primary feature-value. It is an example of another extended feature-value. It is an example of another semantic tag selection model based on the maximum entropy model. It is an example of a meaning tag provision result. It is a block diagram which shows the principal part of the word meaning tag provision apparatus concerning the 3rd Embodiment of this invention. It is an example of a priority database.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 10 ... Word meaning tag provision apparatus, 1 ... Arithmetic processing part, 11 ... Object sentence analysis part, 11A ... Morphological analysis part, 11B ... Syntax-and-meaning analysis part, 11C ... Analysis result integration part, 12 ... Feature-value extraction part, 12A ... Primary feature quantity extraction unit, 12B ... feature quantity expansion unit, 13 ... semantic tag determination unit, 13A ... semantic tag combination search unit, 13B ... semantic tag assignment order determination unit, 14 ... semantic tag output unit, 2 ... storage unit, 20 ... Program, 21 ... Word dictionary, 22 ... Semantic tag selection model, 23 ... Granting priority order database, 3 ... Input / output I / F unit, 4 ... Communication I / F unit, 5 ... Operation input unit, 6 ... Screen display unit , X ... target sentence, Y ... output sentence, M ... recording medium.

Claims (7)

A word meaning tag assigning device that assigns a meaning tag indicating the meaning of a target word to each target word included in a target sentence composed of natural language data,
A storage unit that stores a word dictionary including a set of a meaning tag indicating the meaning and a meaning class indicating the concept of the meaning, provided for each meaning of the word;
For a processing target word selected from the target words, a primary feature amount indicating a word co-occurrence relationship between the processing target word and another word included in the target sentence is generated and included in the target sentence A semantic co-occurrence feature amount indicating a semantic co-occurrence relationship between the semantic tag and semantic class obtained by searching the word dictionary of the storage unit for the other word and the processing target word is generated, and the processing target word a feature extraction unit which adds the meaning co-occurrence characteristic amount to the primary characteristic of generating an extended feature amount of the processing target word,
Meaning indicating the most appropriate meaning of the processing target word used in the target sentence among the semantic tags of the processing target word described in the word dictionary based on the extended feature amount of the processing target word A semantic tag determination unit for determining a tag , and
The storage unit stores a weight given to each set of the extended feature amount of the processing target word and the semantic tag when determining the probability that each semantic tag is selected by the maximum entropy method,
The semantic tag determining unit multiplies the weight given to each semantic tag corresponding to the extended feature amount, with respect to the extended feature amount of the processing target word, and uses the maximum entropy method to determine the processing target word and each semantic tag. And calculating a probability for each combination, and determining a semantic tag having the highest probability among semantic tags corresponding to the processing target word as a semantic tag of the processing target word . In the word meaning tag grant apparatus of Claim 1,
The feature amount extraction unit includes:
A primary feature amount extraction unit that generates a primary feature amount indicating the word co-occurrence relationship for each word co-occurrence relationship between the processing target word and another word included in the target sentence;
A word dictionary in the storage unit is searched for each of the other target words included in the target sentence, and a semantic co-occurrence relationship indicating a semantic co-occurrence relationship with the processing target word is obtained for each of the obtained semantic tags and semantic classes. A word meaning tag comprising: a feature amount expansion unit that generates each occurrence feature amount and generates an extended feature amount of the target word from the semantic co-occurrence feature amount and a primary feature amount of the processing target word Granting device. In the word meaning tag grant apparatus of Claim 1 ,
The storage unit stores an assignment order database indicating an assignment order of semantic tags for the target word;
When the semantic tag determining unit determines a semantic tag for a plurality of processing target words, the semantic tag determining unit acquires a processing order of the processing target words by searching a storage order database of the storage unit, and performs each processing based on the processing order. A semantic word tagging apparatus characterized in that a semantic tag is sequentially determined for a target word. In the word meaning tag grant apparatus of Claim 1,
The morphological analysis unit that performs morphological analysis on the target sentence, the syntax-semantic analysis unit that performs syntax-separation analysis on the target sentence, and the types of analysis results in the morpheme analysis unit and the syntax-semantic analysis unit are integrated for each same analysis target. A word meaning tagging apparatus, further comprising: a target sentence input unit including an analysis result integration unit that outputs the target sentence together with the target sentence to the primary feature amount extraction unit. A word meaning tag assigning method for assigning a meaning tag indicating the meaning of a target word to each target word included in a target sentence composed of natural language data,
A storage step of storing in the storage unit a word dictionary composed of a combination of a meaning tag indicating the meaning and a meaning class indicating the classification of the meaning, provided for each meaning of the word;
A feature amount extraction unit generates a primary feature amount indicating a word co-occurrence relationship between the processing target word and another word included in the target sentence for the processing target word selected from the target words, Semantic co-occurrence feature quantity indicating a semantic co-occurrence relationship between a semantic tag and a semantic class obtained by searching the word dictionary in the storage unit for other words included in the target sentence and the processing target word is generated. A feature amount extraction step of generating the extended feature amount of the processing target word by adding the semantic co-occurrence feature amount to the primary feature amount of the processing target word;
Based on the extended feature amount of the processing target word, the semantic tag determination unit selects the most processing target word used in the target sentence from the semantic tags of the processing target word described in the word dictionary. A semantic tag determination step for determining a semantic tag indicating an appropriate meaning , and
The storing step stores a weight given to each set of the extended feature amount of the processing target word and the semantic tag when obtaining the probability that each semantic tag is selected by the maximum entropy method,
The semantic tag determination step multiplies the weight given to each semantic tag corresponding to the extended feature amount by multiplying the extended feature amount of the processing target word by the maximum entropy method and each semantic tag. And calculating a probability for each combination of the above and determining a semantic tag having the highest probability among semantic tags corresponding to the processing target word as a semantic tag of the processing target word . The program for making a computer perform each step of the word meaning tag provision method of Claim 5 . A recording medium on which the program according to claim 6 is recorded.

Images