JP2007219947A - Causal relation knowledge extraction device and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a causal relation knowledge extraction device and program, capable of extracting causal relation between nouns (noun pair) even in the case that a key word doesn't appear or verbs in parallel words don't have a common object. <P>SOLUTION: A noun pair extraction part 2 extracts a noun pair from text data, and a feature extraction part 3 extracts features of a syntactic structure of text data including the noun pair and features of an attribute of the noun pair to generate a three-item set. Data without teacher where the presence or the absence of causal relation is not shown and data with teacher where the presence or the absence of causal relation is shown are stored by a tagging part 4. A machine learning part 5 uses the stored data to determine about causal relation and stores the result. A text analysis interface part 6 discriminates the presence or the absence of causal relation of the noun pair on the basis of the result data with respect to text data being an object of text analysis. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to information extraction and natural language processing for digitized text data, and in particular, by using natural language processing technology, pairs of nouns and causal relationships having causal relationships can be identified from text data. The present invention relates to a technique for extracting a syntactic structure.

  Currently, in digital broadcasting services, a large amount of text data such as data broadcasting and caption broadcasting is multiplexed and broadcast. If a device that receives digital broadcasts can constantly monitor these text data and extract and store useful information, it will be possible to realize a television that answers the viewer's questions. Under such circumstances, research for automatically extracting a relationship such as “cause-result” between words (hereinafter referred to as “causal relationship”) is underway.

  For example, Non-Patent Document 1 proposes a causal relationship extraction method for Japanese text, using the word “for” as a clue word. In Non-Patent Document 2, assuming that a parallel phrase exists in one sentence and the verbs in the parallel phrase have a common object, it is easy to establish a causal relation. An extraction method has been proposed.

Inui Takashi, Inui Kentaro, Matsumoto Yuji, “Automatic Acquisition of Causal Relationship Knowledge from Document Set Based on Connection Sign“ For ””, Information Processing, Vol. 45, No. 3, pp. 919-933 (2004) Kentaro Torizawa, “Automatic Extraction of“ Commonsense ”Inference Rules from Corpus”, 9th Annual Conference of the Association for Natural Language Processing, pp. 318-321 (2003)

  However, in the method of Non-Patent Document 1 described above, there is a problem that the causal relationship cannot be extracted when no clue word appears. In the method of Non-Patent Document 2, there is a problem that the causal relationship cannot be extracted when the verbs in the parallel phrase do not have a common object.

  Therefore, the present invention has been made to solve the above problems, and the purpose of the present invention is to provide a noun even when a clue word does not appear or when a verb in a parallel word does not have a common object. It is an object of the present invention to provide a causal relationship knowledge extraction apparatus and program capable of extracting a causal relationship between spaces (noun pairs).

  In order to solve the above-described problem, the causal relationship knowledge extraction device according to the present invention is a device that extracts the presence or absence of a causal relationship as knowledge from noun pairs included in text data, and extracts noun pairs included in text data. A noun pair extracting unit that extracts the syntax structure of the text data and the attributes of the noun pair as feature data for the text data including the noun pair extracted by the noun pair extracting unit, and the feature extraction A tagging unit that adds a tag indicating whether or not a causal relationship exists in the noun pair for a part of the feature data extracted by the unit, and feature data to which the tag is added by the tagging unit is stored as supervised data. A storage unit in which feature data not tagged by the tagging unit is stored as unsupervised data; Based on the supervised data and unsupervised data is characterized by comprising a causality determination unit determines the presence or absence of a causal relationship noun pairs in the 該教 nurses no data.

  In the causal relationship knowledge extracting apparatus according to the present invention, the feature extraction unit generates a tree structure indicating a dependency relation of the noun pair for the text data including the noun pair, and the tree structure generation PSE generation means for separating nodes constituting the tree structure of text data generated by the means into independent words and function words and generating PSE (Preorder String Expression) indicating the syntax structure of the text data; and the noun pair A ternary set for generating a ternary set as feature data using a superordinate concept extracting means for extracting a superordinate concept of the PSE, a PSE generated by the PSE generating means, and a superordinate concept extracted by the superordinate concept extracting means And generating means.

  The causal relationship determination unit calculates the probability of the presence or absence of a causal relationship in the data with respect to supervised data stored in the storage unit by an EM algorithm, and the supervised data and unsupervised data stored in the storage unit Calculating the probability of occurrence of a syntactic structure having the same or a similarity equal to or higher than a predetermined value, and for the supervised data and the unsupervised data stored in the storage unit, the same attribute as the noun pair included in the data Based on these probabilities, the probability of occurrence of causality in the feature data, the probability of causality in the syntactic structure included in the feature data, and the causality in the noun pair included in the feature data are calculated based on these probabilities. It is preferable to determine the presence or absence of a causal relationship by calculating the probability of the presence or absence of a relationship.

  The causal relationship knowledge extracting apparatus according to the present invention further inputs new text data, extracts noun pairs included in the text data, and extracts the syntax structure of the text data and the attributes of the noun pairs as feature data. A text determination unit for determining the presence or absence of a causal relationship in a noun pair included in the new text data based on the probability of the presence or absence of a causal relationship calculated by the causal relationship determination unit.

  The text determination unit uses the extracted feature data, and based on the probability of the presence or absence of a causal relationship in the syntax structure included in the feature data calculated by the causal relationship determination unit, the syntax structure included in the extracted feature data It is preferable to calculate the probability of the presence or absence of a causal relationship in, and determine the presence or absence of the causal relationship based on the probability.

  The text determination unit is included in the extracted feature data based on the probability of the presence or absence of a causal relationship in a noun pair included in the feature data calculated by the causal relationship determination unit using the extracted feature data. It is preferable to calculate the probability of the presence or absence of a causal relationship in the noun pair and determine the presence or absence of the causal relationship based on the probability.

  Although the present invention has been described as a causal relationship knowledge extraction device, the present invention can be substantially realized as a program executed by a computer constituting the causal relationship knowledge extraction device. A knowledge extraction program is also included. That is, the causal relationship knowledge extraction program according to the present invention is a causal relationship knowledge extraction program by a device that extracts the presence or absence of a causal relationship between noun pairs included in text data, and is included in the text data in the computer constituting the device. A process of extracting a noun pair to be processed, a process of extracting a syntactic structure of text data including the noun pair and an attribute of the noun pair as feature data, and a causal relationship in the noun pair with respect to a part of the extracted feature data. Causality in noun pairs included in unsupervised data based on the process of assigning presence / absence as a tag, supervised data that is feature data to which the tag is attached, and unsupervised data that is feature data to which no tag is attached And a process for determining whether or not there is a relationship.

  As described above, according to the present invention, the causal relationship in the noun pair is extracted even when there is no clue word in the text data or when the verb in the parallel word does not have a common object. It becomes possible.

The best mode for carrying out the present invention will be described below in detail with reference to the drawings.
〔Constitution〕
First, the configuration of the causal relationship knowledge extraction apparatus according to the embodiment of the present invention will be described. FIG. 1 is a block diagram showing the configuration of the causal relationship knowledge extraction apparatus. The causal relationship knowledge extraction device 1 includes a noun pair extraction unit 2, a feature extraction unit 3, a tagging unit 4, a machine learning unit 5, a text analysis interface unit 6, and storage units 7-1 to -4. Unsupervised data in which the presence / absence of causal relationship is not indicated is stored in the storage unit 7-2 by the storage unit 7-1 storing the text data, the noun pair extraction unit 2, the feature extraction unit 3 and the tagging unit 4. The supervised data indicating the presence or absence of the causal relationship is stored in the storage unit 7-3. Further, the machine learning unit 5 determines the causal relationship with the unsupervised data, and the output data including the probability is stored in the storage unit 7-4. Further, the text analysis interface unit 6 determines whether or not there is a causal relationship for the text data to be analyzed by using the output data stored in the storage unit 7-4.

  The noun pair extraction unit 2 inputs text data from the storage unit 7-1, divides the data into morphemes by morphological analysis, and extracts noun pairs. Since the method for extracting the noun pair (extraction method by morphological analysis) is known, the description thereof is omitted here.

  The feature extraction unit 3 extracts, from the noun pair extracted by the noun pair extraction unit 2, features on the structure (syntax structure) of text data including the noun pair and features of attributes of the noun pair. FIG. 2 is a block diagram showing a configuration of the feature extraction unit 3 shown in FIG. The feature extraction unit 3 includes a tree structure generation unit 31, a superordinate concept extraction unit 32, a PSE (Preorder String Expression) generation unit 33, a triplet generation unit 34, and a thesaurus storage unit 35. The tree structure generation unit 31 generates a tree structure for a sentence of text data including a noun pair based on the syntax analysis result. Since this parsing method is known, the description is omitted here. For details, see the literature of “Kudo et al.,“ Dependency Analysis by Chunking Stage Application ”, Information Processing, Vol. 43, No. 6, pp. 1834-1842 (2002)”. The PSE generation unit 33 generates a PSE that can express the tree structure from the tree structure generated by the tree structure generation unit 31. The superordinate concept extraction means 32 searches the thesaurus storage unit 35 and extracts superordinate concepts of each noun pair. The ternary set generating means 34 uses the PSE generated by the PSE generating means 33 and the superordinate concept extracted by the superordinate concept extracting means 32 to determine the structural features of the text data and the features of the noun pair attributes. Generate the triplet shown.

  The tagging unit 4 selects, by the user's operation, a ternary group that specifies the presence or absence of a causal relationship from the ternary group generated by the feature extraction unit 3, and the causal relationship is selected for the selected ternary group. The presence / absence is designated (tagged) and stored as supervised data in the storage unit 7-3. The ternary set not tagged by the tagging unit 4 is stored in the storage unit 7-2 as unsupervised data. In this case, the ternary group is divided into unsupervised data and supervised data, a large amount of unsupervised data exists, and a small amount of supervised data exists.

  The machine learning unit 5 reads a large amount of unsupervised data in which the causal relationship is not specified from the storage unit 7-2, reads out a small amount of supervised data in which the causal relationship is specified from the storage unit 7-3, and has a teacher. Based on the structural characteristics of text data in the data and the characteristics of the attributes of the noun pair, the probability of the causal relationship between the supervised data and the unsupervised data is calculated and stored as output data 7-4.

  The text analysis interface unit 6 inputs text data to be subjected to text analysis, and determines whether or not a noun pair having a causal relationship exists using the output data stored in the storage unit 7-4. .

[Operation]
Next, the operation of the causal relationship knowledge extraction apparatus 1 shown in FIG. 1 will be described. First, the noun pair extraction unit 2 extracts the noun pair by dividing the text data input from the storage unit 7-1 into morphemes by morphological analysis. Here, a combination of all nouns appearing in one sentence is defined as a noun pair.

  Then, the feature extraction unit 3 analyzes the dependency relationship between clauses in one sentence of the text data including the noun pair by syntactic analysis for the noun pair extracted by the noun pair extraction unit 2. Since the syntax analysis method is known, the description is omitted here. Further, the feature extraction unit 3 extracts a syntactic structure between the noun pairs based on the syntax analysis result. For example, in the sentence “Atherosclerosis leads to stroke.”, The dependency relationship between the noun pair “arteriosclerosis” and “stroke” is a tree structure with the nodes shown in FIG. 3 as nodes. It can be expressed by In FIG. 3, nodes surrounded by a square indicate that the parent is modified. For example, “arteriosclerosis” modifies “when it happens”. Similarly, in the sentence “hidden obesity leads to diabetes”, the tree structure of FIG. Shows the dependency relationship between. That is, the tree structure generation means 31 of the feature extraction unit 3 generates a tree structure as shown in FIG. 3 or FIG. 4 based on the syntax analysis result of the sentence including the noun pair.

  Then, the PSE generation means 33 of the feature extraction unit 3 generates a PSE from the tree structure. During the generation process, each node of the tree structure is separated into independent words (nouns, verbs, adjectives, adverbs, adjective verbs, conjunctions, etc.) and function words (particles, auxiliary verbs, etc.). For example, the node “arteriosclerosis” shown in FIG. 3 is separated into an independent word “arteriosclerosis” and a function word “ga”. Also, this separated function word is inserted into the tree structure as an upper node of the original node. FIG. 5 shows a new tree structure after separating function words from the tree structure shown in FIG. 3, that is, a tree structure having independent words and function words as nodes. The PSE generation unit 33 generates a PSE from a new tree structure from which function words are separated as shown in FIG. Then, the PSE generation unit 33 replaces each of the noun pairs with “noun 1” and “noun 2” for the generated PSE. Since a method for generating a PSE from a tree structure is known, description thereof is omitted here. For details, refer to the document “Fablizio Luccio et al.,“ Exact Rooted Subtree Matching in Subscriber Time ”, Technical Report TR-01-14 (2001)”.

For example, the PSE generation unit 33 separates the function words from the tree structure of FIG. 3 generated by the tree structure generation unit 31 to generate the new tree structure of FIG. 5, and the PSE shown below from the new tree structure Is generated.
PSE = {“connect”, “to”, “get up”, “ga”, “arteriosclerosis”, 0, 0, 0, 0, “to”, “stroke”, 0, 0, 0}
Here, the PSE generating unit 33 replaces the target of the noun pair with “noun 1” and “noun 2” in the order of appearance for the generated PSE. For example, in the above example, it is as follows.
PSE = {“connect”, “to”, “get up”, “ga”, “noun 1”, 0, 0, 0, 0, “ni”, “noun 2”, 0, 0, 0}
Similarly, when function words are separated from the tree structure of FIG. 4 and a PSE is generated from the new tree structure and replaced with “noun 1” and “noun 2”, the result is as follows.
PSE = {“connected”, “ga”, “noun 1”, 0, 0, “ni”, “noun 2”, 0, 0, 0}
The PSE can be restored to the original tree structure by word order and element “0”.

  Returning to FIG. 3, the superordinate concept extraction means 32 extracts the features of the superordinate concept of the noun. Specifically, for a noun that is the target of a noun pair, an existing classification vocabulary table or the like stored in the thesaurus storage unit 35 is searched, and each superordinate concept is extracted. In this case, when the superordinate concept can be uniquely determined on the thesaurus, the superordinate concept is extracted, and when the superordinate concept cannot be uniquely determined such as having a plurality of attributes, the notation itself is extracted as the superordinate concept. In addition, since no superordinate concept can be extracted when there is no superordinate concept, the noun itself is treated as a superordinate concept. For example, since the noun pairs “stroke” and “arteriosclerosis” both have a superordinate concept of “disease / physical condition”, the superordinate concept extraction means 32 extracts “disease / physical condition” as a superordinate concept.

Then, the ternary set generation unit 34 generates a ternary set by using the PSE generated by the PSE generation unit 33 and the superordinate concept extracted by the superordinate concept extraction unit 32. For example, the ternary set generation means 34 is as follows for the noun pair “arteriosclerosis” “stroke” with the sentence “arteriosclerosis occurs, which leads to stroke”.
Ternary set = <{sickness / physical condition}, {sickness / physical condition}, {“connected”, “to”, “get up”, “ga”, “noun 1”, 0, 0, 0, 0, “to” , "Noun 2", 0, 0, 0}>
Thus, the triplet is a characteristic of the structure of the text data {“connected”, “to”, “occurs”, “ga”, “noun 1”, 0, 0, 0, 0, “to” , “Noun 2”, 0, 0, 0} and {sickness / physical condition}, {sickness / physical condition} which are the characteristics of the attributes of the noun pair.

  Then, the tagging unit 4 tags whether or not the expression of the sentence in the noun pair has a causal relationship with respect to a part of the ternary set generated by the feature extraction unit 3 by a user operation. And the tagged three-term set is stored as supervised data in the storage unit 7-3. The ternary tuples that are not tagged are stored in the storage unit 7-2 by the feature extraction unit 3 as unsupervised data. Thereby, a large amount of unsupervised data and a small amount of supervised data are generated.

ここで、ｔ_ｉは３項組、Ｐ（ｃ_１｜ｔ_ｉ）は３項組ｔ_ｉが因果関係を持つ確率、Ｐ（ｃ_０｜ｔ_ｉ）は３項組ｔ_ｉが因果関係を持たない確率、Ｐ（ｔ_ｉ）は３項組ｔ_ｉが出現する確率、Ｐ（ｃ_１）は因果関係を持つ３項組が出現する確率、Ｐ（ｃ_０）は因果関係を持たない３項組が出現する確率、Ｐ（ｔ_ｉ｜ｃ_ｊ）はクラスｃ_ｊのときに３項組ｔ_ｉが出現する確率をそれぞれ示す。式（１）で表された確率Ｐ（ｃ_ｊ｜ｔ_ｉ）の値が大きいクラスｃ_ｊ（ｃ_０またはｃ_１）を、因果関係の有無の判定結果とする。すなわち、Ｐ（ｃ_０｜ｔ_ｉ）＞Ｐ（ｃ_１｜ｔ_ｉ）のときは３項組ｔ_ｉについて因果関係無しを判定し、Ｐ（ｃ_０｜ｔ_ｉ）＜Ｐ（ｃ_１｜ｔ_ｉ）のときは３項組ｔ_ｉについて因果関係有りを判定する。同じ確率であるときは予め設定された判定結果とする。 Then, the machine learning unit 5 reads the unsupervised data from the storage unit 7-2 and the supervised data from the storage unit 7-3. The probability that the ternary set of unsupervised data has a causal relationship is calculated using the EM algorithm. Calculated by the machine learning used. In this case, the following formulas (1) and (2) are used. Based on this probability, the presence or absence of a causal relationship can be determined. The probability that the ternary group has a causal relationship and the probability that it does not have it are expressed by the following equations.

Here, t _i is a ternary set, P (c ₁ | t _i ) is a probability that the ternary set t _i has a causal relationship, and P (c ₀ | t _i ) is a ternary set t _i has a causal relationship. P (t _i ) is the probability that a ternary set t _i will appear, P (c ₁ ) is the probability that a ternary set with causality will appear, and P (c ₀ ) is the 3 term that has no causal relationship The probability that a pair appears, P (t _i | c _j ), indicates the probability that a ternary set t _i will appear when class c _j is used. A class c _j (c ₀ or c ₁ ) having a large value of the probability P (c _j | t _i ) expressed by the expression (1) is set as a determination result of the causal relationship. That is, when P (c ₀ | t _i )> P (c ₁ | t _i ), it is determined that there is no causal relationship for the ternary set t _i , and P (c ₀ | t _i ) <P (c ₁ | t _{In the case of i} ), it is determined that there is a causal relationship for the ternary set t _i . When the probability is the same, the determination result is set in advance.

ここで、ＣＰ_ｔｉは３項組ｔ_ｉに含まれる２つの名詞間の構文構造、すなわちテキストデータの構文構造を、ＳＰ_ｔｉは３項組ｔ_ｉに含まれる名詞ペアを、Ｐ（ＣＰ_ｔｉ｜ｃ_ｊ）はクラスｃ_ｊのときに３項組ｔ_ｉに含まれる２つの名詞間の構文構造ＣＰ_ｔｉが出現する確率を、Ｐ（ＳＰ_ｔｉ｜ｃ_ｊ）はクラスｃ_ｊのときに３項組ｔ_ｉに含まれる名詞ペアＳＰ_ｔｉが出現する確率をそれぞれ示す。 The probability P (t _i | c _j ) of the occurrence of a ternary set in class c _j shown in equation (1) is expressed by the following equation.

Here, _{CP ti} syntax structure between two nouns included in 3-tuple _{t i,} i.e. the syntactic structure of the text data, the noun pair _{SP ti} is contained in the 3-tuple _{t i,} P _{(CP ti} | c _j ) is the probability of the occurrence of a syntactic structure CP _ti between two nouns included in the ternary set t _i for class c _j , and P (SP _ti | c _j ) is ternary for class c _j indicating the probability of noun pair _{SP ti} included in the set _{t i} appears respectively.

  The machine learning unit 5 performs machine learning by the EM algorithm using these equations (1) and (2). The EM algorithm is a process of performing the iterative learning that maximizes the likelihood for incomplete data whose internal state is unknown, and estimating the internal state. Since the EM algorithm is known, the description is omitted here. For details, see "Kamel Nigam et al.," Text Classification from Labeled and Unlabeled Document using EM ", machine learning, Vol. 39, No. 2/3, pp. 103-134 (2000). I want.

Next, the process of the machine learning unit 5 will be described in detail. FIG. 6 is a flowchart for explaining the processing of the machine learning unit 5. First, the machine learning unit 5 targets the supervised data read from the storage unit 7-3, and the initial of the class c _j to which the triplet t _i belongs (c ₁ if causal, c ₀ if not) The probability P (c ₁ | t _i ) is calculated by the following equation (step S6-1). This step is the E step in the EM algorithm.

ここで、｜ＣＰ｜は名詞ペアを含むテキストデータの構文構造の総数を、｜ＳＰ｜は名詞ペアの総数を、｜Ｔ｜は３項組の総数を示す。Ｎ（ＳＰ，ｔ_ｋ）は３組項ｔ_ｋに名詞ペアを含むか否かを表す関数を示し、含むときのみ１の値となる。ｓｉｍ’（ＣＰ_ｔｉ，ＣＰ_ｔｋ）は名詞ペアを含むテキストデータの構文構造の類似性を示し、以下に示す式（６）による類似度の計算により求められる。類似度が０．５より大きいときは、ｓｉｍ’（ＣＰ_ｔｉ，ＣＰ_ｔｋ）を式（６）の計算結果であるｓｉｍ（ＣＰ_ｔｉ，ＣＰ_ｔｋ）とし、０．５以下のときは、ｓｉｍ’（ＣＰ_ｔｉ，ＣＰ_ｔｋ）を０とする。 Next, the machine learning unit 5 targets a set of all text data including unsupervised data read from the storage unit 7-2 and supervised data read from the storage unit 7-3, under the class c _j . in, probability _{P CP ti} (syntactic structure of the text data including the two nouns included in 3-tuple _{t i)} appears _| nouns included in _(CP ti _c j), and _{SP ti} (3-tuple _{t i} probability pair) appears _{P (SP} ti | _{c j)} calculating each according to the following equation (step S6-2). This step is the M step in the EM algorithm.

Here, | CP | indicates the total number of syntactic structures of text data including noun pairs, | SP | indicates the total number of noun pairs, and | T | N (SP, t _k ) represents a function indicating whether or not a noun pair is included in the triplet t _k, and takes a value of 1 only when included. Sim ′ (CP _ti , CP _tk ) indicates the similarity of the syntax structure of text data including noun pairs, and is obtained by calculating the similarity according to the following equation (6). When the similarity is greater than _{0.5, sim '(CP ti, CP} tk) which is a calculation result of formula _{(6) sim (CP ti,} CP tk) and, when less than 0.5, sim' _Let (CP _ti , CP _tk ) be 0.

ｗｃ（ｐ_ｉ）は、ＰＳＥを示すｐ_ｉに出現する名詞ペアの２つの名詞の属性及び要素「０」以外の単語数を示す。ｃｏｍ（Ｐ_１，Ｐ_２）は、そのうちのＰＳＥの構造を考慮した共通単語数を示す。例えば、図３及び図４の木構造から機能語を分離して得た２つのＰＳＥであるＰ_１，Ｐ_２では、ｗｃ（ｐ_１）＝５、ｗｃ（ｐ_２）＝３、単語「つながります」「が」「に」が共通単語であるから共通単語数ｃｏｍ（Ｐ_１，Ｐ_２）＝３となり、ｓｉｍ（Ｐ_１，Ｐ_２）＝６／８＝０．７５となる。 The similarity sim (P ₁ , P ₂ ) of the syntactic structure between the above-mentioned noun pairs is calculated by the following equation.

wc (p _i) shows two of the number of words other than the attributes and elements "0" of the noun of noun pairs that appear to p _i indicating the PSE. com (P ₁ , P ₂ ) indicates the number of common words in consideration of the PSE structure. For example, in P ₁ and P ₂ which are two PSEs obtained by separating the function words from the tree structure of FIGS. 3 and 4, wc (p ₁ ) = 5, wc (p ₂ ) = 3, and the word “connection” Since “Masu”, “GA” and “NI” are common words, the number of common words is com (P ₁ , P ₂ ) = 3, and sim (P ₁ , P ₂ ) = 6/8 = 0.75.

そして、機械学習部５は、式（７）の結果を利用して、因果関係を持つまたは持たない３項組３項組ｔ_ｉが出現する確率Ｐ（ｃ_ｊ）を以下の式により計算する（ステップＳ６−４）。

ここで、｜ｃ｜は、分類すべきクラスの数を示し、この場合は２である。 In step 6-2, the machine learning unit 5 calculates the probability P (CP _ti | c _j ) that the syntactic structure CP _ti appears and the probability P (SP _ti | c _j ) that the noun pair SP _ti appears in the formula (4). After calculating according to (5), using these results, the expected value of the probability P (c _j | t _i ) of the ternary group t _i having or not having a causal relationship is calculated by the following equation (step S6-3).

Then, the machine learning unit 5 calculates the probability P (c _j ) of occurrence of the ternary group ternary group t _i having or not having the causal relationship using the result of the equation (7) by the following equation. (Step S6-4).

Here, | c | indicates the number of classes to be classified, and is 2 in this case.

The machine learning unit 5 uses the result of Expression (8) to change the probability P (c _j ) of occurrence of a ternary set t _i having or not having a causal relationship and a certain threshold (for example, 1 .0 × 10 ⁻³ ) (step S6-5). When the amount of change in the probability P (c _j ) is equal to or greater than a certain threshold value, the process returns to step 6-2, and the unsupervised data and the new probability P (c _j | t _i ) calculated in step 6-3 are used. For a set of all text data including supervised data, the probability P (CP _ti | c _j ) that a syntax structure appears and the probability P (SP _ti | c _j ) that a noun pair appears are _expressed by the above-described equations ( 4) Calculate according to (5), respectively (step S6-2). Then, steps 6-2 to 6-5 are repeated until the change amount of the probability P (c _j ) becomes smaller than a certain threshold value in step 6-5. Then, when the amount of change in the probability P (c _j ) is smaller than a certain threshold, the machine learning unit 5 determines whether the last calculated ternary set t _i has or does not have a causal relationship (3 terms). Probability of existence of causal relationship in set t _i ) P (c _j | t _i ), probability P of occurrence of syntax structure CP _ti (CP _ti | c _j ), and probability P of appearance of noun pair SP _ti (SP _ti | c _j ).

In Table 1, the original sentence (part) designated as a ternary group having a causal relationship by the user and the ternary group calculated by the processing of the machine learning unit 5 shown in FIG. 6 have a causal relationship. The probability P (c ₁ | t _i ) is shown. This is a 16-program “Kyoto Health” featured in the topic of circulatory system. The closed caption 2180 sentence used in the program is used as text data, and 1495 ternary groups are generated. This is an example of a related experiment. Tagging of presence / absence of causality is performed by user operation to 149 ternary groups of one program randomly selected from 16 programs, and the machine learning unit 5 uses the probability P (c ₁ | T _i ) is calculated. According to Table 1, since the respective probabilities P (c ₁ | t _i ) are close to 1, the causal relationship knowledge extraction apparatus 1 determines P (c ₁ | t _i )> P (c ₀ | t _i ). to, 3-tuple t _i shown in Table 1 can be determined as having a causal relationship.

ここで、確率Ｐ（ｃ₁｜ＣＰ_ｔｉ）は、３項組ｔ_ｉに含まれる構文特徴ＣＰ_ｔｉが因果関係を持つ確率を示す。また、確率Ｐ（ｃ_０｜ＣＰ_ｔｉ）は、３項組ｔ_ｉに含まれる構文特徴ＣＰ_ｔｉが因果関係を持たない確率を示す。因果関係知識抽出装置１は、この値を用いて予め設定された値と比較する等により、因果関係を持つときの特徴的な構文構造を判定することができる。 Furthermore, the machine learning unit 5 calculates the probability P (c _j | CP _ti ) using the following equation from the probability P (CP _ti | c _j ) finally calculated in step 6-2 shown in FIG. calculate.

Here, the probability P (c ₁ | _{CP ti)} indicates the probability that the syntax characteristics _{CP ti} has a causal relation included in the 3-tuple _{t i.} Also, the probability _P (c 0 | _{CP ti)} indicates the probability that the syntax characteristics _{CP ti} has no causal relationships included in the 3-tuple _{t i.} The causal relationship knowledge extraction apparatus 1 can determine a characteristic syntax structure having a causal relationship by using this value and comparing it with a preset value.

ここで、確率Ｐ（ｃ_１｜ＳＰ_ｔｉ）は、３項組ｔ_ｉに含まれる名詞ペアＳＰ_ｔｉが因果関係を持つ確率を示す。確率Ｐ（ｃ_０｜ＳＰ_ｔｉ）は、３項組ｔ_ｉに含まれる名詞ペアＳＰ_ｔｉが因果関係を持たない確率を示す。因果関係知識抽出装置１は、この値を用いて予め設定された値と比較する等により、因果関係を持つ名詞ペアを判定することができる。 Further, the machine learning unit 5 calculates the probability P (c _j | SP _ti ) from the probability P (SP _ti | c _j ) finally calculated in Step 6-2 shown in FIG. calculate.

Here, the probability _P (c 1 | _{SP ti)} indicates the probability that the noun pair _{SP ti} has a causal relation included in the 3-tuple _{t i.} The probability P (c ₀ | SP _ti ) indicates the probability that the noun pair SP _ti included in the ternary set t _i has no causal relationship. The causal relationship knowledge extraction apparatus 1 can determine a noun pair having a causal relationship by using this value and comparing it with a preset value.

  Next, the processing of the text analysis interface unit 6 will be described in detail. FIG. 7 is a flowchart for explaining the processing of the text analysis interface unit 6. First, the text analysis interface unit 6 inputs text data to be subjected to text analysis, divides it into morphemes by morphological analysis, and extracts noun pairs (step S7-1). The method for extracting the noun pair is the same as that by the noun pair extracting unit 2 shown in FIG.

  Then, the text analysis interface unit 6 extracts the structural features of the text data including the noun pairs and the attribute features of the noun pairs for all the extracted noun pairs (step S7-2). Specifically, by syntactic analysis, dependency relationships between clauses in text data including noun pairs are analyzed to generate a tree structure, a PSE is generated from the tree structure, and a superordinate concept of noun pairs using a thesaurus To extract. Then, using the PSE and the superordinate concept, a ternary set indicating the structural characteristics of the text data and the characteristics of the attributes of the noun pairs is generated. In this way, the text analysis interface unit 6 generates ternary groups for all noun pairs. The method for generating the triplet is the same as that by the feature extraction unit 3 shown in FIGS.

Then, the text analysis interface unit 6 uses the probability P (c _j | CP _ti ) shown in the equation (9) having the same syntax feature as the ternary syntax feature CP _ti in the noun pair to be analyzed, A search is performed from the output data stored in the storage unit 7-1. Then, the probability P (c ₁ | CP _ti ) of the search result is compared with a preset threshold value (for example, 0.5) (step S7-4), and this probability P (c ₁ | CP _ti ) is greater than the threshold value. When it is larger, it is determined that the noun pair has a causal relationship (step S7-4). On the other hand, when the probability P (c _j | CP _ti ) is less than or equal to the threshold value, it is determined that the noun pair has no causal relationship (step S7-5). Steps 7-4 to 7-6 are performed for all noun pairs (step S7-3).

In the flowchart shown in FIG. 7, the presence / absence of a causal relationship with respect to a noun pair is compared with the probability P (c ₁ | CP _ti ) shown in Expression (9) in step 7-4 and a preset threshold value. The probability P (c ₁ | t _i ) shown in the equation (7) is used instead of the probability P (c ₁ | CP _ti ) shown in the equation (9). Also good.

  As described above, according to the causal relationship knowledge extraction apparatus 1, a large amount of unsupervised data in which the causal relationship is not specified and a small amount of supervised data in which the causal relationship is specified are generated, and the EM algorithm is used. The probability of the presence or absence of causality was calculated for all noun pairs by machine learning. By using this probability, it is possible to determine the presence or absence of a causal relationship. That is, it becomes possible to extract a noun pair having a causal relationship and a sentence structure having a causal relationship. In addition, it is possible to determine whether or not there is a causal relationship with unknown text data.

  In addition, for example, when such a causal relationship knowledge extraction device 1 is applied to a broadcast receiving device that receives broadcast waves, the broadcast receiving device can provide reliable information sources such as closed captions transmitted by broadcasting. By constantly monitoring and analyzing the text data, the causal knowledge data can be automatically accumulated. Thereby, it becomes possible to automatically learn the knowledge about the causal relationship that a human has. Therefore, by using the causal knowledge learned in this way, it is possible to construct a system capable of automatically responding to advanced questions of the “why” type.

  The causal relationship knowledge extraction apparatus 1 is configured by a computer including a volatile storage medium such as a CPU and a RAM, a non-volatile storage medium such as a ROM, an interface, and the like. The functions of the noun pair extraction unit 2, the feature extraction unit 3, the tagging unit 4 and the machine learning unit 5 included in the causal knowledge extraction apparatus 1 are realized by causing the CPU to execute a program describing these functions. Is done. These programs can also be stored and distributed in a storage medium such as a magnetic disk (floppy disk, hard disk, etc.), optical disk (CD-ROM, DVD, etc.), semiconductor memory, or the like.

  The present invention has been described with reference to the embodiment. However, the present invention is not limited to the above embodiment, and various modifications can be made without departing from the technical idea thereof. For example, the causal relationship knowledge extraction device 1 shown in FIG. 1 is configured by one computer device, but is not limited to this, and only the text analysis interface unit 6 is provided in another computer device, and the network You may comprise so that it may connect via. Further, the noun pair extraction unit 2 and the feature extraction unit 3 may be provided in different computer devices for each processing unit and connected via a network, or the storage units 7-1 to 7-4 or the like. A part of the above may be provided in another computer apparatus.

It is a block diagram which shows the structure of the causal relationship knowledge extraction apparatus by embodiment of this invention. It is a block diagram which shows the structure of the feature extraction part of FIG. It is a figure which shows the production example 1 of a tree structure. It is a figure which shows the production example 2 of a tree structure. It is a figure which shows the example of the deformed tree structure. It is a flowchart explaining the process of a machine learning part. It is a flowchart explaining the process of a text analysis interface part.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Causal relationship knowledge extraction device 2 Noun pair extraction part 3 Feature extraction part 4 Tagging part 5 Machine learning part 6 Text analysis interface part 7 Memory | storage part 31 Tree structure production | generation means 32 High-order concept extraction means 33 PSE production | generation means 34 3 term generation Means 35 Thesaurus storage unit

Claims (7)

It is a device that extracts the presence or absence of causal relationships as knowledge about noun pairs included in text data,
A noun pair extraction unit for extracting noun pairs included in the text data;
For text data including noun pairs extracted by the noun pair extraction unit, a feature extraction unit that extracts the syntax structure of the text data and the attributes of the noun pairs as feature data;
A tagging unit that gives a tag of presence or absence of a causal relationship in the noun pair for a part of the feature data extracted by the feature extraction unit;
Feature data to which a tag is given by the tagging unit is stored as supervised data, and feature data to which no tag is given by the tagging unit is stored as unsupervised data;
A causal relationship determination unit for determining the presence or absence of a causal relationship in a noun pair included in the unsupervised data based on supervised data and unsupervised data stored in the storage unit. Relational knowledge extraction device. In the causal relationship knowledge extraction device according to claim 1,
The feature extraction unit
For text data including noun pairs, a tree structure generating means for generating a tree structure indicating the dependency relationship of the noun pairs;
PSE generation means for separating nodes constituting the tree structure of the text data generated by the tree structure generation means into independent words and function words and generating PSE (Preorder String Expression) indicating the syntax structure of the text data; ,
Superordinate concept extracting means for extracting superordinate concepts of the noun pair;
A ternary set generation unit that generates a ternary set as feature data using the PSE generated by the PSE generation unit and the higher level concept extracted by the higher level concept extraction unit; Relational knowledge extraction device. In the causal relationship knowledge extraction device according to claim 1 or 2,
The causal relationship determination unit calculates the probability of the presence or absence of a causal relationship in the data with respect to supervised data stored in the storage unit by an EM algorithm, and the supervised data and unsupervised data stored in the storage unit Calculating the probability of occurrence of a syntactic structure having the same or a similarity equal to or higher than a predetermined value, and for the supervised data and the unsupervised data stored in the storage unit, the same attribute as the noun pair included in the data Based on these probabilities, the probability of occurrence of causality in the feature data, the probability of causality in the syntactic structure included in the feature data, and the causality in the noun pair included in the feature data are calculated based on these probabilities. A causal relationship knowledge extraction apparatus characterized by calculating the probability of presence or absence of a relationship and determining the presence or absence of a causal relationship. In the causal relationship knowledge extraction device according to claim 3,
Further, new text data is input, a noun pair included in the text data is extracted, a syntactic structure of the text data and an attribute of the noun pair are extracted as feature data, and the causal relationship calculated by the causal relationship determination unit A causal relationship knowledge extraction device comprising a text determination unit that determines the presence or absence of a causal relationship in a noun pair included in the new text data based on the probability of the presence or absence. In the causal relationship knowledge extraction device according to claim 4,
The text determination unit uses the extracted feature data, and based on the probability of the presence or absence of a causal relationship in the syntax structure included in the feature data calculated by the causal relationship determination unit, the syntax structure included in the extracted feature data The causal relationship knowledge extraction apparatus characterized in that the probability of the presence or absence of a causal relationship is calculated and the presence or absence of the causal relationship is determined based on the probability. In the causal relationship knowledge extraction device according to claim 4,
The text determination unit uses the extracted feature data, and based on the probability of the presence or absence of a causal relationship in the noun pair included in the feature data calculated by the causal relationship determination unit, the noun pair included in the extracted feature data The causal relationship knowledge extraction apparatus characterized in that the probability of the presence or absence of a causal relationship is calculated and the presence or absence of the causal relationship is determined based on the probability. A causal relationship knowledge extraction program by a device that extracts the presence or absence of a causal relationship between noun pairs included in text data, the computer constituting the device,
Processing to extract noun pairs contained in text data;
Processing for extracting the syntactic structure of text data including the noun pair and the attributes of the noun pair as feature data;
A process for assigning a tag of presence or absence of a causal relationship in a noun pair for a part of the extracted feature data;
A process of determining the presence or absence of a causal relationship in a noun pair included in unsupervised data based on supervised data that is feature data to which the tag is attached and unsupervised data that is feature data to which no tag is attached. A causal knowledge extraction program to be executed.

Images