JP6482084B2 - Grammar rule filter model learning device, grammar rule filter device, syntax analysis device, and program - Google Patents

Description

  The present invention relates to a grammatical rule filter model learning device, a grammatical rule filter device, a syntax analysis device, and a program for performing sentence syntax analysis.

  Conventionally, a statistical syntax analysis apparatus including a model learning apparatus and an analysis apparatus is known (Non-Patent Document 1). The model learning device outputs a grammar rule set and a model weight vector from a set of sentence-syntax tree pairs. The analysis device receives the grammar rule set and the weight vector output from the model learning device, and outputs a syntax tree having the highest score for the sentence to be analyzed.

Jenny Rose Finkel, Alex Kleeman, and Christopher D Manning.Efficient, feature-based, conditional random field parsing.In ACL, volume 46, pages 959-967, 2008.

  However, in the above-described conventional statistical syntax analysis apparatus, a large number of solution candidates can be considered by combining grammatical rules, so scoring all solution candidates requires a long calculation time.

  In the present invention, a grammar rule filter model learning device, a grammar rule filter device, a syntax analysis, which can reduce the grammar rules in order to perform sentence syntax analysis at high speed, is made to solve the above problems. An object is to provide an apparatus and a program.

  To achieve the above object, the grammatical rule filter model learning device according to the present invention is based on learning data consisting of a set of pairs of a sentence x and a syntax tree y representing a syntax analysis result of the sentence x. For each of the data syntax trees y, a syntax tree simplification that generates a simplified syntax tree y￣ by converting each label different from a label representing a word and a label representing a part of speech into a virtual label. And a set R￣ of grammatical rules r￣ extracted from the simplified syntax tree y￣ for each of the learning data syntax trees y and the grammar rules r 変 換 converted using the virtual labels For each of the pair of learning data, a simplified grammar rule set extraction unit for generating a conversion function for converting the grammar rule to a pre-conversion grammar rule r using the virtual label And simplification Score based on the feature vector f￣ (x, y￣) extracted from the syntax tree y￣ and the weight vector w￣ corresponding to the feature vector f￣ (x, y ，). And a simplified syntax tree weight vector update unit that updates the weight vector w￣ according to the score.

  The grammatical rule filter device according to the present invention is simplified by converting each of a label representing a word and a label different from a label representing a part of speech included in a syntax tree y representing a sentence parsing result into a virtual label. Corresponding to a set R￣ of grammatical rules r￣ extracted from the parsed syntax tree y￣, and a feature vector fｘ (x, y 特 徴) extracted from the sentence x and the simplified syntax tree y￣. For each of the grammar rule r 予 め included in the set R￣ of the grammar rule r￣, a score calculation unit that calculates a reference score based on the previously learned weight vector w￣ and the input sentence x A score that includes the grammar rule r￣ that represents the weight vector w 規則, the input sentence x, and the parsing result of the input sentence x, and that is calculated based on the simplified syntax tree y￣; , Depending on the reference score, A grammar rule adding unit that converts the legal rule r￣ to a grammar rule r before conversion using the virtual label and adds the grammar rule r to the set R ′ of grammar rules r not using the virtual label; It is configured to include.

  The syntax analysis device according to the present invention is a feature quantity extracted from the set R ′ of the grammar rules r obtained by the grammar rule filter device, the sentence x, and the syntax tree y representing the result of the syntax analysis of the sentence x. Based on the previously learned weight vector w corresponding to the vector f (x, y) and the input sentence x, the grammar rule r included in the set R ′ of the grammar rules r is used to determine the input sentence x. The input sentence having a maximum score calculated based on the feature vector f (x, y) extracted from the syntax tree y representing the parsing result and the input sentence x, and the weight vector w It includes an analysis unit that searches for a syntax tree y that represents the result of parsing x.

  The program of the present invention is a program for causing a computer to function as each unit constituting the grammar rule filter model learning device, the grammar rule filter device, or the syntax analysis device.

  As described above, according to the grammatical rule filter model learning apparatus and program of the present invention, a simplified syntax tree y￣ is generated for each of the syntax trees y of learning data, and a set of grammar rules r￣ is set. R￣ is extracted, and for each pair of learning data, a score is calculated based on the feature vector f 特 徴 (x, y￣) of the pair and a weight vector w￣, and the score is calculated according to the score. By updating the weight vector w￣, it is possible to learn a grammar rule filter that can reduce grammar rules in order to perform sentence parsing at high speed.

  According to the grammatical rule filter device and program of the present invention, it corresponds to a set R￣ of grammatical rules r￣ extracted from a simplified syntax tree y￣ and a vector f 量 (x, y￣) of feature quantities. A reference score is calculated based on the weight vector w ベ ク ト ル learned in advance and the input sentence x, and the grammar rule r￣ representing the parsing result of the input sentence x is calculated for each of the grammar rules r￣. The grammar rule r 構 文 is converted into the grammar rule r before conversion according to the score calculated based on the parsed tree y￣ and the reference score, and added to the set R ′ of grammar rules r. As a result, grammar rules can be reduced in order to perform sentence parsing at high speed.

  According to the syntax analysis apparatus and program of the present invention, sentence syntax analysis can be performed at high speed.

It is a figure which shows an example of the sentence used as input. It is a figure which shows an example of a syntax tree. It is a block diagram which shows the functional structure of the syntax analysis apparatus which concerns on embodiment of this invention. It is a block diagram which shows the structure of the learning part of the syntax analyzer which concerns on embodiment of this invention. It is a figure which shows an example of the set of grammar rules. It is a block diagram which shows the structure of the rule filter model learning part of the syntax analysis apparatus which concerns on embodiment of this invention. It is a figure which shows an example of the simplified syntax tree. It is a figure which shows an example of the set of the grammar rules extracted from the simplified syntax tree. It is a block diagram which shows the structure of the rule filter part of the syntax analyzer which concerns on embodiment of this invention. It is a flowchart figure of the rule filter model learning process routine in the syntax analysis apparatus which concerns on embodiment of this invention. It is a flowchart figure of the syntax analysis process routine in the syntax analysis apparatus which concerns on embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

<Outline of the present embodiment>
First, an outline of an embodiment of the present invention will be described. The present embodiment relates to a statistical syntax analysis apparatus used in fields such as natural language processing. The syntax analysis device is a device that receives a word string as an input and outputs a tree structure representing the syntax structure of the word string. An input word string (with an index) means a sentence as shown in FIG. The syntax tree for this is as shown in FIG. The syntax tree is hierarchically constructed using a word label, a part-of-speech label (NN, V, DT, etc.) that appears only immediately above the word label, and a phrase label (S, VP, NP, etc.).

  The statistical parsing apparatus uses a set (learning data) of a sentence and its correct syntax tree pair to build a grammar rule set and a statistical model for scoring the sentence and syntax tree pair in advance. The grammar rules here are:

NN → I

A rule with one word as a child, one part of speech label as a parent, and

S → NP VP
VP → V NP

Such that one or more parts of speech or phrase labels are children and one phrase label is the parent.

Say the law. When a sentence is input to the syntax analysis apparatus, a syntax tree having the highest score is constructed and output using a grammar rule set and a statistical model. The output syntax tree is used in application tasks in the field of natural language processing such as machine translation.

  The syntax analysis apparatus according to the present embodiment includes a grammar rule filter model learning unit and a grammar rule filter unit. This device is a device in which a grammatical rule filter model learning unit and a grammatical rule filter unit are coupled to a conventional statistical syntax analysis device.

  The rule filter model learning unit receives a set of sentence and syntax tree pairs and a grammar rule set output from a conventional model learning device, and inputs a new grammar rule set and a new grammar rule as elements of the grammar rule set. A conversion function, a weight vector, and a super parameter to be converted into are output.

  The rule filter unit receives a new grammar rule set, a conversion function, a weight vector, and a super parameter that are output from the rule filter model learning device, and outputs a filtered grammar rule set for the input sentence.

<Configuration of syntax analysis apparatus according to embodiment of the present invention>
Next, the configuration of the syntax analysis apparatus according to the embodiment of the present invention will be described. As shown in FIG. 3, the syntax analysis apparatus 100 according to the present embodiment stores a CPU, a RAM, a rule filter model learning process routine, which will be described later, and a program for executing the syntax analysis process routine and various data. And a computer including a ROM. Functionally, the syntax analysis apparatus 100 includes an input unit 10, a calculation unit 20, and an output unit 90 as shown in FIG.

  The input unit 10 uses a set of sentence and syntax tree pairs as learning data.

Accept. However, (x, y) is a pair of sentence x and its correct syntax tree y.

  Further, the input unit 10 receives an input sentence x that is a target of syntax analysis.

  The arithmetic unit 20 includes a learning unit 22, a grammar rule storage unit 24, a rule filter model learning unit 26, a simple grammar rule storage unit 28, a rule filter unit 30, and an analysis unit 32.

  Based on the learning data, the learning unit 22 learns a set of grammar rules for performing chart-type syntax analysis and a weight vector corresponding to a vector of feature amounts that can be extracted from the sentence x and the grammar rule r.

  Here, chart type syntax analysis will be described.

  A method of constructing a syntax tree of the sentence x using the grammar rule set R for the sentence x is called chart type parsing. Statistical parsing is

(1)

The chart type parsing accomplishes this. here

Is a candidate set of syntax trees that can be constructed from the grammar rule set R for the sentence x, r is the grammar rule used to construct the syntax tree y, and f (x, r) can be extracted from the sentence x and the rule r A function that returns a vector of feature amounts, w represents a weight vector composed of weights for weighting each feature amount. Further, f (x, y) is the sum of feature quantity vectors f (x, r) that can be extracted from each rule r and sentence x included in the syntax tree y. However, r is not a normal grammar rule,

As shown in Fig. 2, the index of the partial word string of x covered by the rule is given on the syntax tree (the rule in the above example is used at the position covering a telescope in Fig. 2). Show). In the following, it is assumed that the grammatical rule r used to construct the syntax tree y is given such an index of partial word strings.

  As shown in FIG. 4, the learning unit 22 includes a grammar rule set extraction unit 220 and a weight vector update unit 222.

  The grammar rule set extraction unit 220 extracts a set R of grammar rules r from each of the syntax trees y of the learning data P and stores them in the grammar rule storage unit 24.

Specifically, the grammar rule r is extracted from all the syntax trees y _i (1 ≦ I ≦ | P |) included in the learning data P. A set of extracted grammar rules r is written as R. For example, a set of grammar rules shown in FIG. 5 is extracted from the two syntax trees shown in FIG.

For each pair (x _i , y _i ) of the learning data P, the weight vector update unit 222 features vector f (x _i , y) extracted from the sentence x _{i of the} pair and the syntax tree y _{i. i} ) and a weight vector w are used to calculate a score, and the weight vector w is updated according to the score.

  Specifically, first, all the values of the weight vector w are set to zero.

Then, for each pair (x _i , y _i ) εP, the syntax tree ^ y _i of the sentence x _i is obtained according to the following equation (2).

(2)

(2) As a result of seeking syntax tree ^ y _i by equation ^ when y _i ≠ y _i, sentence x _i, syntax tree y _i, and the syntax tree ^ y _i obtained in the above (2) The weight vector w is updated according to the following equation (3).

(3)

  The above procedure is repeated a finite number of times, and the weight vector update unit 222 stores the finally obtained weight vector w in the grammar rule storage unit 24.

  Based on the learning data P and the set R of grammar rules r stored in the grammar rule storage unit 24, the rule filter model learning unit 26 inputs a new grammar rule set R￣ and a new grammar rule. A conversion function, weight vector w￣, and superparameter γ to be converted into elements of the rule set are learned.

  As shown in FIG. 6, the rule filter model learning unit 26 includes a syntax tree simplification unit 260, a simplified grammar rule set extraction unit 262, and a simplified syntax tree weight vector update unit 264.

Based on the learning data P, the syntax tree simplification unit 260 converts each of the syntax tree y _i of the learning data P into a virtual label X with a label different from a label representing a word and a label representing a part of speech. Transform to produce a simplified syntax tree y _i ￣.

Specifically, for the syntax tree y _i , all labels except words and parts of speech are converted into virtual labels called X. This conversion is called syntax tree simplification. FIG. 7 shows a simplified example of the syntax tree in the left diagram of FIG. All of (x _i, y _i) for the syntax tree y _i in ∈P, the results of the simplified and y _i ¯,

And

The simplified grammar rule set extraction unit 262 extracts a set R￣ of grammar rules r￣ from the simplified syntax tree y _iに つ い_て for each of the syntax trees y _i of the learning data P, and the grammar rule storage unit 24. Is used to convert the grammar rule r￣ converted using the virtual label X into the grammar rule r before conversion using the virtual label X, using the set R of grammar rules r stored in FIG. A conversion function conv is generated, and a set R￣ of grammar rules r￣ and a conversion function conv are stored in the simple grammar rule storage unit 28.

Specifically, the grammatical rule r ￣ is extracted from all the simplified syntax trees y _i ￣ (1 ≦ I ≦ | P |) included in the learning data P. FIG. 8 shows the grammar rule r￣ extracted from the simplified syntax tree of FIG.

  The set of extracted grammar rules r￣ is R￣. Also, the grammar rule r￣ extracted from the simplified syntax tree y￣ is

As described above, there is a correspondence relationship with the element r of the set R of the grammar rules r extracted by the learning unit 22. The function conv is generated as a function that takes a simplified grammar rule r￣ using a virtual label X as an argument and returns a set of grammar rules r in the corresponding set R of grammar rules r.

For each pair (x _i , y _i ) of the learning data P, the simplified syntax tree weight vector update unit 264 is simplified by the sentence x _{i of the} pair (x _i , y _i ) and the syntax tree simplification unit 260. Based on a feature vector f￣ (x, y _i ￣) extracted from the generalized syntax tree y _{i 、} and a weight vector w￣ corresponding to the feature vector f￣ (x, y￣). The score is calculated, and the weight vector w ベ ク ト ル is updated according to the score.

  In particular,

The branch hunting model is learned using the set R￣ of grammatical rules r￣. First, all values of the weight vector w are set to 0, and the super parameter γ is set to an arbitrary value from 0 to 1.

Then, for each pair (x _i , y _i ￣) εP, a score tx _i (γ) is calculated according to the following equation (4).

(4)
However,

Here, f￣ (x _i , r￣) is a function that returns a vector of feature values that can be extracted from the sentence x _i and the simplified rule r￣.

As a result of calculating the above equation (4), as a result of the score tx _i (γ), the weight vector w￣ and the feature vector f 量 (x _i , y￣), the score w￣ · f￣ (x _i , y￣)

At this time, the weight vector w￣ is updated according to the following equation (5) using the sentence x _i , the syntax tree y _i ￣, and the syntax tree y _i ￣ ^* obtained by the above equation (4).

(5)

  here,

And a syntax tree containing the rule r￣

Searching for the one with the highest score.

  The above procedure is repeated a finite number of times, and the simplified syntax tree weight vector updating unit 264 stores the learned weight vector w￣ in the simplified grammar rule storage unit 28. Further, the super parameter γ set with an arbitrary value is stored in the simple grammar rule storage unit 28.

  The rule filter unit 30 receives the input sentence x received by the input unit 10, the set R 集合 of grammar rules r￣ stored in the simple grammar rule storage unit 28, the conversion function conv, the weight vector w￣, and the super parameter γ. As an input, a set R ′ of filtered grammar rules r is obtained for the input sentence x.

  As shown in FIG. 9, the rule filter unit 30 includes a score calculation unit 300 and a grammar rule addition unit 302.

Based on the input sentence x, the set R と of grammatical rules r￣, the weight vector w￣, and the hyperparameter γ, the score calculation unit 300 performs the reference score t for the input sentence x according to the following equation (6). Calculate _x (γ).

(6)

The grammar rule adding unit 302 represents, for each grammar rule r￣ included in the set R￣ of grammar rules r￣, the grammar rule r representing the weight vector w￣, the input sentence x, and the parsing result of the input sentence x. According to the Viterbi score calculated based on the syntax tree y 木 including ￣ and the reference score t _x (γ), the grammar rule r￣ is converted into a grammar rule r before conversion using the virtual label X. And is added to a set R ′ of grammar rules r not using the virtual label X.

  Specifically, first, the set R ′ is initialized with an empty set.

  And for each grammar rule r 規則 ∈R￣,

When

And Here, r￣ is an indexed rule of the input sentence x, and this index is inherited by the grammar rule included in the set that is the return value of the conversion function conv.

  Then, the grammar rule adding unit 302 outputs the set R ′ of grammar rules r to the analyzing unit 32.

  The analysis unit 32 receives the input sentence x, the set R ′ of the grammar rules r, and the weight vector w stored in the grammar rule storage unit 24, and includes them in the set R ′ of the grammar rules r according to the above equation (1). Is calculated on the basis of a feature vector f (x, y) extracted from the syntax tree y representing the result of parsing the input sentence x, the input sentence x, and the weight vector w. A syntax tree ｙ representing the syntax analysis result of the input sentence x having the maximum score is searched and output by the output unit 90.

<Operation of the syntax analysis apparatus according to the embodiment of the present invention>
Next, the operation of the syntax analysis apparatus 100 according to the embodiment of the present invention will be described. When the syntax analysis apparatus 100 receives learning data P that is a set of sentence and syntax tree pairs (x _i , y _i ) through the input unit 10, the syntax analysis apparatus 100 performs rule filter model learning processing illustrated in FIG. 10. The routine is executed.

  First, in step S100, the learning data P is read. In step S 102, a set R of grammar rules r is extracted from each of the syntax trees y included in the learning data P and stored in the grammar rule storage unit 24.

In step S104, the weight vector w is initialized. In step S106, each pair (x _i , y _i ) included in the learning data P is determined based on the set R of grammar rules r extracted in step S102 and the weight vector w initialized or updated. in contrast, in accordance with the above equation (2), carried out the syntax analysis, determine the syntax tree ^ y _i of the sentence x _i, ^ y _i ≠ when y _i, sentence x _i, syntax tree y _i, and the (2 The weight vector w is updated in accordance with the above equation (3) using the syntax tree ^ y _i obtained by the equation).

  In step S108, it is determined whether or not the process of step S106 has been repeatedly performed a predetermined number of times. If the process of step S106 has not been repeated a predetermined number of times, the process returns to step S106. On the other hand, when the process of step S106 is repeatedly executed a predetermined number of times, the process moves to step S110.

In step S110, based on the learning data P, for each of the syntax trees y _i of the learning data P, each of a label different from a label representing a word and a label representing a part of speech is converted into a virtual label X, Generate a simplified syntax tree y _i ￣.

In the next step S112, a set R￣ of grammar rules r￣ is extracted from the syntax tree y _i簡 単 simplified for each of the syntax trees y _{i in} step S110, and the grammar stored in the grammar rule storage unit 24 is extracted. Using the rule set R, a conversion function conv for converting the grammar rule r￣ converted using the virtual label X into the grammar rule r before conversion using the virtual label X is generated. The set R 集合 of grammar rules r￣ and the conversion function conv are stored in the simple grammar rule storage unit 28.

In step S114, the weight vector w￣ is initialized. In the next step S116, each pair (x _i , y _i ) of the learning data P is extracted from the sentence x _{i of the} pair (x _i , y _i ) and the simplified syntax tree y _i ￣. Based on the feature vector fγ (x, y _i ￣) and the initialized or updated weight vector w￣, the score tx _i (γ) is calculated according to the above equation (4), and the score tx _{If i} (γ) is larger than the score w￣ · f￣ (x _i , y￣) calculated from the weight vector w￣ and the feature vector f￣ (x _i , y￣), the sentence x _i , syntax Using the tree y _i ￣ and the syntax tree y _i ￣ ^* obtained in the above equation (4), the weight vector w 更新 is updated according to the above equation (5).

  In step S118, it is determined whether the process of step S116 has been repeatedly performed a predetermined number of times. If the process of step S116 has not been repeated a predetermined number of times, the process returns to step S116. On the other hand, when the process of step S116 is repeatedly performed a predetermined number of times, the rule filter model learning process routine is terminated, and the learned weight vector w￣ is stored in the simple grammar rule storage unit 28.

  When the syntax analysis apparatus 100 receives the input sentence x by the input unit 10, the syntax analysis apparatus 100 executes a syntax analysis processing routine shown in FIG.

In step S120, the input sentence x is read. In step S122, the input sentence x, the set R￣ of grammar rules r￣ stored in the simple grammar rule storage unit 28, the weight vector w￣, the super parameter γ, Based on the above, the reference score t _x (γ) for the input sentence x is calculated according to the above equation (6).

In step S124, each grammar rule r 規則 included in the set R￣ of grammar rules r￣ is calculated based on the weight vector w￣, the input sentence x, and the syntax tree y￣ including the grammar rule r￣. When the Viterbi score to be obtained is equal to or higher than the reference score t _x (γ), the conversion function conv converts the grammar rule r 集合 into a set of grammar rules r before conversion using the virtual label X. , The virtual label X is not added to the set R ′ of grammar rules r.

  In the next step S126, the input sentence x, the set R ′ of the grammar rules r, and the weight vector w stored in the grammar rule storage unit 24 are input to the set R ′ of grammar rules r according to the above equation (1). The score calculated based on the vector f (x, y) of the feature quantity extracted from the syntax tree y representing the parsing result of the input sentence x using the included grammar rule r and the weight vector w is maximum. The syntax tree y ^ of the input sentence x is searched. In step S128, the syntax tree ＾ obtained in step S126 is output as a syntax analysis result by the output unit 90, and the syntax analysis processing routine is terminated.

  As described above, according to the syntax analysis apparatus according to the embodiment of the present invention, a simplified syntax tree y￣ is generated for each syntax tree y of the learning data P, and the grammar rule r￣ is generated. A set R￣ is extracted, and a score for each pair (x, y) of the learning data P is calculated based on a feature vector f￣ (x, y￣) and a weight vector w￣ for the pair. By updating the weight vector w￣ according to the score, it is possible to learn a grammar rule filter model that can reduce grammar rules in order to perform sentence syntax analysis at high speed.

  Further, based on the set R 計算 of grammar rules r￣ extracted from the simplified syntax tree y￣, the learned weight vector w￣, and the input sentence x, a reference score is calculated, and the grammar rule r For each ￣, the grammar rule r￣ is converted according to the score calculated based on the parse tree y￣ including the grammar rule r￣ that represents the parsing result of the input sentence x and the reference score. Grammatical rules r are converted into grammar rules r and added to the set R ′ of grammatical rules r, so that the grammar rules can be reduced in order to parse the sentence at high speed, and the input sentence can be parsed at high speed. Can do.

  In addition, since the rule filter can reduce the number of grammatical rules having detailed phrase labels used for syntax analysis, the analysis can be performed faster than the conventional statistical syntax analysis apparatus. By improving the parsing speed, it becomes possible to speed up application systems such as machine translation and document summarization.

  Note that the present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  For example, in the above embodiment, the case where the grammar rule filter model learning device and the syntax analysis device are realized by one syntax analysis device has been described as an example. However, the present invention is not limited to this. The model learning device and the syntax analysis device may be provided separately. In this case, the grammatical rule filter model learning device may include the learning unit 22 and the rule filter model learning unit 26. Further, the syntax analysis apparatus may include the rule filter unit 30 and the analysis unit 32.

  Further, in the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium or provided via a network. It is also possible to do.

DESCRIPTION OF SYMBOLS 10 Input part 20 Operation part 22 Learning part 24 Grammar rule memory | storage part 26 Rule filter model learning part 28 Simple grammar rule memory | storage part 30 Rule filter part 32 Analysis part 90 Output part 100 Syntax analyzer 220 Grammar rule set extraction part 222 Weight vector update Unit 260 Syntax tree simplification unit 262 Simplified grammar rule set extraction unit 264 Weight vector update unit 300 Score calculation unit 302 Grammar rule addition unit

Claims (8)

A grammar rule set extraction unit that extracts a set R of grammar rules r from each of the syntax trees y of learning data, which is a set of pairs of a sentence x and a syntax tree y representing a parsing result of the sentence x;
On the basis of the learning data, for each of the parse tree y of the training data, each of the different labels is a label representing the label and part of speech represent words, by converting the virtual label, simplified syntax tree a syntax tree simplification unit for generating y￣,
A set R￣ of grammatical rules r￣ is extracted from a simplified syntax tree y に つ い て for each of the learning data syntax trees y, the grammar rules r before conversion using the virtual labels, and the grammar A correspondence relationship with the grammar rule r￣ converted using the virtual label included in the set R of the grammar rules r extracted by the rule set extraction unit is obtained, and the virtual relationship is determined based on the correspondence relationship. A simplified grammar rule set extraction unit that generates a conversion function for converting a grammar rule r￣ converted using a typical label into a grammar rule r before conversion using the virtual label;
For each of the pairs of learning data, a feature quantity vector f￣ (x, y￣) extracted from the pair sentence x and the simplified syntax tree y￣, and the feature quantity vector f A weight vector update unit for a simplified syntax tree that calculates a score based on a weight vector w に corresponding to ￣ (x, y￣), and updates the weight vector w￣ according to the score;
Grammar rule filter model learning device including For each of the pair of front Symbol learning data, sentence x and the syntax tree y extracted from to be the feature quantity of the vector f (x, y) of the pair and the feature amount vector f (x, y) to The grammatical rule filter model learning device according to claim 1, further comprising a learning unit that calculates a score based on a corresponding weight vector w and updates the weight vector w according to the score.   The simplified syntax tree weight vector updating unit calculates the score for each of the pairs of the learning data, and updates the weight vector w￣ according to the score by a predetermined number of repetitions. The grammatical rule filter model learning device according to claim 1, wherein the grammar rule filter model learning device is repeated. Each of the label representing the word and the label different from the label representing the part of speech included in the syntax tree y representing the sentence parsing result is extracted from the simplified syntax tree y￣ by converting each label into a virtual label. A set R￣ of grammatical rules r￣, a pre-learned weight vector w￣ corresponding to a feature vector f￣ (x, y￣) extracted from the sentence x and the simplified syntax tree y￣ A score calculation unit for calculating a reference score based on the input sentence x,
For each of the grammar rules r￣ included in the set R￣ of the grammar rules r￣, the grammar rule r￣ representing the weight vector w￣, the input sentence x, and the parsing result of the input sentence x is And a grammar rule r before conversion using the virtual label according to the score calculated based on the simplified syntax tree y￣ and the reference score, and the virtual Based on the correspondence with the grammatical rule r￣ converted using the label, the grammatical rule r￣ converted using the virtual label, which has been generated in advance, is converted before the conversion using the virtual label. The grammar rule r￣ is converted to the grammar rule r before conversion using the virtual label by using a conversion function for converting to the grammar rule r, and the virtual label is not used. Additional grammar rules added to the set R ′ of grammar rules r Grammar rules filter device including a part. A grammatical rule filter device according to claim 4 ,
Corresponds to the set R ′ of the grammar rules r obtained by the grammar rule filter device and the feature quantity vector f (x, y) extracted from the sentence x and the syntax tree y representing the result of parsing the sentence x. Based on a weight vector w learned in advance and the input sentence x, a syntax tree y representing a parsing result of the input sentence x using a grammar rule r included in the set R ′ of the grammar rules r, and A syntax tree y representing the result of parsing of the input sentence x having the maximum score calculated based on the feature vector f (x, y) extracted from the input sentence x and the weight vector w. A parser that includes an analysis unit that searches for ^. Computer program to function as each section of the grammar rule filter model learning equipment according to any one of claims 1-3. A program for a computer to function as each unit of the grammatical rules filter equipment according to claim 4, wherein. A program for a computer to function as each unit of the syntax analysis apparatus 請 Motomeko 5 wherein.