JP6291440B2 - Parameter learning method, apparatus, and program - Google Patents

Description

  The present invention relates to a parameter learning method, apparatus, and program.

Rhetorical Structure Theory (RST) is a theory for capturing the logical structure (rhetorical structure) of a document (for example, Non-Patent Document 1). A tree representing a rhetorical structure based on RST is referred to as a rhetorical structure theory based discourse tree (RST-DT). An example of RST-DT is shown in FIG. Also, giving a rhetorical structure tree to a raw text to which no annotation is given is called discourse structure analysis. The given document is divided into the minimum units (Elementary Discourse Unit: EDU) in the discourse structure (in FIG. 9, the range from e1 to e10 corresponds to it). Each EDU is assigned a label of satellite (Stellite: S) or nucleus (Nuclues: N), and S always has a relationship of modifying N. Further, labels representing rhetorical relationships are assigned between S and N and between N and N. For example, a relation label “Background” is assigned between e ₁ and e ₂ .

  In the RST-DT, until the entire document becomes one node, the N or S label given between the nodes and the rhetorical label are used as one node, and the label assignment and the node generation are performed recursively. Root is a virtual node that represents the entire document.

  HILDA (for example, Non-Patent Document 2) is a representative algorithm for analyzing a RST-DT when a document is given. HILDA is one of the most prioritized search methods using the greedy method, and analyzes a given sentence into RST-DT using the following procedure.

(Step 1) A given document is divided into EDUs.

(Step 2) Support Vector Machine is used to determine which of the adjacent nodes is most likely to be combined, and the adjacent nodes are combined into one node after giving a label.

(Step 3) If the entire node is a single node, the combined tree is returned; otherwise, the process returns to Step 2.

William C, Mann and Sandra A. Thompson, “Rhetorical structure theory: Toward a functional theory of text organization”, 1988, Text, 8 (3), p.243-281 H. Hernault, H. Prendinger, David A. duVerle, and M. Ishizuka, “HILDA: A Discourse Parser Using Support Vector Machine Classification”, 2010, In Dialogue & Discourse, 2010 (3), p.1-33

  However, conventional discourse structure analysis techniques represented by HILDA are not very robust against search errors. For example, if the relationship between e5 and e6 in FIG. 9 is given as “Elaboration”, the rhetorical structure label “Contrast” may not be given correctly when it is further combined with e4.

  The present invention has been made in view of the above circumstances, and an object thereof is to provide a parameter learning method, apparatus, and program capable of obtaining parameters for accurately performing discourse structure analysis.

  In order to achieve the above object, a parameter learning method of the present invention is a parameter learning method in a parameter learning device including a learning input unit, a parameter learning unit, and an iterative determination unit, wherein the learning input unit includes a plurality of learning input units. For each of the learning documents, each of the character string units in the learning document, a root node representing the whole of the learning document, and a sequence of at least one character string unit of the learning documents A rhetorical structure tree based on the rhetorical structure of each sequence of character strings in the learning document, representing a hierarchical structure with each node as a node, and representing a modification relationship and a relationship label between the sequences of character strings. And when each of the correct rhetorical structure trees corresponding to the learning document is received, the parameter learning unit performs the correct answer for each of a plurality of learning documents. A column of subtrees included in the rhetorical structure tree and a subtree of the rhetorical structure tree corresponding to the learning document selected using a weight vector for a feature vector extracted from the column of the subtree Corresponding to the column of subtrees included in the correct rhetorical structure tree, the score difference calculated using the feature vector and the weight vector being maximized, and the learning document Updating the weight vector based on a feature vector extracted from each of the subtree columns included in a pair with the subtree column of the rhetorical structure tree, and the iteration determination unit is predetermined. Repeating the updating by the parameter learning unit a predetermined number of times.

  The parameter learning device according to the present invention includes, for each of a plurality of learning documents, each of the character string units in the learning document, a root node representing the whole learning document, and among the learning documents, A sequence of at least one character string unit representing a hierarchical structure with each node as a node, and a modification relationship between the character string unit sequences and a relationship label; A rhetorical structure tree based on each rhetorical structure and receiving each correct rhetorical structure tree corresponding to the learning document; and for each of a plurality of learning documents, the correct rhetorical structure A sequence of subtrees included in the tree, and a sequence of subtrees of the rhetorical structure tree corresponding to the learning document, selected using a weight vector for a feature vector extracted from the sequence of subtrees. Among the subtrees included in the correct rhetorical structure tree, the difference between the scores calculated using the feature vector and the weight vector is maximized, and the document corresponding to the learning document A parameter learning unit that updates the weight vector based on a feature vector extracted from each of the subtree columns included in a pair with a subtree column of the rhetorical structure tree, and a predetermined number of times, An iterative determination unit that repeats updating by the parameter learning unit.

  Further, the parameter learning unit of the present invention is generated by combining a subtree sequence included in the correct rhetorical tree with a pair of adjacent subtrees in the previously selected subtree sequence. , Of the set of subtrees included in the correct rhetorical structure tree, the subtree having the maximum score calculated using the feature vector and the weight vector extracted from the subtree sequence Generated by combining adjacent subtree pairs in each of the subtree columns selected previously with respect to the subtree column of the rhetorical structure tree corresponding to the learning document. Selecting a row of subtrees having the top k scores calculated using the feature vector and the weight vector extracted from the subtree column, Subtrees included in the correct rhetorical structure tree Generates a pair of a subtree column selected with respect to each of the top k subtree columns selected for the subtree column of the rhetorical structure tree corresponding to the learning document. A sequence of subtrees included in the correct rhetorical structure tree in which the difference between the scores calculated using the feature vector and the weight vector among the generated pairs is maximized; The weight vector is updated based on a feature vector extracted from each of the subtree columns included in a pair with the subtree column of the rhetorical structure tree corresponding to the learning document. Can do.

  The program of this invention is a program for functioning a computer as each part of the discourse structure analysis apparatus of this invention.

  As described above, according to the parameter learning method, apparatus, and program of the present invention, each of a plurality of learning documents is extracted from the subtree sequence included in the correct rhetorical structure tree and the subtree sequence. Of the score calculated using the feature vector and the weight vector out of the pair of the subtree of the rhetorical structure tree corresponding to the learning document, which is selected using the weight vector for the feature vector Extracted from each of the subtree columns included in the pair of the subtree included in the correct rhetorical tree with the maximum difference and the subtree column corresponding to the learning document. By updating the weight vector based on the feature vector, it is possible to obtain an effect that a parameter for performing discourse structure analysis with high accuracy can be obtained.

It is a block diagram which shows one structural example of the discourse structure analysis apparatus of embodiment of this invention. It is a figure which shows an example of the algorithm of discourse structure analysis. It is a block diagram which shows one structural example of the parameter learning apparatus of embodiment of this invention. It is a figure which shows an example of the algorithm which produces | generates the pair from which a score becomes the maximum. It is a figure which shows an example of the algorithm which learns the weight vector w. It is a flowchart which shows the content of the learning process routine in the parameter learning apparatus of embodiment of this invention. It is a flowchart which shows the content of the largest pair calculation process routine in the parameter learning apparatus of embodiment of this invention. It is a flowchart which shows the content of the analysis process routine in the discourse structure analysis apparatus of embodiment of this invention. It is explanatory drawing for demonstrating the rhetorical structure tree (RST-DT) based on a rhetorical structure.

<Overview>
First, an outline of an embodiment of the present invention will be described. Embodiments of the present invention relate to the analysis of discourse structures between grammatical elements in a given document. This technique is a technique for analyzing a discourse structure between grammatical elements of an entire document without an annotation as a tree. The point of the embodiment of the present invention is that the most prioritized search in discourse structure analysis is extended to beam search and robust to search errors, and the easiest priority search is extended to beam search to learn parameters. It is.

  The embodiment according to the present invention performs a discourse structure analysis using beam search in order to capture a discourse structure more accurately. HILDA, which is an existing method, is a method based on the greedy method, and if a decision at a certain point in time (for example, combining e5 and e6 in FIG. 9 above and giving the label “Evidence”, etc.) It may adversely affect subsequent decisions. Therefore, in the embodiment of the present invention, search errors are reduced by using beam search for HILDA which is a greedy method. The embodiment of the present invention is divided into two steps: a step of learning an optimum parameter and a step of analyzing a discourse structure of a document input using the optimum parameter.

  First, the step of learning the optimum parameter will be described. At this stage, the feature vector and the initial parameter extracted from the discourse structure annotated document are input to the parameter learning unit. The parameter learning unit learns the optimum parameters for the discourse structure analysis and outputs the parameters as learned parameters.

  Next, the stage of analyzing the discourse structure of the input document using optimum parameters will be described. At this stage, the input document is divided into EDU units. Next, a feature vector is extracted from the EDU column, and is passed to the discourse structure analysis unit together with the learned parameters obtained in the previous stage. The discourse structure analysis unit outputs RST-DT as a result of the discourse structure analysis for the documents input based on them.

Here, RST-DT represents a hierarchical structure in which the root node represents the entire document, and each of at least one character string unit sequence of the document is a node, and the character string unit sequence. It is a rhetorical structure tree based on the rhetorical structure of each series of character string units of a document, which represents a modification relationship and a relation label between them.
The character string unit of RST-DT corresponds to the smallest unit (Elementary Discourse Unit: EDU) in the document.

<System configuration of discourse structure analyzer>
Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram showing a discourse structure analyzing apparatus 100 according to an embodiment of the present invention. The discourse structure analyzing apparatus 100 is composed of a computer including a CPU, a RAM, and a ROM storing a program for executing a discourse structure analyzing process routine, and is functionally configured as follows. .

  As shown in FIG. 1, the discourse structure analyzing apparatus 100 according to the present embodiment includes an input unit 10, a parameter database 20, a calculation unit 30, and an output unit 40.

  In the discourse structure analyzing apparatus 100, when a document to be analyzed is input, the discourse structure analysis of the document is performed.

  The input unit 10 receives input of a document to be analyzed. Note that the input document is a document including at least one sentence.

  The parameter database 20 stores a weight vector w learned by a parameter learning device 200 described later.

  The calculation unit 30 performs a discourse structure analysis on the analysis target document received by the input unit 10. In addition, the calculation unit 30 includes an EDU division unit 32, a feature extraction unit 34, and a discourse structure analysis unit 36.

  The EDU division unit 32 divides the analysis target document received by the input unit 10 into EDUs. For example, the EDU division unit 32 uses a classifier such as SVM between words in the document to determine whether the word is divided or not, and outputs a document divided into EDUs.

The feature extraction unit 34 extracts a feature vector from a subtree sequence that is an EDU sequence obtained by the EDU division unit 32 or a subtree sequence generated by the discourse structure analysis unit 36 described later. For example, the feature extraction unit 34 sets f (S) as a feature quantity vector when combining two RST-DT subtrees S _i and S _{i + 1} from a column S of the RST-DT subtree. Among the feature vectors, typical ones are listed below.

(1) Whether the number of words included in the subtree S _i of RST-DT is 5 or less.
(2) Whether subtrees S _i and S _{i + 1 of} RST-DT are included in the same sentence.
(3) Whether the head of the subtree S _i of RST-DT starts with “Because”.
(4) The number of EDUs included in the subtrees S _i and S _{i + 1 of the} RST-DT.
(5) whether the part of speech of the head word of the RST-DT subtree _{S i} of a verb.
(6) Is the rhetorical relation label of the top node of the subtree S _{i + 1} of the RST-DT “Evidence”?

  For example, in FIG. 9 above, the column [e1, e2, e3, e4, e5-6, e7, e8, e9, e10] of the RST-DT sub-tree is S, and e4 and e5-6 are combined. The feature amount will be described. Here, e5-6 represents a node obtained by combining e5 and e6 with the rhetorical structure label “Evidence”. The actual text of e4 is "Only the midday sun at tropical latitudes is warm enough to thaw ice on occasion," and the actual text of e5-6 is "but any liquid water formed that way would evaporate almost instantly because of the low atmospheric pressure. " At this time, the characteristic amount (relation label is “Contrast”) when combining e4 and e5-6 is

(1) Since the number of words included in e4 is 16, the number of words is larger than 5.
(2) e4 and e5-6 are included in the same sentence.
(3) The beginning of e4 does not start with “Because”.
(4) The number of EDUs included in e4 and e5-6 is three.
(5) The part of speech of e4 is the verb.
(6) The rhetoric label of the top node of e5-6 is “Evidence”.
Therefore, f (S) = [0, 1, 0, 3, 1, 1].

  The discourse structure analysis unit 36 calculates the weight vector stored in the parameter database 20 for each of a plurality of subtree columns generated based on a subtree column that is an EDU column or a previously generated subtree column. Based on w and the feature quantity vector f (S) extracted by the feature extraction unit 34, a score for a subtree column is calculated. Then, the discourse structure analysis unit 36 stores each column of subtrees having the highest k scores in the array beam based on each score for each column of the plurality of subtrees.

  Specifically, the discourse structure analyzing unit 36 enumerates a column of the RST-DT subtree anew from the column of the RST-DT subtree according to the function expand shown in the following equation (1).

  Here, the set of rhetorical relation labels is L, the sequence including the subtree of RST-DT is S, the set of pairs of nuclei or satellites is NS = {(Nucleus, Satellite), (Satellite, Nucleus), (Nucleus, Nucleus) )}. The function build receives as inputs S and an index iε {1,..., Length (S) −1}, rhetorical relationship label lεL, a nucleus or satellite pair (ns1, ns2) εNS. Also, the function build returns a new RST-DT subtree obtained by combining two RST-DT subtrees Si and Si + 1. At that time, a pair of rhetorical label 1 and a nucleus or satellite is given.

FIG. 2 shows a pseudo code representing the specific processing contents of the discourse structure analysis unit 36. The function top _k (Z) is a function that holds k candidates with the highest score from the set (Z) of columns of the subtree of RST-DT. Each RST-DT subtree column zεZ holds the inner product of the weight vector w and the feature vector extracted from z as a score.

  Algorithm 1 shown in FIG. 2 receives a document x = [e1, e2,..., En] divided into a weight vector w and an EDU. The discourse structure analysis unit 36 expands the subtree column candidates using the function expand, and holds the k candidate subtree column with the highest score in the array beam according to the following equation (2). Note that the array beam is a two-dimensional array, and the beam [i] stores an array in which a column of the i-th subtree having the highest score is stored. That is, the beam [i] [j] stores the j-th subtree in the column with the highest i-th subtree.

  The finally obtained beam [0] [0] stores one tree having the highest score, that is, RST-DT. The discourse structure analyzing unit 36 outputs the finally obtained beam [0] [0] as an RST-DT that is an analysis result.

  The output unit 40 outputs the RST-DT output by the discourse structure analyzing unit 36 as an analysis result.

  If the weight vector w is known, the discourse structure analysis based on the easy search using the beam search can be performed using Algorithm 1 shown in FIG. However, the weight vector w is not known. Therefore, in the present embodiment, the weight vector w is obtained based on the structured perceptron. A structured perceptron is an example of a learning algorithm.

<System configuration of parameter learning device>
FIG. 3 is a block diagram illustrating the parameter learning device 200 according to the embodiment of this invention. The parameter learning device 200 is configured by a computer including a CPU, a RAM, and a ROM that stores a program for executing a parameter learning processing routine described later, and is functionally configured as follows. Yes.

  The learning input unit 50 receives input of a plurality of learning data. Specifically, the learning input unit 50 learns, for each of a plurality of learning documents, a combination of each EDU in the learning document and a correct RST-DT corresponding to the learning document. Accept as data.

  The learning calculation unit 60 learns a weight vector w for performing discourse structure analysis based on the plurality of learning data received by the learning input unit 50. The learning calculation unit 60 includes a learning database 62, a learning feature extraction unit 64, a parameter learning unit 66, and an iterative determination unit 68.

  The learning database 62 stores a plurality of learning data received by the learning input unit 50.

  The learning feature extraction unit 64 is based on a subtree sequence that is a sequence of EDUs included in each of the plurality of learning data stored in the learning database 62 or a subtree sequence generated by the parameter learning unit 66 described later. Similarly to the feature extraction unit 34, a feature quantity vector is extracted.

  The parameter learning unit 66 selects the learning data by using the column of the subtree included in the correct RST-DT and the weight vector w for the feature vector extracted from the column of the subtree. A pair with a column of the RST-DT sub-tree corresponding to is generated. Then, the parameter learning unit 66 includes a subtree included in the correct RST-DT in which the difference between the scores calculated using the feature vector and the weight vector w is maximized among the generated pairs. The weight vector w is updated based on the feature vector extracted from each of the subtree columns included in a pair of the column and the subtree of the RST-DT corresponding to the learning document. The parameter learning unit 66 repeats the above processing for each of a plurality of learning data.

  Here, details of the processing of the parameter learning unit 66 will be described. First, in accordance with Algorithm 2 shown in FIG. 4, the parameter learning unit 66 includes a sequence of subtrees included in the correct RST-DT in which the difference between the scores calculated using the feature vector and the weight vector w is maximized. Then, a pair with a column of the RST-DT subtree corresponding to the learning document is selected.

  Specifically, the parameter learning unit 66 is adjacent to the subtree sequence selected by the parameter learning unit 66 last time with respect to the subtree sequence included in the correct RST-DT according to the following equation (3). Of the set of subtree columns included in the correct RST-DT generated by combining only one matching subtree pair, the feature amount extracted from the subtree column by the learning feature extraction unit 64 A column of the subtree having the maximum score calculated using the vector f (o) and the weight vector w is selected. Here, the column of the subtree included in the correct RST-DT is referred to as oracle o.

  Next, in accordance with the above equation (2), the parameter learning unit 66 sets the adjacent subtree in each of the subtrees selected last time with respect to the subtree of the RST-DT corresponding to the learning document. Among the set of sub-tree columns generated by combining the pairs, the top k scores are calculated using the feature vector f (S) and the weight vector w extracted from the sub-tree column. Select a row of subtrees.

  Next, the parameter learning unit 66, in accordance with the following equation (4), selects the subtree sequence selected for the subtree sequence included in the correct RST-DT and the RST-DT corresponding to the learning document. A pair with each of the top k subtree columns selected for the subtree column is generated.

  The parameter learning unit 66 repeats the above processing until the subtree sequence included in the correct RST-DT becomes one tree.

  The parameter learning unit 66 then corrects the correct RST-DT in which the difference between the scores calculated using the feature vector and the weight vector w among the generated pairs is maximized according to the following equation (5). A pair of a subtree sequence included in the RST-DT subtree sequence corresponding to the learning document is selected.

  Then, the parameter learning unit 66 is included in a pair of a subtree column included in the correct RST-DT selected according to the above formula (5) and a subtree column of the RST-DT corresponding to the learning document. The weight vector w is updated based on the feature vector extracted from each of the columns of the partial tree. The update processing in the parameter learning unit 66 is shown in Algorithm 3 in FIG.

In the structured perceptron used in the present embodiment, every time one correct RST-DT (t) is given, the weight is updated based on the combination of Oracle o and the predicted RST-DT (S). . The set of Oracle o and the predicted RST-DT (S) is obtained by the function max-violation-pair described in Algorithm 2 in FIG. The function max-violation-pair stores the pair of predicted RST-DT obtained by Oracle and beam search in pairs, and the difference between the scores of Oracle o and predicted RST-DT (S) obtained by beam search is This function returns the largest tuple. The function expand-oracle used in the function max-violation-pair is a function that expands candidates as in the function expand but returns only candidates included in the correct RST-DT (t). In the structured perceptron, the weight vector w is updated using a set obtained from the function max-violation-pair. Further, the structured perceptron, the subtree of the resulting RST-DT pair (o, S) is extracted from each feature vector f (o), the weight vector by adding the difference f (S) ∈R ^M Update.

  The iteration determination unit 68 repeats the update by the parameter learning unit 66 a predetermined number of times. The parameter learning unit 66 stores the updated weight vector w in the parameter database 70 when it has been updated a predetermined number of times.

  The parameter database 70 stores the weight vector w updated by the parameter learning unit 66.

<Operation of parameter learning device>
Next, the operation of the parameter learning device 200 according to the present embodiment will be described. First, when a plurality of learning data is input to the parameter learning device 200, the input plurality of learning data is stored in the learning database 62 by the parameter learning device 200. Then, the parameter learning device 200 executes a learning process routine shown in FIG.

  First, in step S100, one learning data is read from a plurality of learning data stored in the learning database 62 and set.

  Next, in step S102, the parameter learning unit 66, with respect to the learning data set in step S100, the subtree sequence included in the correct RST-DT and the RST-DT subtree corresponding to the learning document. Create a pair with the column. Note that the RST-DT subtree column corresponding to the learning document is selected using the weight vector w for the feature vector extracted from the subtree column. This step S102 is realized by the maximum pair calculation processing routine shown in FIG.

<Maximum pair calculation processing routine>
First, in step S200, the array pairs is initialized, and an array storing each of the EDUs included in the learning data set in step S100 is stored in the oracle and the array beam.

  Next, in step S202, it is determined whether or not there is one array element stored in the oracle set in step S200 or the oracle updated in the previous step S206. If the number of elements stored in oracle is not one, the process proceeds to step S204. On the other hand, if one tree having the highest score is generated by the processing of steps S204 to S206 described later and the number of elements of the array stored in oracle is one, the process proceeds to step S214.

  In step S204, the correct answer is obtained from the sequence of the subtree stored in the oracle updated in the previous step S206 with respect to the sequence of the subtree included in the correct RST-DT of the learning data by the function expand-oracle. Each new subtree sequence included in each RST-DT is generated. Then, the learning feature extraction unit 64 extracts a feature vector f (o) from each newly generated subtree sequence.

  In step S206, calculation is performed using the feature vector f (o) extracted in step S204 and the weight vector w from the sequence of the new subtree generated in step S204 according to the above equation (3). The column of the subtree that gives the maximum score is selected and stored in oracle.

  In step S208, by using the function expand, from each of the subtree columns stored in the beam updated in the previous step S210, with respect to the subtree column of the RST-DT corresponding to the learning document of the learning data, Generate a set of subtree columns. Then, the learning feature extraction unit 64 extracts a feature quantity vector f (S) for each of the subtree columns S included in the newly generated subtree column set.

  In step S210, the feature vector f (S) extracted in step S208 and the weight vector w are calculated from the set of columns of the subtree generated in step S208 according to the equation (2). The column of the subtree having the top k scores is selected and stored in the array beam.

  In step S212, according to the above equation (4), a pair of the subtree stored in the oracle updated in step S206 and the subtree stored in the array beam updated in step S210 is paired. Each is generated, stored in the array pairs, and the process returns to step S202.

  In step S214, the correct RST in which the difference between the scores calculated using the feature vector and the weight vector w among the pairs stored in the array pairs in step S212 is maximized according to the above equation (5). The pair of the subtree included in the DT and the subtree column of the RST-DT corresponding to the learning document is output, and the maximum pair calculation processing routine is terminated.

  Next, returning to the learning processing routine, in step S104, the parameter learning unit 66 is based on the feature vector extracted from each column of the subtree included in the pair based on the pair output in step S102. , Update the weight vector w.

  In step S106, it is determined whether or not the processing in steps S100 to S104 has been executed for all of the plurality of learning data stored in the learning database 62. When the processes of steps S100 to S104 are executed for all of the plurality of learning data stored in the learning database 62, the process proceeds to step S108. On the other hand, if there is learning data that has not been subjected to the processing of steps S100 to S104, the process returns to step S100.

  In step S108, it is determined whether or not the processing in steps S100 to S106 has been repeated a predetermined number of times. When the processes in steps S100 to S106 are repeated a predetermined number of times, the process proceeds to step S110. On the other hand, if the processes in steps S100 to S106 are not repeated a predetermined number of times, the process returns to step S100.

  In step S110, the weight vector w obtained in the process of step S104 is stored in the parameter database 70, and the learning process routine is terminated.

<Operation of discourse structure analyzer>
Next, the operation of the discourse structure analyzing apparatus 100 of the present embodiment will be described. First, when the weight vector w stored in the parameter database 70 of the parameter learning device 200 is input to the discourse structure analyzing device 100, it is stored in the parameter database 20. When the input document to be analyzed is input to the discourse structure analyzing apparatus 100, the discourse structure analyzing apparatus 100 executes an analysis processing routine shown in FIG.

  First, in step S300, the input unit 10 receives an input document to be analyzed.

  In step S302, the EDU divider 32 divides the input document to be analyzed received in step S300 into EDUs.

  In step S304, the discourse structure analysis unit 36 reads the weight vector w stored in the parameter database 20.

  In step S306, the discourse structure analyzing unit 36 stores an array storing each of the EDUs obtained in step S302 in the array beam.

  In step S308, the discourse structure analysis unit 36 has one array element stored in the array beam [0] set in step S306 or the array beam [0] updated in the previous step S312. It is determined whether or not. If the number of elements in the array stored in the array beam [0] is not one, the process proceeds to step S310. On the other hand, if the number of elements of the array stored in the array beam [0] is one by the processes of steps S310 to S312 described later, the process proceeds to step S314.

  In step S310, the discourse structure analysis unit 36 uses the function expand to calculate the subtree from each of the subtree columns stored in the array beam set in step S306 or the array beam updated in the previous step S312. Generate a set of columns. Then, the feature extraction unit 34 extracts a feature quantity vector f (S) for each of the subtree columns S included in the generated set of subtree columns.

  In step S312, the discourse structure analysis unit 36 has the top k scores calculated using the feature vector f (S) and the weight vector w extracted in step S310 according to the above equation (2). A row of subtrees is selected, stored in the array beam, and the process returns to step 308 above.

  In step S314, out of the array beam updated in step S312, RST-DT stored in beam [0] [0] is output as an analysis result by the output unit 40, and the analysis processing routine ends.

  As described above, according to the parameter learning apparatus of the present embodiment, for each of a plurality of learning documents, the subtree sequence included in the correct RST-DT and the features extracted from the subtree sequence The difference between the scores calculated using the feature quantity vector and the weight vector w among the pairs of the RST-DT subtree columns corresponding to the learning document, which are selected using the weight vector w for the quantity vector. Extracted from each of the subtree columns included in a pair of the subtree column included in the correct RST-DT and the RST-DT subtree column corresponding to the learning document. By updating the weight vector w based on the vector, the weight vector w for accurately performing the discourse structure analysis can be obtained.

  According to the discourse structure analysis apparatus of the present embodiment, the discourse structure analysis can be performed with high accuracy by performing the discourse structure analysis using the weight vector w obtained by the parameter learning apparatus.

  Also, by using the parameter learning device and the discourse structure analysis device of the present embodiment, it is possible to perform robust analysis against search errors by performing the most prioritized search based on the beam search, and more accurate discourse structure analysis. Is possible.

  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, although a case has been described with the present embodiment where an EDU in a document is each RST-DT node, each node can also be represented as a character string other than an EDU. In that case, the EDU dividing unit 32 divides the document into the character string units, and constructs an RST-DT in which the character string units are represented as nodes.

  Further, the parameter learning device and the discourse structure analyzing device of the present embodiment can be applied not only to English but also to other languages such as Japanese.

  The learning database 62 and the parameter database 70 may be provided outside the parameter learning device and connected to the parameter learning device via a network. The parameter database 20 may be provided outside the discourse structure analyzing apparatus and connected to the discourse structure analyzing apparatus via a network.

  Further, the document input to the input unit 10 may be in a form that has already been divided into sentences or EDUs. In that case, the processing of the EDU division unit 32 is omitted.

  Moreover, although the case where the parameter learning device and the discourse structure analyzing device are configured as separate devices has been described as an example in the above embodiment, the parameter learning device and the discourse structure analyzing device may be configured as one device. Good.

  The parameter learning device and the discourse structure analyzing device described above have a computer system inside. However, if the “computer system” uses a WWW system, a homepage providing environment (or display environment) is also available. Shall be included.

  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.

DESCRIPTION OF SYMBOLS 10 Input part 20, 70 Parameter database 30 Operation part 32 EDU division part 34 Feature extraction part 36 Discourse structure analysis part 40 Output part 50 Learning input part 60 Learning operation part 64 Learning feature extraction part 62 Learning database 66 Parameter learning part 68 Iteration Determination Unit 100 Discourse Structure Analysis Device 200 Parameter Learning Device

Claims (4)

A parameter learning method in a parameter learning device including a learning input unit, a parameter learning unit, and an iterative determination unit,
The learning input unit, for each of a plurality of learning documents, each of the character string units in the learning document, the root node represents the whole learning document, and at least of the learning documents Each of the character string unit series of the learning document that represents a hierarchical structure with each of the character string unit series as a node, and represents a modification relationship and a relation label between the character string unit series Each of the rhetorical structure tree based on the rhetorical structure and the correct rhetorical structure tree corresponding to the learning document.
The parameter learning unit
For each of the multiple learning documents
A part of the rhetorical structure tree corresponding to the learning document selected using a subtree included in the correct rhetorical structure tree and a weight vector for a feature vector extracted from the subtree line A pair of subtrees included in the correct rhetorical structure tree, wherein the difference between the scores calculated using the feature vector and the weight vector is maximized, and the learning tree Updating the weight vector based on a feature vector extracted from each of the subtree columns included in a pair with a subtree column of the rhetorical structure tree corresponding to a document;
The iteration determination unit repeating the update by the parameter learning unit a predetermined number of times;
A parameter learning method including: The step of the parameter learning unit updating the weight vector includes:
A part included in the correct rhetorical structure tree generated by combining a pair of adjacent subtrees in the previously selected subtree column with respect to the subtree string included in the correct rhetorical structure tree Selecting a column of a partial tree having a maximum score calculated using the feature vector and the weight vector extracted from the column of the partial tree from a set of trees;
For each column of subtrees of the rhetorical structure tree corresponding to the learning document, a sequence of subtrees generated by combining pairs of adjacent subtrees in each of the subtree columns selected previously. From the set, select a column of the subtree having the top k scores calculated using the feature vector and the weight vector extracted from the column of the subtree;
A column of the subtree selected for the subtree column included in the correct rhetorical structure tree and the top k selected for the subtree column of the rhetorical tree corresponding to the learning document Repeat to generate a pair with each of the subtree columns
Among the generated pairs, a subtree sequence included in the correct rhetorical structure tree having a maximum score difference calculated using the feature vector and the weight vector, and the learning document The parameter learning method according to claim 1, wherein the weight vector is updated based on a feature vector extracted from each column of the subtrees included in a pair with a subtree column of the rhetorical structure tree corresponding to. For each of the plurality of learning documents, each of the character string units in the learning document and a root node represents the whole of the learning document, and the sequence of at least one character string unit of the learning documents A rhetorical structure based on the rhetorical structure of each sequence of character strings in the learning document, representing a hierarchical structure with each of the nodes as a node, and representing a modification relationship and a relationship label between the sequences of the character string An input unit that accepts each correct rhetorical structure tree corresponding to the learning document,
For each of the multiple learning documents
A part of the rhetorical structure tree corresponding to the learning document selected using a subtree included in the correct rhetorical structure tree and a weight vector for a feature vector extracted from the subtree line A pair of subtrees included in the correct rhetorical structure tree, wherein the difference between the scores calculated using the feature vector and the weight vector is maximized, and the learning tree A parameter learning unit that updates the weight vector based on a feature vector extracted from each column of the subtree included in a pair with a subtree column of the rhetorical structure tree corresponding to a document;
An iterative determination unit that repeats the update by the parameter learning unit a predetermined number of times;
A parameter learning device.   The program for functioning a computer as each part of the parameter learning apparatus of Claim 3.

Images