JP2016062181A - Weight vector learning device, summary generation device, method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To allow quality of a single document summary to be improved.SOLUTION: A feature quantity extraction part 24 extracts a feature quantity every pair of a text unit included in an inputted single document. A discourse analysis part 26, on the basis of the extracted feature quantity every pair, an initial value of a weight vector, or value of the weight vector updated previously, generates a prediction discourse analytic tree. A summary extraction part 28, on the basis of the generated prediction discourse analytic tree, generates a summary of a single document. A summary comparison part 30 compares the generated summary with a correct answer of a summary prepared in advance with respect to a single document. A weight vector update part 32, when as a result of comparison, there is a difference between the summary and the correct answer of the summary, updates the value of the weight vector. A determination repetition part 34, repeats the generation by the discourse analysis part, generation by the summary extraction part, the comparison by the summary comparison part, and update by the weight vector update part until a predetermined termination condition is satisfied.SELECTED DRAWING: Figure 4

Description

  The present invention relates to a weight vector learning apparatus, method, and program, and more particularly, a weight vector learning apparatus, method, and program for learning a weight vector for summarizing an input document, and an input document. The present invention relates to a summary generation apparatus, method, and program.

  Abstracts have been an important research subject for many years in the field of natural language processing and have become increasingly important in recent years due to the rapid increase in information. Depending on the number of documents handled, summaries can be classified into two types: single document summaries and multiple document summaries. Single document summaries deal only with a single document, and multiple document summaries deal with a set of related documents.

  Non-Patent Document 1 is a previous study on single document summarization, which considers a document as a rhetorical structure theory discourse tree (hereinafter abbreviated as “RST-DT”). A text unit is selected according to the extracted priority rank. A document compression method that directly models the probability of summarization given RST-DT by a noisy channel model is also known (Non-Patent Document 2).

  By considering the summary as an optimization problem for a given discourse analysis tree, both discourse relationships and optimization can be considered. Recently, it is also known that a summary is formulated using an integer linear programming problem as a knapsack problem for a tree structure (Non-patent Document 3). According to the reported results, it is possible to surpass the state-of-the-art method when based on the correct discourse analysis tree. Non-Patent Document 4 is one of the most advanced analysis methods, and is also used in the discourse analysis of the present invention.

Marcu, Daniel. “Improving summarization through rhetorical parsing tuning.” The 6th Workshop on Very Large Corpora. 1998. Daum III, Hal, and Daniel Marcu. “A noisy-channel model for document compression.” Proceedings of the 40th Annual Meeting on Association for Computational Linguistics. Association for Computational Linguistics, 2002. Hirao, Tsutomu, et al. “Single-Document Summarization as a Tree Knapsack Problem.” Proceedings of the 2013 Conference on Empirical Methods in Natural Language. Association for Computational Linguistics. 2013. McDonald, Ryan, Koby Crammer, and Fernando Pereira. “Online large-margin training of dependency parsers.” Proceedings of the 43rd Annual Meeting on Association for Computational Linguistics. Association for Computational Linguistics, 2005.

  The methods of Non-Patent Documents 1 and 2 consider discourse relations, but have a problem that optimization may fail.

  In addition, discourse analysis trees (correct discourse analysis trees) generated by humans based on some predefined principles are not intended for use in practical tasks such as summarization. In such a case, there is a problem that the answer is not always correct.

  Therefore, the weight vector used for discourse analysis developed based on annotated data may be inappropriate for summarization. Discourse analysis aims to generate an parse tree that matches the annotated corpus as much as possible, but the purpose of the summarization is to obtain a high-quality summary, and the accuracy of the discourse parse tree is not necessarily related. This is because there is not.

  In short, discourse relations play an important role in extracting high-quality summaries, but it is difficult to create practical summaries in a simple way that uses discourse analysis trees as they are. There is a problem.

  The present invention provides a weight vector learning apparatus, method, and program that can solve the above-mentioned problems and can learn a discourse analysis weight vector that improves the quality of a single document summary. The purpose is to do.

  Another object of the present invention is to provide a summary generation apparatus, method, and program capable of improving the quality of a single document summary.

  In order to achieve the above object, a weight vector learning device according to a first invention includes a feature amount extraction unit that extracts a feature amount for each pair of text units included in an input single document, and the feature Based on the feature value for each pair extracted by the quantity extraction unit and the initial value of the weight vector or the value of the weight vector updated last time, the most important text unit of the single document is determined as a root node. Dependent structure of each text unit of the single document, wherein each text unit of the single document is a node, and nodes corresponding to the text units having a modification relationship are connected by an edge. A discourse analysis unit that generates a predictive discourse analysis tree that is a discourse analysis tree based on the above; a summary extraction unit that generates a summary of the single document based on the predictive discourse analysis tree generated by the discourse analysis unit; A summary comparison unit that compares the summary generated by the summary extraction unit with a correct summary prepared in advance for the single document; the summary generated by the summary extraction unit by the summary comparison unit; When there is a difference between the summary and the correct answer of the summary prepared in advance for one document, the weight vector update unit that updates the value of the weight vector and the discourse analysis unit A repetition determination unit that repeats generation, generation by the summary extraction unit, comparison by the summary comparison unit, and update by the weight vector update unit until a predetermined end condition is satisfied. Yes.

  A weight vector learning method according to a second invention includes a feature amount extraction unit, a discourse analysis unit, a summary extraction unit, a summary comparison unit, a weight vector update unit, and an iterative determination unit. A weight vector learning method in the apparatus, wherein the feature amount extraction unit extracts a feature amount for each pair of text units included in an input single document, and the discourse analysis unit includes the feature amount. Based on the feature value for each pair extracted by the extraction unit and the initial value of the weight vector or the value of the weight vector updated last time, the most important text unit of the single document is set as a root node. And each text unit of the single document is defined as each node, and nodes corresponding to the text units having a modification relationship are connected by edges. A predicted discourse analysis tree that is a discourse analysis tree based on a structure is generated, and the summary extraction unit generates a summary of the single document based on the predicted discourse analysis tree generated by the discourse analysis unit, and the summary comparison The summary compares the summary generated by the summary extraction unit with the correct summary prepared in advance for the single document, and the weight vector update unit is generated by the summary extraction unit by the summary comparison unit. When the summary and the correct answer of the summary prepared in advance for the single document are different, if the summary and the correct answer of the summary are different, the value of the weight vector is updated, and the iterative determination unit The generation by the discourse analysis unit, the generation by the summary extraction unit, the comparison by the summary comparison unit, and the update by the weight vector update unit until a predetermined end condition is satisfied Return Ri.

  According to the first and second inventions, the feature amount extraction unit extracts a feature amount for each pair of text units included in the input single document, and the discourse analysis unit extracts each pair. The most important text unit of a single document is the root node and each text of the single document is based on the feature quantity of the text and the initial value of the weight vector or the previously updated weight vector value. Generate a predictive discourse analysis tree, which is a discourse analysis tree based on the dependency structure of each text unit of a single document, in which the unit is each node and the nodes corresponding to each text unit having a modification relationship are connected by an edge. The summary extraction unit generates a summary of a single document based on the generated predicted discourse analysis tree, and the summary comparison unit compares the generated summary with the correct summary prepared in advance for the single document. And heavy When the vector update unit compares the summary generated by the summary extraction unit with the summary correct answer prepared in advance for a single document, and there is a difference between the summary and the summary correct answer, the weight vector value is set. The repetitive determination unit repeats the generation by the discourse analysis unit, the generation by the summary extraction unit, the comparison by the summary comparison unit, and the update by the weight vector update unit until a predetermined end condition is satisfied.

  As described above, according to the first and second inventions, the feature quantity is extracted for each pair of text units included in the input single document, and the feature quantity and the weight vector for each extracted pair are extracted. Or a previously updated weight vector value, a predicted discourse analysis tree is generated, a single document summary is generated based on the generated predicted discourse analysis tree, and the generated summary If there is a difference between the summary and the correct answer of the summary prepared in advance for a single document, there is a difference between the summary and the correct answer of the summary. It is possible to learn discourse analysis weight vectors that improve

  In the first and second aspects of the invention, the weight vector update unit compares the summary generated by the summary extraction unit with a correct summary prepared for the single document by the summary comparison unit. As a result, when there is a difference between the summary and the correct answer of the summary, the predicted discourse analysis tree node corresponding to the text unit appearing only in the correct answer of the summary is moved up, and the text unit appearing only in the summary A new discourse analysis tree is generated by lowering the corresponding node of the predicted discourse analysis tree, and the discourse analysis tree generated for the single document based on the weight vector is the new discourse analysis tree. The value of the weight vector can be updated.

  According to a third aspect of the present invention, there is provided a summary generation device including a feature amount extraction unit that extracts a feature amount for each pair of text units included in an input single document, and each pair extracted by the feature amount extraction unit. A discourse analysis unit that generates a discourse analysis tree based on the feature amount of the first and the weight vector values updated by the weight vector learning device of the first invention, and a discourse analysis tree generated by the discourse analysis unit And a summary extractor for generating a summary of the single document.

  A summary generation method according to a fourth invention is a summary generation method in a summary generation apparatus including a feature amount extraction unit, a discourse analysis unit, and a summary extraction unit, wherein the feature amount extraction unit is input The feature amount is extracted for each pair of text units included in the single document, and the discourse analysis unit extracts the feature amount for each pair extracted by the feature amount extraction unit and the weight vector of the second invention. A discourse analysis tree is generated based on the value of the weight vector updated by the learning method, and the summary extraction unit calculates a summary of the single document based on the discourse analysis tree generated by the discourse analysis unit. Generate.

  According to the third and fourth inventions, the feature amount extraction unit extracts a feature amount for each pair of text units included in the input single document, and the discourse analysis unit extracts each pair. And a discourse analysis tree is generated based on the feature vector and the weight vector value updated by the weight vector learning device or the weight vector learning method, and the summary extraction unit generates a single discourse analysis tree based on the generated discourse analysis tree. Generate document summaries.

  As described above, according to the third and fourth inventions, the feature amount is extracted for each pair of text units included in the input single document, and the feature amount and the weight vector for each extracted pair are extracted. A discourse analysis tree is generated based on the weight vector value updated by the learning device or the weight vector learning method, and a single document summary is generated based on the generated discourse analysis tree. Summarization quality can be improved.

  Moreover, the program of this invention is a program for functioning a computer as each part which comprises said weight vector learning apparatus and summary production | generation apparatus.

  As described above, according to the weight vector learning device, method, and program of the present invention, feature quantities are extracted for each pair of text units included in an input single document, and each extracted pair is extracted. Generates a predicted discourse parse tree based on the feature quantity and the initial value of the weight vector or the previously updated weight vector value, and generates a single document summary based on the generated predicted discourse parse tree If there is a difference between the summary and the correct summaries prepared for a single document, and the summaries are different from the correct answers, the updating of the weight vector value is repeated. It is possible to learn discourse analysis weight vectors that improve the quality of single document summaries.

  Further, according to the summary generation device, method, and program of the present invention, the feature amount is extracted for each pair of text units included in the input single document, and the extracted feature amount for each pair; It is possible to improve the quality of a single document summary by generating a discourse analysis tree based on the learned weight vector values and generating a single document summary based on the generated discourse analysis tree. it can.

It is a figure which shows the detail of the algorithm which shows the general system of MIRA. It is a figure which shows the detail of the algorithm of weight vector learning. It is a figure which shows the detail of a summary algorithm. It is a block diagram which shows the functional structure of the weight vector learning apparatus which concerns on embodiment of this invention. It is a block diagram which shows the functional structure of the summary production | generation apparatus which concerns on embodiment of this invention. It is a flowchart figure which shows the weight vector learning process routine in the weight vector learning apparatus which concerns on embodiment of this invention. It is a flowchart figure which shows the summary production | generation processing routine in the summary production | generation apparatus which concerns on embodiment of this invention. It is a figure which shows the result about RST-DT. It is a figure which shows the result about TAC-2011.

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

<Principle of Embodiment of the Present Invention>
First, the principle of the weight vector learning device according to the embodiment of the present invention will be described. The weight vector learning apparatus according to the present embodiment is composed of two main parts. The first is a discourse analysis part, and the second is a summary part. The originality of the weight vector learning device according to the present embodiment is its comprehensiveness.

  In this embodiment, each time a weight vector used in discourse analysis is trained in discourse analysis training, a new predicted discourse analysis tree for learning data is generated. In addition, a summary is extracted based on a new predicted discourse parse tree generated each time it is trained, and is compared with a standard summary representing a summary of correct answers generated by humans. Then, the summary based on the predicted discourse analysis tree is compared with the standard summary, the predicted discourse analysis tree is adjusted based on the difference between the two based on the comparison result, and the adjusted discourse analysis tree is set as a new discourse analysis tree. Then, based on the predicted discourse analysis tree, the new discourse analysis tree, and the weight vector, a weight vector used in a new discourse analysis is generated as learning data for repeated iterations. After repeating the iteration several times, the weight vector used in the updated discourse analysis optimized for the specific task of summarization can be obtained. The weight vector used in the updated discourse analysis that can be obtained is used for the discourse analysis of the document, and is based on the discourse analysis tree generated from the document and the weight vector used in the discourse analysis. A summary can be extracted. Note that the weight vectors used in the updated discourse analysis are optimized for summarizing with higher accuracy than that of high discourse analysis accuracy. This overcomes the gap between discourse analysis and summarization.

  Next, task-oriented discourse analysis will be described. Although a plurality of types of discourse analysis algorithms are known, a dynamic bottom-up discourse analysis algorithm (Non-Patent Document 4) is adopted in the present embodiment. The discourse analysis algorithm was originally designed for sentence analysis, but can easily be shifted to discourse analysis.

Let the input document be X = x ₁ , x ₂ ,. . . , _Xn . Here, x _i indicates the i-th sentence in the input document. If there is a dependency from the sentence x _{i to the} sentence x _j , it is expressed as (i, j). In this case, x _i is a head and x _j is subordinate. Considering a sentence unit as one node, if each node has only one head, it is assumed that the dependency relationship is independent between different pairs. Assigning a score to each node pair and indicating the probability that there is a dependency between the two nodes, the problem of discourse analysis of the document is to connect all the nodes and the sum of all the scores of the edges is the highest It becomes a problem to discover trees. This will find a maximum spanning tree in a directed graph. The Eisner algorithm requires two binary variables to record direction and element integrity during discourse analysis, and discourse analysis takes O (n ³ ) time (Non-Patent Document 4).

  Here, what is important is to define scores for all possible edges s. A sentence pair (i, j) can be expressed by a feature vector f (i, j) and a score s (i, j) expressed by the following equation.

  The main point of the algorithm shown in FIG. 1 that shows the general scheme of MIRA is to define weight vectors for features. In the present embodiment, the weight vector is used using MIRA.

To train.

  In the present embodiment, the procedure shown in FIG. 1 is repeated several times, and the final weight vector is stored in the entire history.

By using the arithmetic mean of, excessive adaptation to the learning data is avoided.

  Next, a case where the weight vector is updated according to the training example will be described. The conventional discourse analysis method uses weight vectors according to the correct discourse analysis tree.

Was trying to update. That is, the weight vector is updated according to the following formula (1) and the following formula (2).

here,

Is a predicted discourse analysis tree, y _i is a correct discourse analysis tree, a loss function is a Hamming function, and is expressed by the following equation (2). Note that the correct discourse analysis tree is generated by a human based on a pre-defined principle.

  In the present embodiment, instead of the correct discourse analysis tree of the above formula (1) and the above formula (2), a new discourse analysis tree is used, and according to the following formula (3) and the following formula (4): Train the weight vector.

In the above equation (3), the correct discourse analysis tree y _i of the above equation (1) is replaced with a new discourse analysis tree y ^* _i . New discourse analysis tree y ^* _i is the predicted discourse analysis tree

Depending on the difference between the summary generated based on the standard summary and the standard summary,

It is generated by moving up and down each of the nodes contained within. In the adjustment of the discourse analysis tree, attention is paid to nodes corresponding to text units appearing in the summary. Details are shown in Algorithm 2 shown in FIG.

  Note that the corresponding nodes in “4)” and “5)” in Algorithm 2 shown in FIG. 2 are sentences, sections, or other basic elements in the document. Define. In the present embodiment, the RST-DT uses a basic document unit (a unit shorter than a sentence and generally corresponds to a section; hereinafter abbreviated as “EDU” (Elementary Discourse Unit)). By repeating the process of algorithm 2 until the repetition condition is satisfied, the arithmetic average of the history of the weight vectors can be used as the final weight vector. The weight vector thus obtained is particularly optimized for summary generation.

  Here, nodes corresponding to text units appearing only in the standard summary are moved up. This is because, in Algorithm 3, which is the summarization algorithm shown in FIG. 3, high scores are assigned to higher nodes (nodes close to the root) in the discourse analysis tree when a summary is generated. Also, the node corresponding to the text unit that appears only in the summary generated based on the predicted discourse analysis tree is carried down. This is because lower scores are assigned to lower nodes in the discourse analysis tree when a summary is generated.

  Algorithm 3 shown in FIG. 3 is a summarization algorithm based on a discourse analysis tree in which constraints are weakened to consider discourse relations.

  Every text unit is assigned an initial score according to the word contained in the text unit and the position in the discourse analysis tree of the text unit. These text units are re-ranked according to their distance from the text units in the summary. If the node corresponding to the text unit included in the summary is set close, a high score is assigned and the possibility of being selected as the text unit included in the summary at the next iteration increases. Thus, in this embodiment, a consistent text spread is generated under relatively weak constraints. The ranking and selection is repeated until a summary is obtained that satisfies a predetermined termination condition. Here, it is necessary to adjust two parameters. α∈ (0,1] indicates the influence of the role on the node.As the role for the node, the score of each head node increases according to what depends on it, and the score of the dependent node decreases. The influence of the distance to the node in the summary is shown: the closer to the node corresponding to the text unit included in the current summary, the more the score is determined by the value of β.

  Next, feature vectors used in the present embodiment will be described. In this embodiment, feature vectors are used to represent edges between text units. Many features have been developed through previous research. For example, there are a set of words (bag of words), a total frequency (token frequency), a part-of-speech tag, a generation rule, a co-occurrence, and the like, all of which can be used in this embodiment. In the present embodiment, attention is not paid to feature selection used for feature vectors. Therefore, in this embodiment, in order to confirm the efficiency of the algorithms 2 and 3 described above, some simple types of features shown in the following (a) to (d) are used.

(A) word: whether the current element contains a word (pronoun, noun, verb, modulo) of a specific part of speech tag (b) length: number of words in the element (c) relative position: pair Whether (d) word pair: vector product between two elements

  Note that the number of features used for the feature vector varies depending on the data used. In the present embodiment, feature selection is performed using a training set of RST-DT. After deleting all features with a frequency less than 5, a total of 5000 or more features were obtained. After feature extraction, each pair of text units will be represented by a binary vector. Each weight of the weight vector is a weight of each feature of the feature vector. The relative position represents the context of the text unit in the document.

<Configuration of Weight Vector Learning Device According to Embodiment of the Present Invention>
Next, the configuration of the weight vector learning device according to the embodiment of the present invention will be described. As shown in FIG. 4, the weight vector learning device 100 according to the embodiment of the present invention includes a CPU, a RAM, a ROM for storing a program and various data for executing a weight vector learning processing routine described later, Can be configured with a computer including Functionally, the weight vector learning apparatus 100 includes an input unit 10, a calculation unit 20, and an output unit 90 as shown in FIG.

  The input unit 10 receives one or more single documents and a standard summary for each single document, and stores the standard summary for each single document in the standard summary storage unit 22.

  The calculation unit 20 includes a standard summary storage unit 22, a feature amount extraction unit 24, a discourse analysis unit 26, a summary extraction unit 28, a summary comparison unit 30, a weight vector update unit 32, a repetition determination unit 34, And a weight vector storage unit 36.

  The standard summary storage unit 22 stores a standard summary for each single document received by the input unit 10.

  The feature quantity extraction unit 24 extracts, for each single document, a feature vector composed of a plurality of feature quantities for each pair composed of a combination of text units included in the single document. In the present embodiment, the text unit is a sentence unit. Specifically, for each pair, for each of the text unit corresponding to the upper node as the head of the pair and the text unit corresponding to the lower node subordinate to the upper node of the pair, Each of the features (a) to (c) is extracted, and the cartesian product of the word set included in the text unit corresponding to the upper node and the text unit corresponding to the lower node is extracted from the above (d) ).

  The discourse analysis unit 26 extracts the initial value of the weight vector stored in a memory (not shown) or the value of the weight vector updated in the previous process and the feature amount extraction unit 24 for each single document. Based on each of the feature vectors for each pair of text units of the single document, the most important text unit of the single document as a root node and each text unit of the single document as A predicted discourse analysis tree, which is a discourse analysis tree based on the dependency structure of each text unit of a single document, is generated by connecting nodes corresponding to each text unit having a modification relationship with edges.

  The summary extraction unit 28, for each single document, the predicted discourse analysis tree of the single document generated by the discourse analysis unit 26, the parameter α indicating the influence of the role on the node, and the distance between the nodes in the summary. Based on the parameter β indicating the influence, ranking is performed for each text unit included in the single document, a ranked list R is obtained, and the text unit included in the summary is obtained from the obtained ranked list R. The selection is repeated until the number of words included in the generated summary is equal to or greater than a predetermined threshold.

  The summary comparison unit 30 compares, for each single document, the summary of the single document generated by the summary extraction unit 28 and the standard summary of the single document stored in the standard summary storage unit 22. Then, based on the comparison result, the predicted discourse analysis tree of the single document is adjusted to generate a new discourse analysis tree. Specifically, in the predicted discourse analysis tree, the node corresponding to the text unit appearing only in the standard summary is moved up, and the node corresponding to the text unit appearing only in the summary generated by the summary extraction unit 28 is moved down.

  The weight vector update unit 32, for each single document, based on the predicted discourse analysis tree of the single document, the new discourse analysis tree of the single document, and the weight vector, The weight vector is updated according to the above equation (4) and stored in a memory (not shown).

  The repetition determination unit 34 determines whether or not the weight vector is updated in the weight vector update unit 32 for the single document to be processed. If the weight vector has been updated, the discourse analysis unit 26 and the summary extraction unit 28, the processes of the summary comparison unit 30 and the weight vector update unit 32 are repeated. If not updated, the repetition process is terminated, and the average of the history of the updated weight vectors is calculated as the updated weight vector. And stored in a memory (not shown). In addition, when the repetition determination unit 34 finishes the repetition process for all target single documents, the repetition weighting unit 34 calculates the average of the updated weight vectors stored in the memory (not shown) as the weight after learning. The vector is stored in the weight vector storage unit 36 as a vector and is output to the output unit 90.

  The weight vector storage unit 36 stores the learned weight vector acquired by the repetition determination unit 34.

<Configuration of Summary Generation Device According to Embodiment of the Present Invention>
Next, the configuration of the summary generation device according to the embodiment of the present invention will be described. As shown in FIG. 5, summary generation apparatus 200 according to the embodiment of the present invention includes a CPU, a RAM, and a ROM that stores a program and various data for executing a summary generation processing routine described later. Can be configured with a computer. Functionally, the summary generation apparatus 200 includes an input unit 210, a calculation unit 220, and an output unit 290 as shown in FIG.

  The input unit 210 accepts a single document that is a summary generation target.

  The calculation unit 220 includes a feature amount extraction unit 224, a discourse analysis unit 226, a summary extraction unit 228, and a weight vector storage unit 236.

  The feature quantity extraction unit 224 extracts feature vectors composed of a plurality of feature quantities for each pair of combinations of text units included in a single document, like the feature quantity extraction unit 24 in the weight vector learning device 100. To do.

  The discourse analysis unit 226 is based on the value of the weight vector stored in the weight vector storage unit 236 and each feature vector of each pair of text units of the single document extracted by the feature amount extraction unit 224. Generate a discourse analysis tree.

  The summary extraction unit 228 is based on a single document discourse analysis tree generated by the discourse analysis unit 226, a parameter α indicating the influence of the role on the node, and a parameter β indicating the influence of the distance to the node in the summary. A ranking of each text unit included in a single document, obtaining a ranked list R, and selecting a text unit included in the summary from the obtained ranked list R; A summary is generated by repeating until the number of words is equal to or greater than a predetermined threshold, and is output to the output unit 290.

  The weight vector storage unit 236 stores the same weight vector as the weight vector stored in the weight vector storage unit 36 of the weight vector learning device 100.

<Operation of Weight Vector Learning Device According to Embodiment of the Present Invention>
Next, the operation of the weight vector learning device 100 according to the first embodiment of the present invention will be described. When the input unit 10 receives each single document and the standard summary for each single document, the standard summary for each single document is stored in the standard summary storage unit 22 and the weight vector learning device 100. Executes the weight vector learning processing routine shown in FIG.

  First, in step S100, each standard summary of a single document received by the input unit 10 stored in the standard summary storage unit 22 is read.

  Next, in step S102, for each single document received by the input unit 10, a feature vector composed of a plurality of feature amounts is extracted for each pair of text units included in the single document.

  Next, in step S104, an initial value of a weight vector stored in a memory (not shown) is read.

  Next, in step S106, for the single document to be processed, the feature vector for each pair of text units acquired in step S102 and the initial value of the weight vector acquired in step S104, or the previous processing in step S112. A predicted discourse analysis tree is generated based on the obtained weight vector value.

  In step S108, a summary is generated for the single document to be processed based on the predicted discourse analysis tree acquired in step S106.

  Next, in step S110, for the single document to be processed, the standard summary for the single document acquired in step S100 is compared with the summary acquired in step S108, and based on the comparison result, A new discourse analysis tree is generated by moving each of the nodes included in the predicted discourse analysis tree acquired in step S106 up and down.

  Next, in step S112, for the single document to be processed, the initial value of the weight vector acquired in step S104, or the value of the weight vector acquired in step S112 in the previous process, and the predicted discourse obtained in step S106. Based on the analysis tree and the new discourse analysis tree acquired in step S110, the value of the weight vector is updated in accordance with the above equations (3) and (4), and the value of the weight vector is stored in the memory (illustrated). (Omitted).

  Next, in step S114, it is determined whether or not a predetermined end condition is satisfied. If the end condition is satisfied, the process proceeds to step S116. If the end condition is not satisfied, step S114 is performed. The process proceeds to S106, and the processes from Step S106 to Step S114 are repeated.

  Next, in step S116, for the single document to be processed, the average of the weight vector values for each repetition obtained in step S112 is obtained as the updated weight vector value and stored in a memory (not shown). Remember.

  Next, in step S118, it is determined whether or not the processing from step S106 to step S116 has been completed for all the single documents received by the input unit 10. If the processing from step S106 to step S116 has been completed for all single documents, the process proceeds to step S120, and if the processing from step S106 to step S116 has not been completed for all single documents. The single document to be processed is changed, and the processing from step S106 to step S118 is repeated.

  Next, in step S120, the average of the values of the updated weight vectors acquired in step S116 is acquired as the learned weight vector values.

  Next, in step S122, the learned weight vector value acquired in step S120 is stored in the weight vector storage unit 36 and output to the output unit 90, and the weight vector learning processing routine is terminated.

<Operation of Summary Generation Device According to Embodiment of the Present Invention>
Next, the operation of the summary generation apparatus 200 according to the embodiment of the present invention will be described. When the input unit 210 accepts a single document, the summary generation apparatus 200 executes a summary generation processing routine shown in FIG.

  First, in step S200, the same weight vector value as that stored in the weight vector storage unit 236 and stored in the weight vector storage unit 36 of the weight vector learning device 100 is read.

  Next, in step S202, a feature vector is extracted for each pair of text units included in the single document acquired by the input unit 210, as in step S102.

  Next, in step S204, a discourse analysis tree is generated based on the weight vector value acquired in step S200 and the feature vector for each pair of text units acquired in step S202.

  Next, in step S206, a summary is generated based on the discourse analysis tree acquired in step S204.

  Next, in step S208, the summary acquired in step S206 is output from the output unit 290, and the summary generation processing routine ends.

<Experimental result>
Next, the results of experiments on RST-DT and TAC2011 summary task data will be described. The RST-DT contains 385 documents (training 347, test 38) from WSJ. Of the 385 documents, 30 have a summary written by humans and 150 have a summary that includes a manually selected EDU. Long documents take a long time to analyze discourse, and short documents do not contain useful information, so we excluded 30 documents with manually written summaries and documents that were too long or too short. 100 documents were selected for training and 30 documents were selected as test data.

  In order to adjust the parameters of the summarization algorithm, a small development set containing 10 documents was created, and the parameters were adjusted according to the learning data. Each parameter (α, β, number of repetitions) is set to (0.7, 0.1, 6).

  Furthermore, the system in the present embodiment was inspected using the TAC2011 multi-document summary task data. There are 44 document sets, and each set contains two folders, a summary folder A and an updated summary folder B. Each folder contains 10 documents. It is. The system in this embodiment is designed only for single document summarization. A medium-sized document is selected from each folder A as a representative of the entire document set.

  The method described in this embodiment was compared with a state-of-the-art tree-structure-based single document summarization system. FIG. 8 shows the results for RST-DT. DST (Discourse Supervised Tree-based summarization) is a system in this embodiment, and TKP (Tree Knapsack Problem) is a method of the tree structure knapsack problem method of Non-Patent Document 3. M is also derived from Non-Patent Document 3, which implements the method of Non-Patent Document 1. MEAD is a baseline system provided by the MEAD toolkit, which uses a ranking method and MMR (Maximal Marginal Relevance) to extract summaries. Yes. G attached here means that a summary is generated based on a correct discourse analysis tree, and H means that a discourse analysis tree generated with a weight vector is used. The results were analyzed with the ROUGE toolkit.

  The method described in this embodiment is superior to any other method in terms of performance, and is not only based on a syntax tree obtained by automatic parsing, but also a correct syntax tree generated manually. The same applies to those based on. The weight vector used in the discourse analysis in the present embodiment is learned for summarization, and succeeds in improving the quality of the summarization. The correct ROUGE score (automated assessment scale for summaries based on word recall. Higher values between 0 and 1 are better) suggests that the correct discourse parse tree is not suitable for summarization as a whole. Proven not. In addition, experiments were also performed on TAC-2011 data in order to verify the performance of the system according to the present embodiment. FIG. 9 shows the result.

  In FIG. 9, the whole sentence is used as a basic element. By using a small basic element, a short summary sentence can be obtained. In order to verify this assumption, SPADE is used to divide the sentence into EDUs, and the result based on EDU (DST-EDU) is shown in FIG. EDU-based systems do not improve performance, but are still comparable to multi-document summarization methods.

  As described above, according to the weight vector learning device according to the embodiment of the present invention, the feature vector is extracted for each pair of text units included in the input single document, and for each extracted pair Based on the feature vector and the initial value of the weight vector or the previously updated weight vector value, a predicted discourse analysis tree is generated, and a single document summary is generated based on the generated predicted discourse analysis tree. If the summarization and the correct answer of the summary are different as a result of comparing the generated summary and the correct answer of the summary prepared in advance for a single document, it is determined in advance that the value of the weight vector is updated. By repeating until the end condition is satisfied, it is possible to learn a discourse analysis weight vector that improves the quality of the single document summary.

  Further, according to the summary generation device according to the embodiment of the present invention, the feature vector is extracted for each pair of text units included in the input single document, and the extracted feature vector for each pair; It is possible to improve the quality of a single document summary by generating a discourse analysis tree based on the learned weight vector values and generating a single document summary based on the generated discourse analysis tree. it can.

  In addition, by selecting a plurality of sentences including important information and considering a discourse relationship between sentences, it is possible to obtain a consistent and high-quality summary.

  In addition, learning data is dynamically generated by adjusting the discourse analysis tree according to the summary provided by a human. In other words, the system architecture in the present embodiment is very different from the conventional method, and the learning data is different for each iteration.

  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 present embodiment, the case where the text unit is a sentence has been described, but the present invention is not limited to this, and the text unit may be a section or another basic element in a document. Here, rhetoric structure theory (RST) uses EDU as “text unit”. In a compound sentence, the main clause and the subordinate clause are each one EDU, so the EDU is almost equivalent to a “single sentence” in Japanese.

  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 Standard summary memory | storage part 24 Feature-value extraction part 26 Discourse analysis part 28 Summary extraction part 30 Summary comparison part 32 Weight vector update part 34 Repeat determination part 36 Weight vector memory | storage part 90 Output part 100 Weight vector learning apparatus 200 summary generation device 210 input unit 220 calculation unit 224 feature amount extraction unit 226 discourse analysis unit 228 summary extraction unit 236 weight vector storage unit 290 output unit

Claims (7)

A feature amount extraction unit that extracts a feature amount for each pair of text units included in the input single document;
Based on the feature value for each pair extracted by the feature amount extraction unit and the initial value of the weight vector or the value of the weight vector updated last time, the most important text unit of the single document is determined. Each text unit of the single document having a root node, each text unit of the single document as each node, and nodes corresponding to each other between the text units having a qualifying relationship are connected by an edge. A discourse analysis unit that generates a predictive discourse analysis tree that is a discourse analysis tree based on the dependency structure;
A summary extraction unit that generates a summary of the single document based on the predicted discourse analysis tree generated by the discourse analysis unit;
A summary comparison unit that compares the summary generated by the summary extraction unit with a correct summary prepared in advance for the single document;
As a result of comparing the summary generated by the summary extraction unit by the summary comparison unit and the correct answer of the summary prepared in advance for the single document, there is a difference between the summary and the correct answer of the summary, A weight vector updating unit for updating the value of the weight vector;
A repeat determination unit that repeats generation by the discourse analysis unit, generation by the summary extraction unit, comparison by the summary comparison unit, and update by the weight vector update unit until a predetermined end condition is satisfied,
A weight vector learning device.   The weight vector update unit compares the summary generated by the summary extraction unit with the summary correct answer prepared in advance for the single document by the summary comparison unit. And a node of the predicted discourse analysis tree corresponding to the text unit appearing only in the correct answer of the summary is raised, and a node of the predicted discourse analysis tree corresponding to the text unit appearing only in the summary is lowered. A new discourse analysis tree is generated, and the value of the weight vector is updated so that the discourse analysis tree generated for the single document based on the weight vector becomes the new discourse analysis tree. The described weight vector learning device. A feature amount extraction unit that extracts a feature amount for each pair of text units included in the input single document;
A discourse analysis unit that generates a discourse analysis tree based on the feature amount of each pair extracted by the feature amount extraction unit and the value of the weight vector updated by the weight vector learning device according to claim 1. When,
A summary extraction unit that generates a summary of the single document based on the discourse analysis tree generated by the discourse analysis unit;
A summary generation device. A weight vector learning method in a weight vector learning device including a feature amount extraction unit, a discourse analysis unit, a summary extraction unit, a summary comparison unit, a weight vector update unit, and a repetition determination unit,
The feature amount extraction unit extracts a feature amount for each pair of text units included in the input single document,
The discourse analysis unit, based on the feature value for each pair extracted by the feature value extraction unit and the initial value of the weight vector or the value of the weight vector updated last time, The single text unit in which the important text unit is a root node, each text unit of the single document is a node, and nodes corresponding to the text units having a modification relationship are connected by an edge. Generate a predictive discourse analysis tree, which is a discourse analysis tree based on the dependency structure of each text unit of the document,
The summary extraction unit generates a summary of the single document based on the predicted discourse analysis tree generated by the discourse analysis unit;
The summary comparison unit compares the summary generated by the summary extraction unit with the correct answer of the summary prepared in advance for the single document,
The weight vector update unit compares the summary generated by the summary extraction unit by the summary comparison unit with the correct summary prepared in advance for the single document. If there is a difference, the value of the weight vector is updated,
The repetition determination unit repeats the generation by the discourse analysis unit, the generation by the summary extraction unit, the comparison by the summary comparison unit, and the update by the weight vector update unit until a predetermined end condition is satisfied. Weight vector learning method.   Updating the weight vector value by the weight vector updating unit compares the summary generated by the summary extraction unit with the correct summary prepared for the single document by the summary comparison unit. As a result, if there is a difference between the summary and the correct answer of the summary, the node of the predicted discourse analysis tree corresponding to the text unit appearing only in the correct answer of the summary is moved up, and corresponding to the text unit appearing only in the summary Generating a new discourse analysis tree in which the nodes of the predicted discourse analysis tree are lowered, and the discourse analysis tree generated for the single document based on the weight vector is the new discourse analysis tree. 5. The weight vector learning method according to claim 4, wherein the weight vector value is updated. A summary generation method in a summary generation device including a feature amount extraction unit, a discourse analysis unit, and a summary extraction unit,
The feature amount extraction unit extracts a feature amount for each pair of text units included in the input single document,
The discourse analysis unit, based on the feature amount for each pair extracted by the feature amount extraction unit and the value of the weight vector updated by the weight vector learning method according to claim 4 or 5, Produces
The summary extraction unit generates a summary of the single document based on a discourse analysis tree generated by the discourse analysis unit.   A program for causing a computer to function as each unit constituting the weight vector learning apparatus according to claim 1 or 2, or the summary generation apparatus according to claim 3.