JP4106470B2 - Solution data editing processing apparatus and processing method - Google Patents

Description

  The present invention relates to edit processing of editable solution data and automatic summarization processing using a machine learning method using solution data in processing for automatically summarizing sentences by a machine learning method.

  In recent years, with the development of information technology, automatic summarization processing of sentences using computers has become popular. However, it can be considered that there is a difference in the tendency of the desired summary result depending on personal preference and use of the summary result.

  For example, in Non-Patent Document 1 below, Table 4 shows the recall rate and relevance rate as a measure of mutual evaluation with respect to the results obtained by summarizing each of a plurality of persons by extracting important sentences. As is clear from Table 4 of Non-Patent Document 1, in the case of the process of extracting 20 sentences, the evaluation (reproduction rate and relevance rate) of person-person (evaluator mutual) is evaluated by each evaluator A, B, C. The degree of coincidence is 50 to 70%, which is not so high, and it can be estimated that there are individual differences in the evaluation of the summary results.

In the following Non-Patent Document 2, Table 4 shows that the accuracy of cross-validation is the best for processing sets A, B, and C in the important sentence extraction processing by the Support Vector Machine. . The cross-validation shown in Table 4 of Non-Patent Document 2 can be identified with the process by the same evaluator, and it is unclear whether the persons who created the sets A, B, and C are the same, but at least the same time or the same It can be seen that it is better to create learning data by a person.
Itoyamahiko et al., "Important sentence extraction for lectures", Proc. Of the 7th Annual Conference of the Language Processing Society of Japan, 2001, pp.305-308 Tsutomu Hirao et al., “Important sentence extraction by Support Vector Machine”, Information Science Society Basic Paper 63-16, Information Society, 2001, pp.121-127

  In this way, it can be considered that there are individual differences and usage differences in the evaluation of the summary results. Therefore, in the automatic summarization process using the machine learning method, the summarization based on the same evaluation is not performed, but the user is specially selected. There is a need for a simplified summary. Therefore, the user must be able to freely edit the solution data to be a teacher.

  An object of the present invention is to realize a solution data editing process in which a user can arbitrarily edit a summary result or an evaluation of the summary result as solution data used in the machine learning method. And, it is to realize an automatic summarization processing technology that can perform summarization specialized for each user by using a machine learning method using the solution data.

  In order to achieve the above-described object, the present invention provides a method for the user to summarize information about what kind of summary result the user highly appreciated for the machine learning process using solution data prepared in advance. And edit its rating. In addition, by feedback of solution data edited by the user, the characteristics for each user are learned in the machine learning process, and the user-specific summarization can be performed.

The present invention is a solution data editing processing apparatus for editing solution data used in a process of automatically summarizing text as document data by a machine learning method, 1) a text storage means for storing text as document data; 3) summary display means for displaying the text acquired from the text storage means on a display device, extracting sentence data in a range specified by the user from the text, and displaying it as a user-specified summary of the text; Is information indicating the characteristics of the summary used as an evaluation, and whether the importance is attached to the fact that the short sentence is emphasized as a summary and the fact that the expression includes quantity is included in the summary Quantitative expression-oriented nature that indicates, method-oriented nature that indicates whether the summary includes expressions about the technique, About two or more properties including two or more properties, either a style-oriented property that indicates whether or not a particular style is emphasized, or a readability-oriented property that indicates whether a summary is easy to read , An item for inputting an evaluation value for each property, and an evaluation giving means for receiving an input of the evaluation value of the user for each property with respect to the user-specified summary; Solution data storage means for performing 5) a problem with the text and the user-specified summary, giving an evaluation value input by the user to the problem as a solution, generating solution data, and extracting a sentence from the text key sentence selection process to the selection of the state candidate condensates, important passage election to the state of the selection of any clauses removed phrase and candidate condensates from the text Processing, or the modification of the statement text in accordance with a predetermined transformation rule carried out any one of process variants a process for the summary candidates while the deformation generates a summary candidate of the text, the text and the summary A candidate candidate that is a summary candidate other than the user-specified summary is used as a problem, and solution data is generated by giving a bad evaluation indicating that the summary candidate is not the user-specified summary to the problem as a solution, It comprises an evaluation customization means for outputting solution data whose solution is an evaluation value input by a user and solution data whose solution is the bad evaluation to the solution data storage means.

  The present invention also provides an automatic summarization processing device for automatically summarizing text, which is document data, using a machine learning method, and includes 1) text storage means for storing the text and the text summary, and 2) the summary. Summary display processing means for displaying on a display device, 3) evaluation setting processing means for accepting an input of a user's evaluation for the summary and evaluating the summary, and 4) for the problem composed of the text and the summary Solution data output processing means for storing solution data generated by giving the evaluation as a solution in solution data storage means; 5) extracting a set of feature features of the problem and the solution from the solution data; Machine learning processing means for storing in the learning result data storage means learning result data for learning what kind of solution is likely to be obtained from the set, and 6) summary object Summary candidate generation processing means for inputting a text and generating a summary candidate from the input text; and 7) what kind of solution is likely to be obtained from the feature set by extracting a set of features from the input text and the summary candidate 8) summary candidate-estimated solution pair generation processing means for generating a pair of summary candidate and estimated solution based on the learning result data, and 8) an estimated solution is determined from the summary candidate-estimated solution pair Summarization selection processing means for selecting a pair with good evaluation and having the highest certainty factor, and summarizing the pair of candidate candidates.

  According to the present invention, a summary of text stored in the text storage means is displayed on a display device, and an input of a user's evaluation for the summary is received and the summary is evaluated. Then, solution data generated by assigning the evaluation as a solution to the problem composed of the text and the summary is stored in a solution data storage means, and a set of feature features of the problem and the solution are determined from the solution data. A pair is extracted, and learning result data obtained by learning what kind of solution is likely to be obtained from the pair is stored in the learning result data storage unit.

  After that, a text to be summarized is input, a summary candidate is generated from the input text, a set of features is extracted from the input text and the summary candidate, and what kind of solution is likely to be obtained from the set of features is learned. Estimate based on the result data, and generate a pair of summary candidate and estimated solution (summary candidate-estimated solution pair). Then, from the summary candidate-estimated solution pair, a pair whose estimated solution has a predetermined good evaluation and the highest certainty factor is selected, and the summary candidate of the pair is used as a summary.

  This makes it possible to machine-learn a summary that the user considers good using the user's evaluation of the displayed summary, and to perform a user-specific summary on the text that is input thereafter.

  Alternatively, the present invention includes 1) text storage means for storing text, 2) text display processing means for displaying the text on a display device, and 3) extracting sentence data in a range designated by the user from the text. A summary editing processing means for providing a user-specified summary of the text, and 4) an automatic summary generation process of either a process of generating a summary based on a predetermined rule or a process of generating a summary using a machine learning method Solution data generated by generating a summary of a text and giving a predetermined good evaluation as a solution indicating a good summary selected by the user for a problem composed of the text and the user-specified summary; To solve a problem composed of the text and the summary generated by the automatic summary generation process, which is composed of parts other than the user-specified summary. Solution data output processing means for storing, in a solution data storage means, solution data generated by giving a predetermined bad evaluation indicating that it is not the user-specified summary as a solution; 5) from the solution data, A machine learning processing means for extracting a set of the set and the solution, and storing learning result data in the learning result data storage means that learns what kind of solution is likely to be generated from the set; 6) summary candidate generation processing means for inputting a text to be summarized and generating a summary candidate from the input text; and 7) extracting a set of features from the input text and the summary candidate and how from the feature set. A summary candidate-estimated solution pair generation processing means for estimating whether a solution is likely to be a solution based on the learning result data and generating a pair of a summary candidate and an estimated solution; 8) the summary candidate-estimation Of a pair, the estimated solutions and confidence predetermined good evaluation to select the best pair, and a summary selection processing means to summarize the summary candidates of the pair.

  In the present invention, the text stored in the text storage means is displayed on a display device, and sentence data in a range specified by the user is extracted from the text to obtain a user-specified summary of the text. Then, the summary of the text is generated by an automatic summary generation process of either a process of generating a summary based on a predetermined rule or a process of generating a summary using a machine learning method, and the text and the user-specified summary are Solution data generated by giving a predetermined good evaluation indicating that the problem is a good summary selected by the user as a solution, a summary by the text and the automatic summary generation process, and the user Solution data storage means stores solution data generated by giving a predetermined bad evaluation indicating that it is not the user-specified summary to a problem composed of parts other than the specified summary. Furthermore, a set of the feature of the problem and the solution is extracted from the solution data, and learning result data obtained by learning from the set what kind of solution is likely to become a learning result data Store in the storage means. After that, a text to be summarized is input, a summary candidate is generated from the input text, a set of features is extracted from the input text and the summary candidate, and what kind of solution is likely to be obtained from the set of features is learned. Estimate based on the result data, and generate a pair of summary candidate and estimated solution (summary candidate-estimated solution pair). Then, from the summary candidate-estimated solution pair, a pair whose estimated solution has a predetermined good evaluation and the highest certainty factor is selected, and the summary candidate of the pair is used as a summary.

  As a result, a part extracted by the user from the displayed text can be machine-learned as a summary that the user has evaluated as good, and then the user-specific summary can be performed on the text that has been input thereafter.

  Alternatively, the present invention includes: 1) text storage means for storing text and a summary of the text; 2) summary display processing means for displaying the summary on a display device; and 3) accepting an input of user evaluation for the summary. Evaluation setting processing means for evaluating the summary; and 4) generating solution data to which the evaluation set by the user is given as a solution to the problem composed of the text and the summary and storing the solution data in the solution data storage means And 5) extracting a set of feature features of the problem and a set of solutions or solution candidates from the solution data by using evaluations other than the evaluation that has become the solution among predetermined evaluations as solution candidates. , A feature-solution pair / feature-solution candidate pair extraction processing means having a positive example as a set of feature sets and solutions and a negative example as a set of feature sets and solution candidates, and 6) the extracted sets The teacher signal and Machine learning processing means for storing learning result data in a learning result data storage means for learning what kind of solution or solution candidate and feature is a probability of being a positive example or a probability of being a negative example; 7) summary candidate generation processing means for inputting a text to be summarized and generating a summary candidate from the input text; 8) a set of features from the input text and the summary candidate using the predetermined evaluation as a solution candidate A set of feature candidates and solution candidates, and in the case of the feature set and solution candidate pair, the probability of being positive or negative is estimated based on the learning result data, and the estimated result is estimated A summary candidate-estimated solution pair generation processing means for generating a pair of the summary candidate, the solution candidate pair and the estimated solution as a solution; 9) a solution from the summary candidate / solution candidate pair-estimated solution pair; Is the candidate's good evaluation? The probability of positive cases estimated solutions to select the best pair, and a summary selection processing means to summarize the summary candidates of the pair.

  Alternatively, the present invention includes 1) text storage means for storing text, 2) text display processing means for displaying the text on a display device, and 3) extracting sentence data in a range designated by the user from the text. A summary editing processing means for providing a user-specified summary of the text, and 4) an automatic summary generation process of either a process of generating a summary based on a predetermined rule or a process of generating a summary using a machine learning method Solution data generated by generating a summary of a text and giving a predetermined good evaluation as a solution indicating a good summary selected by the user for a problem composed of the text and the user-specified summary; To solve a problem composed of the text and the summary generated by the automatic summary generation process, which is composed of parts other than the user-specified summary. Solution data output processing means for storing, in a solution data storage means, solution data generated by giving a predetermined bad evaluation indicating that it is not the user-specified summary as a solution; and 5) the solution of the predetermined evaluation. A set of feature features and solutions or solution candidates of the problem is extracted from the solution data with evaluations other than the evaluation as solution candidates, and a set of the feature set and solution is defined as a positive example and the feature set. Feature-solution pair / feature-solution candidate pair extraction processing means having a pair with a solution candidate as a negative example; 6) What kind of solution or set of solution candidates and features is the extracted pair as a teacher signal Machine learning processing means for storing in the learning result data storage means learning result data that has been learned whether the probability is a positive example or a negative example, and 7) a text to be summarized is input and summarization is performed from the input text. Summarization that generates candidates And 8) generating a set of feature sets and solution candidates from the input text and the summary candidates using the predetermined evaluation as a solution candidate, and generating a set of the feature set and solution candidates. In this case, the probability of being a positive example or a negative example is estimated based on the learning result data, and a pair of the summary candidate and the solution candidate and the estimated solution is generated using the estimated result as an estimated solution. The summary candidate-estimated solution pair generation processing means, and 9) the pair of the summary candidate and the solution candidate-estimated solution pair, the solution candidate has a predetermined good evaluation and the probability of the positive example of the estimated solution is the highest. And summary selection processing means for summarizing the pair of summary candidates.

  Alternatively, the present invention includes 1) text storage means for storing text, 2) text display processing means for displaying the text on a display device, and 3) extracting sentence data in a range designated by the user from the text. Summary edit processing means for providing a user-specified summary of the text; and 4) solution data output processing means for generating solution data having the text as a problem and solving the user-specified summary for the problem and storing the solution data in a solution data storage means 5) A set of feature features and the solution of the problem are extracted from the solution data, and learning result data is learned from the set to learn what kind of solution is likely to become a solution. Machine learning processing means for storing in the result data storage means, 6) inputting a text to be summarized, extracting a set of features from the input text, and collecting the feature set Solution estimation processing means for estimating what kind of solution is likely to be based on the learning result data, and 7) summary selection processing means for outputting the solution estimated by the solution estimation processing means as a summary of the input text With.

  As a result, it is possible to perform machine learning of the summarization process using the user's evaluation on the displayed summaries, and to perform user-specific summarization on the input text.

  Alternatively, the present invention includes 1) text storage means for storing text, 2) text display processing means for displaying the text on a display device, and 3) extracting sentence data in a range designated by the user from the text. Summary edit processing means for providing a user-specified summary of the text; and 4) solution data output processing means for generating solution data having the text as a problem and solving the user-specified summary for the problem and storing the solution data in a solution data storage means And 5) generating a summary of the text by an automatic summary generation process of either a process of generating a summary based on a predetermined rule or a process of generating a summary using a machine learning method, and the user of the summaries A solution candidate other than the specified summary is taken as a solution candidate, and a set of the solution or solution candidate and the feature set of the problem is extracted from the solution data. Feature-solution pair / feature-solution candidate pair extraction processing means with a positive example of a set of feature sets and solutions, and a negative example of a set of feature sets and solution candidates, and 6) a teacher of the extracted sets Machine learning processing means for storing learning result data in a learning result data storage means that learns what kind of solution or solution candidate and feature set has a probability of being a positive example or a probability of being a negative example as a signal And 7) summary candidate generation processing means for inputting a text to be summarized and generating a summary candidate from the input text, and 8) a set of features from the input text and the summary candidate with the summary candidate as a solution candidate A set of feature candidates and a solution candidate, and in the case of a set of the feature set and solution candidate, the probability of being positive or negative is estimated based on the learning result data, and the estimated result is As the estimation solution, the summary candidate and the estimation 9) a summary candidate-estimated solution pair generation processing means for generating a pair of and 9) selecting a pair having the highest probability of a positive example of the estimated solution from the summary candidate-estimated solution pair, and summarizing the summary candidates of the pair And summary selection processing means.

  This makes it possible to machine-learn the summarization process as a summary in which the user has evaluated the portion extracted from the displayed text as being good, and to perform summarization specialized for the user on the input text.

  Each means, function, or element of the processing apparatus according to the present invention can be realized by a computer-executable program. This program can be stored in an appropriate recording medium such as a portable medium memory, semiconductor memory, or hard disk, which can be read by a computer, provided by being recorded on these recording media, or via a communication interface. It is provided by transmission / reception using various communication networks.

  According to the present invention, since the user can arbitrarily set the evaluation for the summary result that is the solution data of the machine learning, the automatic summarization process using the computer is not summarized in one type. , It is possible to perform user-specific summaries.

  In the automatic summarization process using the machine learning method, the user can freely edit the solution data as a teacher in order to enable the summarization specific to the user, instead of performing the summarization based on the same evaluation.

  Although it is considered that the evaluation of the summary changes even for the same person, according to the present invention, the evaluation for the summary result can be set at any time even for the same person, and machine learning is performed using new solution data. By re-doing, it is possible to perform summarization in accordance with the new evaluation attitude.

[First Embodiment]
FIG. 1 shows a configuration example of the processing apparatus of the present invention in the first embodiment.

  The automatic summary processing device 10 includes an evaluation customization unit 110, a solution data storage unit 120, a solution-feature pair extraction unit 121, a machine learning unit 122, a learning result data storage unit 123, a summary candidate generation unit 124, A feature extraction unit 125, a summary candidate-estimated solution pair generation unit 126, and a summary selection unit 128 are provided.

  The evaluation customizing unit 110 is a processing unit that realizes solution data editing processing. The solution data storage unit 120, the solution-feature pair extraction unit 121, the machine learning unit 122, and the learning result data storage unit 123 realize the machine learning processing unit of the automatic summarization processing device shown in the claims. Processing means.

  The evaluation customizing means 110 is means for customizing the summary result and its evaluation for each user, and includes a summary display unit 111 and an evaluation assigning unit 112.

  The summary display unit 111 is a means for displaying a summary result of the prepared text / summary 4 on a display device (not shown in FIG. 1).

  Text / Summary 4 consists of a text and its summary result. Text is document data consisting of one or more articles. The summary result is document data that summarizes the text. The summary results are the ones generated manually, the summary 3 output for the text 2 input by the automatic summary processing device 10, or the summary candidates generated by the summary candidate generation unit 124 and stored in the solution data storage unit 120. Also good.

  The evaluation assigning unit 112 is a unit that assigns an evaluation input by the user to the summary result displayed by the summary display unit 111 or changes an evaluation given in advance to the summary result to an evaluation input by the user. .

  The solution data storage unit 120 is a unit that stores solution data to be used as a teacher when the machine learning unit 122 executes the machine learning method. The solution data storage unit 120 stores, as solution data, a case that is a set of a problem including a text and its summary result and a solution that is an evaluation of the summary result.

  The solution-feature pair extraction unit 121 is a unit that extracts a set of a solution and a set of features for each case stored in the solution data storage unit 120.

  A feature means one unit of information used for analysis. Here, 1) information indicating the smoothness of a sentence, 2) information indicating whether the contents are well expressed, and 3) automatic summarization processing. This is characteristic information used.

  The machine learning unit 122 learns, based on a machine learning method, what kind of solution is likely to be generated from the set of the solution extracted by the solution-feature pair extraction unit 121 and the set of features. The learning result is stored in the learning result data storage unit 123. The machine learning unit 122 may perform processing by any method as long as it is a machine learning method using solution data. Examples of methods include a decision tree method, a support vector method, a parameter tuning method, a simple Bayes method, a maximum entropy method, and a decision list method.

  The learning result data storage unit 123 is a unit that stores the learning result data of the machine learning unit 122.

  The summary candidate generating unit 124 is a means for generating summary candidates from the input text 2 based on a predetermined method. The summary candidate generation unit 124 generates summary candidates using various models such as an important sentence selection model, an important part selection model, a model using a deformation rule, and a model using random generation.

  The feature extraction unit 125 is a unit that extracts a set of features for the summary candidates generated by the text 2 and the summary candidate generation unit 124 and passes them to the summary candidate-estimated solution pair generation unit 126.

  The summary candidate-estimated solution pair generation unit 126 refers to the learning result data in the learning result data storage unit 123 and determines what kind of solution is likely to occur in the case of the set of features passed from the feature extraction unit 125. This is a means for generating a pair (summary candidate-estimated solution pair) 127 of a summary candidate and an estimated solution by estimation. The summary candidate-estimated solution pair generation unit 126 further obtains and assigns to each summary candidate-estimated solution pair 127 a certainty factor (probability) that is the estimated solution.

  The summary selection unit 128 is means for receiving the summary candidate-estimated solution pair 127, selecting the summary candidate-estimated solution pair 127 having the highest certainty factor, and setting the summary candidate as the summary 3.

  To describe the evaluation customization process in the first embodiment, consider a case where three users A, B, and C customize the summary result.

  Assume that the user A evaluates with emphasis on the fact that the summary result includes a description about accuracy. It is assumed that user B emphasizes that the description of the method is included in the summary result, and user C evaluates that the description of both the method and the accuracy is included in the summary result. Further, it is assumed that the evaluation of the summary result is divided into three stages, and a classification destination (evaluation) of evaluation 1 = good, evaluation 2 = none, evaluation 3 = bad is given.

  FIG. 2 shows the flow of evaluation customization processing in the first embodiment.

  First, assume that text / summary 4 is prepared. FIG. 3 shows an example of text / summary 4 text, and FIG. 4 shows an example of summary results. Each of FIGS. 4A to 4C shows three summary results r1, r2, and r3.

  The summary display unit 111 displays the summary result extracted from the text / summary 4 on the display screen (step S1). And the evaluation provision part 112 receives the evaluation which the user input, and makes the input evaluation the solution (evaluation) of the displayed summary result (step S2).

  Here, it is assumed that the user A uses the automatic summary processing device 10. The user A assigns an evaluation of 1 to the summary result r1 in FIG. Then, the evaluation provision part 112 receives a user's input (evaluation 1), and sets evaluation 1 as a solution of case c1.

  Next, when the summary display unit 111 displays the summary result r2 of the case c2 shown in FIG. 4B, it is not extracted that the summary result r2 is related to accuracy. Evaluation 3 is assigned to r2, and the evaluation assigning unit 112 sets evaluation 3 as the solution of case c2.

  Furthermore, when the summary display unit 111 displays the summary result r3 of the case c3 shown in FIG. 4C, the summary result r3 is extracted with respect to accuracy, but is slightly redundant. The evaluation 2 is assigned, and the evaluation assigning unit 112 sets the evaluation 2 as the solution of the case c3.

  Similarly, the case of user B is assumed. User B gives evaluation 3 because the place related to the technique is not extracted for the summary result r1 shown in FIG. 4 (A), and is related to the technique for the summary result r2 shown in FIG. 4 (B). However, evaluation 1 is assigned because it has been extracted, and evaluation 2 is assigned because a portion related to the technique is extracted for the summary result r3 shown in FIG.

  Similarly, the case of user C is assumed. The user C attaches the evaluation 2 to the summary result r1 shown in FIG. 4A, but the evaluation 2 is given because the place related to the method is not extracted, and the result shown in FIG. Although the place related to the method is extracted for the summary result r2, the place related to the accuracy is not extracted, so the evaluation 2 is given, and the summary result r3 shown in FIG. Although it is extracted, it is evaluated as 1 because it is slightly redundant.

  The evaluation assigning unit 112 sets the input evaluation for the summary results r1 to r3 for each of the users B and C as solutions (evaluations) of the respective cases c1 to c3.

  Then, the evaluation customizing unit 110 stores the text given in the text / summary 4 and its summary result and solution in the solution data storage unit 120 as an example (step S3).

  FIG. 5 shows the flow of machine learning processing and automatic summarization processing.

  The solution-feature pair extraction unit 121 extracts a set of a solution and a feature set for each case from the solution data storage unit 120 (step S11).

  The solution-feature pair extraction unit 121, for example, 1) as information indicating the smoothness of the sentence, the existence of the k-gram morpheme sequence in the corpus, the degree of semantic consistency between the received phrases, and 2) the contents As the information indicating whether or not it is well expressed, including the key phrase inclusion rate in the text before the summarization, and 3) as the information used in the automatic summarization, whether the sentence is a sentence or the lead sentence, TF / IDF (TF is a value indicating the frequency or frequency of occurrence of the word in the document, IDF is the reciprocal of the number of documents in which the word appears in a large number of document groups in advance), sentence length , Presence of clue expressions such as proper expressions, conjunctions, and function words are extracted as features.

  Next, the machine learning unit 122 learns by using a machine learning method what kind of feature is likely to be obtained from the set of the solution and the feature set, and the learning result is obtained as learning result data. It memorize | stores in the memory | storage part 123 (step S12).

Here, in the case of the process of the user A, “example: problem → solution” of the solution data stored in the solution data storage unit 120 is
Case c1: Text-summary result r1 → evaluation 1,
Case c2: Text-summary result r2 → evaluation 3,
Case c3: Text-summary result r3 → evaluation 2
Thus, the machine learning unit 122 learns by machine learning which case is evaluated 1 to 3 based on these solution data. For example, from the case c1 → evaluation 1 and the case c3 → evaluation 2, the machine learning unit 122 learns that the evaluation increases when an expression of accuracy, for example, the expression “number + [%]” appears. Here, the expression “number + [%]” is an example of a feature used for learning.

In addition, in the case of user B's processing, “example: problem → solution”
Case c1: Text-summary result r1 → evaluation 3,
Case c2: Text-summary result r2 → evaluation 1,
Case c3: Text-summary result r3 → evaluation 2
Thus, the machine learning unit 122 learns that the evaluation becomes high when a technical term corresponding to a technique such as “cue expression” or “example” appears.

In the case of processing by user C, “example: problem → solution”
“Case c1: Text-summary result r1 → evaluation 2,
Case c2: Text-summary result r2 → evaluation 2,
Case c3: text-summary result r3 → evaluation 1 ”
Thus, the machine learning unit 122 learns so that the evaluation becomes high when both the precision expression and the expression corresponding to the technique appear.

  Also, the shorter the sentence output as the summary result, the better. In each processing, learning is performed such that the shorter the sentence length, the higher the evaluation.

  As a machine learning method, for example, a simple Bayes method, a decision list method, a maximum entropy method, a support vector machine method, or the like is used.

  The Simple Bayes method is a method for estimating the probability of each classification based on Bayes' theorem and obtaining the classification having the largest probability value.

  In the decision list method, features and pairs of classification targets are set as rules, and they are stored in a list in a predetermined priority order. When an input to be detected is given, the list starts with the highest priority. This is a method in which the input data is compared with the feature of the rule, and the classification destination of the rule having the same feature is set as the classification destination of the input.

  In the maximum entropy method, when a set of preset features fj (1≤j≤k) is F, a probability distribution when maximizing an expression that means entropy while satisfying a predetermined conditional expression is obtained. This is a method of obtaining a classification having the largest probability value among the probabilities of each classification obtained according to the probability distribution.

  The support vector machine method is a method of classifying data composed of two classifications by dividing a space by a hyperplane.

The decision list method and the maximum entropy method are described in Reference Document 1 below, and the support vector machine method is described in Reference Document 2 and Reference Document 3 below.
[Reference 1: Maki Murata, Masao Uchiyama, Kiyotaka Uchimoto, Ma Aoi, Hitoshi Isahara, Ambiguity Solving Experiments Using Various Machine Learning Methods, IEICE Language Understanding and Communication Study Group, NCL2001-2, ( 2001)]
[Reference 2: Nello Cristianini and John Shawe-Taylor, An Introduction to Support Vector Machines and Other Kernel-based Learning Methods, (Cambridge University Press, 2000)]
[Reference 3: Taku Kudoh, Tinysvm: Support Vector machines, (http://cl.aist-nara.ac.jp/taku-ku//software/TinySVM/index.html,2000)]
Thereafter, when the text 2 for which a summary is to be obtained is input (step S13), the summary candidate generating unit 124 creates a summary candidate from the text 2 using, for example, a processing model as shown below (step S14).

1) Important sentence selection model The important sentence selection model is a model that summarizes sentences in units and realizes the summary by selecting and leaving only sentences that are considered important. In this model, all sentence selection states should be candidates for solutions. In addition, if all of the solutions are candidates for calculation, the calculation speed will be affected. Using a selection rule prepared in advance, only the selected state of a sentence satisfying this selection rule is set as a solution candidate. That is, the processing load is reduced by reducing the number of candidates according to a predetermined selection rule. The selection rule may be a manual rule.

2) Important part selection model The important part selection model is a model that realizes summarization by deleting unnecessary ones that are smaller than sentences and using unnecessary ones as summarization units. The unit is the same as the 1) important sentence selection model except that the unit is smaller than the sentence. For example, a phrase is used as a sentence smaller than the sentence. In other words, summarization is realized by deleting unnecessary phrases in units of phrases. In the case of this important part selection model, the selection states of all phrases are all candidate solutions. In addition, if all of the candidates are the solution, the calculation speed will be affected. As in the case of 1) important sentence selection model, a selection rule is prepared in advance, and only the selected state of the sentence satisfying this selection rule is prepared. Is a candidate for the solution.

3) Model using transformation rules A model using transformation rules is a model that generates summary results using transformation rules prepared in advance. Deformation rules are obtained by automatic processing or used manually. For example, a transformation rule that rewrites “X and Y” to “X” or “Y” is created, and an input “A and B” is given according to this transformation rule. When this is done, summary candidates such as “A done” or “B done” are generated.

4) Model using random generation The model using random generation is, for example, a model that uses "... Y ..." as a summary candidate when there is an input "... X ...". It is. At this time, X to be replaced may be selected at random or may be designated by a replacement rule prepared in advance. As the replacement rule, a manually generated rule or an automatically acquired rule is used. The replaced expression Y may be selected at random from a set of words or character strings in a certain dictionary, or may be designated by a conversion rule prepared in advance. As with the replacement rule, a conversion rule that is manually generated or automatically acquired is used. At this time, if X and Y are not selected at random, but are selected based on the conversion rule, the model is the same as the model using the deformation rule.

  The feature extraction unit 125 extracts a set of features from the input text 2 and the summary candidate by substantially the same processing as the solution-feature pair extraction unit 121, and passes it to the summary candidate-estimated solution pair generation unit 126 (step S15). .

  Then, the summary candidate-estimated solution pair generation unit 126 estimates, based on the learning result data, what kind of solution is likely to occur in the case of the received feature set, that is, each of the plurality of summary candidates. The solution (evaluation) and its certainty are calculated based on the learning result data, and a pair of summary candidate and estimated solution (summary candidate-estimated solution pair) 127 is generated (step S16).

  Then, the summary selection unit 128 selects the summary candidate-estimated solution pair 127 having the highest certainty value of the estimated solution from the generated summary candidate-estimated solution pair 127, and sets the summary candidate as the summary 3 ( Step S17).

  FIG. 6 shows another configuration example of the processing apparatus of the present invention in the first embodiment.

  In this embodiment, a large number of candidates can be considered as solutions (classification destinations), but there may be a case where the number of types of classification destinations becomes too large to be processed by a general machine learning method. In such a case, in the automatic summarization processing apparatus 20 shown in FIG. 6, the machine learning unit 132 uses a machine learning method that considers only two types of solutions (classification destinations) of positive examples and negative examples in actual machine learning processing. By using it, processing becomes possible.

  In the automatic summarization processing apparatus 20 shown in FIG. 6, information called evaluation can be used as the learning feature of the machine learning unit 132.

  The automatic summarization processing device 20 includes an evaluation customizing unit 110, a solution data storage unit 130, a feature-solution pair / feature-solution candidate pair extraction unit 131, a machine learning unit 132, a learning result data storage unit 133, and a summary. A candidate generation unit 134, a feature-solution candidate extraction unit 135, a summary candidate-estimated solution pair generation unit 136, and a summary selection unit 138 are provided.

  The solution data storage unit 130, the feature-solution pair / feature-solution candidate pair extraction unit 131, the machine learning unit 132, and the learning result data storage unit 133 are machine learning of the automatic summarization processing device shown in the claims. It is a processing means for realizing the processing means. The summary candidate generation unit 134, the feature-solution candidate extraction unit 135, and the summary candidate-estimated solution pair generation unit 136 are processing units that implement the summary candidate generation processing unit shown in the claims.

  The evaluation customizing unit 110 and the summary candidate generating unit 134 perform the same processing as the evaluation customizing unit 110 and the summary candidate generating unit 124 of the automatic summary processing apparatus 10 shown in FIG.

  The feature-solution pair / feature-solution candidate pair extraction unit 131 is a means for extracting a set of a solution or a solution candidate and a set of features for each case stored in the solution data storage unit 130. Here, the solution candidate means a solution candidate other than the solution, and the evaluation set by the user is the solution. A pair of a solution and a feature set is a positive example, and a pair of a solution candidate and a feature set is a negative example.

  The machine learning unit 132 learns the probability of being a positive example or the probability of being a negative example from any set of solutions or solution candidates and feature sets, and any solution or solution candidate and feature set, The learning result is stored in the learning result data storage unit 133.

  The feature-solution candidate extraction unit 135 extracts a set of solution candidates and feature sets for the input text and summary candidate by the same processing as the feature-solution pair / feature-solution candidate pair extraction unit 131. Means.

  The summary candidate-estimated solution pair generation unit 136 obtains a probability that is a positive example or a probability that is a negative example in the case of a set of a candidate solution and a set of features that have been passed, and the solution that has the highest probability of being a positive example. Is a means for generating a pair (summary candidate-estimated solution pair) 137 of the summary candidate and the estimated solution in that case.

  The summary selection unit 138 is means for setting the summary candidate of the summary candidate-estimated solution pair 137 as the summary 3.

  FIG. 7 shows a flow of machine learning processing and automatic summarization processing of the automatic summarization processing device 20.

  The feature-solution pair / feature-solution candidate pair extraction unit 131 extracts a set of a solution or a solution candidate and a feature set for each case from the solution data storage unit 130 (step S21). Then, the machine learning unit 132 determines the probability of being a positive example or the probability of being a negative example from any set of solutions or solution candidates and feature sets. Learning is performed by the machine learning method, and the learning result is stored in the learning result data storage unit 133 (step S22).

  Thereafter, when the text 2 for which a summary is desired is input (step S23), the summary candidate generating unit 134 generates a summary candidate from the text 2 by a predetermined method (step S24). Then, the feature-solution candidate extraction unit 135 extracts a set of feature sets and solution candidates from the input text 2 and summary candidates, and passes them to the summary candidate-estimated solution pair generation unit 136 (step S25).

  The summary candidate-estimated solution pair generation unit 136 estimates the probability of a positive example or a negative example based on the learning result data in the case of a set of the received solution candidate and a feature set, and the probability of being a positive example The summary candidate-estimated solution pair 137 is generated using the candidate with the largest solution as the estimated solution (step S26), and the summary selection unit 138 sets the summary candidate of the summary candidate-estimated solution pair 137 as summary 3 (step S27). ).

  In the first embodiment, the user may give evaluations 1 to 3 to the summary results output during use while using the automatic summary processing device 1 when necessary. Therefore, the user can customize the evaluation of the summary result without feeling the operation burden.

[Second Embodiment]
FIG. 8 shows a configuration example of the processing apparatus of the present invention in the second embodiment. The automatic summarization processing apparatus 30 shown in FIG. 8 includes an evaluation customizing means 140 instead of the evaluation customizing means 110 of the automatic summarization processing apparatus 10 shown in FIG. A processing means is provided.

  The evaluation customizing unit 140 includes a text display unit 141 and a summary editing unit 142.

  The text display unit 141 is means for displaying the text 5 prepared in advance on a display device (not shown in FIG. 8).

  The summary editing unit 142 is a means for extracting a part designated by the user as a summary from the text 5 displayed by the text display unit 141 or changing the expression in the part designated by the user to edit the summary.

  FIG. 9 shows the flow of evaluation customization processing in the second embodiment.

  The text display unit 141 takes in the prepared text 5 and displays it on the display device (step S31). On the displayed text 5, the user is allowed to specify a portion that is considered to be a good summary result, and the range specified by the user is received and extracted (step S32). If the specified range portion is edited, the edited content is accepted, and the specified range portion after editing is used as the summary result (step S33).

  The user designates a range to be summarized on the displayed text by dragging with a pointing device such as a mouse or by specifying a start position and an end position by moving a cursor key. The text display unit 141 displays the specified range separately from the unspecified range by display such as inversion or marking.

  FIG. 10 shows an example of text to be displayed and an example of a range designated by the user A. User A designates that the portion surrounded by the broken line “the experiment was conducted on a novel, and the test sample could be analyzed with an accuracy of a reproduction rate of 84% and a precision of 82%” was a good portion to summarize. . The summary editing unit 142 sets the portion indicated by the broken-line rectangle of the text in FIG. 10 as the summary result.

  In addition, as shown in FIG. 11, the user B may enclose a portion surrounded by a broken line of text “a verb is omitted in a natural language. Restoring this omitted verb may be a dialogue system or a high quality Therefore, in this study, the abbreviated verb is complemented from the surface expression (cue word) and examples. " Further, in the case of the user C, as shown in FIG. 12, a part surrounded by two broken rectangles “in natural language, a verb may be omitted. To restore this omitted verb, Therefore, in this study, we will supplement the omitted verbs with surface expressions (cue words) and examples. " Specifying that the experiment should be a summary of the test sample, it was possible to analyze the test sample with an accuracy of 84% recall and accuracy of 82%. The summary editing unit 142 sets the portions indicated by broken-line rectangles of the text shown in FIGS. 11 and 12 as summary results.

  The range specified by the user may be displayed separately from the text so that the user can edit the contents of the expression within the specified range by deleting any part or changing the expression. . As shown in FIG. 13, the range specified on the text is displayed separately from the text so that words, phrases, etc. within the specified range can be deleted, added and corrected. When the summary determination button is selected by an operation such as clicking, the summary editing unit 142 accepts the selection and sets the contents of the designated range as the summary result. If the cancel button is selected, the contents of the specified range are cleared.

  The summary editing unit 142 stores the text 5 and the summary result in the solution data storage unit 130 together with a predetermined solution (good evaluation) (step S34). Further, in the first embodiment, the evaluation customizing means 140 is a summary generated by the automatic summary processing device 20, a summary candidate generated by the summary candidate generating unit 124 of the automatic summary processing device 20, or a manually generated summary. Also stored in the solution data storage unit 130 is solution data in which a predetermined solution (bad evaluation) is given to a summary other than the summary designated by the user.

  The flow of machine learning processing and automatic summarization processing is the same as the processing flow shown in FIG. Here, the machine learning unit 122 learns, for each user, one of the text shown in FIG. 3, the summary result shown in FIGS. 10 to 12 (that is, a user-specified range), and a case as a solution. .

  FIG. 14 shows another configuration example of the processing apparatus of the present invention in the second embodiment. Even in this embodiment, there may be a case where the number of types of solutions (classification destinations) becomes too large to be processed by a general machine learning method.

  For this reason, in the automatic summarization processing apparatus 40 shown in FIG. 14, the machine learning unit 132 uses a machine learning method that considers only two types of solutions (classification destinations) of positive examples and negative examples in the actual machine learning process. Processing is possible.

  The automatic summarization processing apparatus 40 includes processing means similar to the processing means constituting the automatic summarization processing apparatus 20 shown in FIG. 6, and includes an evaluation customizing means 140 instead of the evaluation customizing means 110.

  In the present embodiment, since the user can designate a range that can be used as a summary on the text, the burden on the user is greater than that in the first embodiment. However, since it is possible to use solution data (teacher) that is closer to the summary result desired by the user, it is possible to learn so that the summary result desired by the user can be output more quickly.

[Third Embodiment]
FIG. 15 shows a configuration example of the processing apparatus of the present invention in the third embodiment. An automatic summary processing device 50 shown in FIG. 15 includes an evaluation customizing unit 150 instead of the evaluation customizing unit 110 of the automatic summary processing device 10 shown in FIG. 1, and constitutes the automatic summary processing device 10 as another processing unit. Processing means similar to the processing means are provided.

  The evaluation customizing unit 150 includes a summary display unit 151 and a property information setting unit 152.

  The summary display unit 151 is a means for displaying a prepared text / summary 4 summary result on a display device (not shown in FIG. 15).

  The property information setting unit 152 is a means for generating a plurality of property information related to the evaluation of the summary result and setting the evaluation for each property information.

  The property information is information on various properties that constitute the evaluation of the summary result. For example, information indicating whether a short sentence is emphasized (short sentence emphasis), and the summary result includes an expression about quantity. On whether or not emphasis is placed on (quantity expression emphasis), information on whether or not the summary results include expression about the technique (method emphasis), and whether or not the style of the summary results is emphasized Information (stylistic emphasis), information about whether or not the summary results are easy to read (readability emphasis).

  The evaluation customizing unit 150 can arbitrarily set a plurality of property information related to the evaluation that the user needs at any time so that the machine learning unit 122 can learn a plurality of properties related to the evaluation of the summary result. Thus, the user's evaluation for the summary result is defined using a plurality of property information.

  In this embodiment, the solution data storage unit 120 is prepared corresponding to the number of property information set by the property information setting unit 152, and machine learning is performed for each property information. Accordingly, the processing means of the solution-feature pair extraction unit 121, the machine learning unit 122, the learning result data storage unit 123, and the summary candidate-estimated solution pair generation unit 126 are provided corresponding to the number of property information.

  FIG. 16 shows the flow of evaluation customization processing.

  The summary display unit 151 displays the summary result extracted from the text / summary 4 (step S41). The property information setting unit 152 displays a plurality of items of property information for the displayed summary result, prompts the user to set the value of each item, setting a new item, etc., and accepts user input (step S42).

  17 and 18 show examples of property information setting screens. In the property information setting screen, a slide bar is provided corresponding to each of the plurality of position information. The user can specify an evaluation for each property information by setting an arbitrary position such as the right or left side of the slide button on the slide bar of each property information. For example, the user evaluates the displayed summary result for each item of property information such as “emphasis on short sentences, importance on quantity expression, importance on methods, importance on style, readability”. , Set by moving the slide button on the slide bar. In FIGS. 17 and 18, it is assumed that the evaluation is set so that the evaluation increases from the left end to the right end of the slide bar. Moreover, the user can freely define what the slide bar means by inputting arbitrary property information beside the slide bar.

  When the summary display unit 151 displays the summary result r1 shown in FIG. 4 (A), as shown in FIG. 17, since the summary result r1 is a short sentence, the user A moves the slide button “emphasis on short sentences” to the right side. Also, because there is an expression related to quantity, the slide button for “quantity expression emphasis” is on the right side, and since the method is not touched, the slide button for “method emphasis” is on the left side, and the style and readability are not so bad. Position the slide buttons for "style style" and "readability" to the right.

  Further, when the summary display unit 151 displays the summary result r2 shown in FIG. 4B, the summary of the summary result r2 is not so short as shown in FIG. Move the button to the left and the other property information is OK, so move the slide button for the other property information to the right.

  Then, the property information setting unit 152 sets the value for each input property information as a solution, and stores the solution, text, and summary result as examples in the solution data storage unit 120 for each property information (step S43).

  Thereafter, the flow of the machine learning process and the automatic summarization process is almost the same as the process flow shown in FIG. Here, the machine learning unit 122 provided for each property information uses the case stored in the solution data storage unit 120 of the corresponding property information as solution data (teacher data). The machine learning unit 122 performs learning for each property information. For example, for the property information “emphasis on short sentences”, the solution of each case is “solution = rightmost” in case c1 (summary result r1) and “solution = left side” in case c2 (summary result r2). it can. The machine learning unit 122 uses these solution data as teacher data to learn when to evaluate with a short sentence emphasis. The same learning is performed for other property information.

  In this embodiment, after the machine learning process, the summary candidate generation unit 124 generates a summary candidate from the input text 2 by a predetermined method, and the feature extraction unit 125 extracts a set of features from the input text 2 and the summary candidate. To do.

  Then, the summary candidate-estimated solution pair generation unit 126 corresponding to each property information estimates, based on the learning result data, what kind of solution is likely to be obtained in the case of the received feature set, and estimates the estimation candidates. A pair with a solution (summary candidate-estimated solution pair) 127 is generated. For example, the summary candidate-estimated solution pair generation unit 126 calculates the estimated solution and the certainty factor of each of the plurality of summary candidates based on the learning result data, and obtains the summary candidate-estimated solution pair 127 for each property information. Generate.

  The summary selection unit 128 receives the user evaluation setting information 7 that sets how much importance is given to the evaluation property information for the summary result, and evaluates the evaluation value of each property information in the summary candidate-estimated solution pair 127 by the user evaluation. Compared with the setting information 7, the most similar summary candidate-estimated solution pair or the most suitable summary candidate-estimated solution pair for the user evaluation setting information 7 is selected, and the summary candidate of the summary candidate-estimated solution pair 127 is selected. Summary 3

  The summary selection unit 128 displays a property information setting screen as shown in FIG. 17, and the user evaluation setting information 7 which is the property of the summary result currently required by the user is displayed on the slide bar of each item of property information. The button position may be changed and set.

  For example, when the slide buttons for "Short sentence emphasis", "Quantity expression emphasis", and "Method emphasis" are moved to the right side, and the "Text style emphasis" and "Readability emphasis" slide buttons are moved to the left side. Means that the user is requesting a summary 3 that conforms to the nature of the evaluation such that the user is as short as possible and the quantitative expression and method are essential, but the style and readability are neglected Setting information 7 is obtained.

  Moreover, the summary selection part 128 may obtain | require the value Total_Score of all the solution combinations as a simple selection method of the summary candidate-estimated solution pair 127, for example using the following formula | equation.

Total_Score = a (emphasis on short sentences) x score (emphasis on short sentences)
+ A (emphasis on quantity expression) x score (emphasis on quantity expression)
+ A (method emphasis) × score (method emphasis)
+ A (style style emphasis) × score (style style emphasis)
+ A (easy to read) x score (easy to read)
However, a (X) is a value obtained from the slide button position of the slide bar of the property information X designated by the user. The slide button has a larger value as it is located on the right side of the slide bar. score (X) is an evaluation value of the property information X calculated based on the learning result data. The summary selection unit 128 selects the summary candidate-estimated solution pair 127 having the largest combination value Total_Score, and outputs the summary candidate as the summary 3.

  In this embodiment, a processing method similar to the processing in the first embodiment is adopted, in which summary results used by the machine learning unit 122 are displayed and evaluated by the user. However, in this embodiment, as in the machine learning method in the second embodiment, the user is allowed to specify a range that is considered to be a summary result from the text, and then the user has a plurality of properties related to the evaluation. May be evaluated using a slide bar or the like, and solution data that is a teacher signal may be collected. In the case of such processing, even the same user can cope with a situation in which a desired summary type differs every time processing is performed. In addition, since it is possible to learn a plurality of property information at the same time, it is possible to reduce the processing load when the user gives an evaluation (solution) as a whole.

  In this embodiment, an automatic summarization processing device 60 having the processing means as shown in FIG. 19 may be used. An automatic summarization processing device 60 in FIG. 19 includes processing means similar to the processing means constituting the automatic summarization processing device 20 shown in FIG. 6, and includes an evaluation customizing means 150 instead of the evaluation customizing means 110.

  The automatic summarization processing device 60 avoids an excessive processing burden in machine learning by using a machine learning method that considers only two types of solutions (classification destinations) of positive examples and negative examples in actual machine learning processes. be able to.

  As mentioned above, although this invention was demonstrated by the embodiment, this invention can be variously deformed in the range of the main point. For example, it is possible to combine any of the first to third embodiments.

[Fourth Embodiment]
FIG. 20 shows a configuration example of the processing of the present invention in the fourth embodiment. 20 includes a solution data storage unit 120, a solution-feature pair extraction unit 121, a machine learning unit 122, a learning result data storage unit 123, a feature extraction unit 125, and a solution estimation unit. 160 and evaluation customizing means 140.

  The solution data storage unit 120, the solution-feature pair extraction unit 121, the machine learning unit 122, the learning result data storage unit 123, the feature extraction unit 125, and the evaluation customization unit 140 of the automatic summary processing device 70 have the same numbers as shown in FIG. Is a means for performing substantially the same processing as the processing means to which is given.

  The solution estimation unit 160 refers to the learning result data in the learning result data storage unit 123 to estimate what kind of solution is likely to occur in the case of a set of features passed from the feature extraction unit 125, and the estimation The solution 161 is a means for summarizing 3.

  In this embodiment, the solution data storage unit 120 stores solution data having a text as a problem and a text summary result as a solution, and the machine learning unit 122 uses a solution-feature pair extracted from the solution data. Do learning. The feature extraction unit 125 extracts the feature of the input text 2 and passes it to the solution estimation unit 160.

  FIG. 21 shows the flow of machine learning processing and automatic summarization processing in the fourth embodiment.

  The solution-feature pair extraction unit 121 extracts a set of a solution and a feature set for each case from the solution data storage unit 120 (step S51). Next, the machine learning unit 122 extracts the solution and feature set. From the set of features, what kind of feature set is likely to become a solution is learned by the machine learning method, and the learning result is stored in the learning result data storage unit 123 (step S52). The processes in steps S51 and S52 are the same as the processes in steps S11 and S12 shown in FIG.

  Thereafter, when the text 2 to be summarized is input (step S53), the feature extraction unit 125 extracts a set of features from the input text 2 by substantially the same processing as the solution-feature pair extraction unit 121, It passes to the solution estimation part 160 (step S54). Then, the solution estimation unit 160 estimates, based on the learning result data, what kind of solution is likely to be generated in the case of the received feature set, and sets the estimated solution 161 as summary 3 (step S55).

  In the present embodiment, machine learning is performed using solution data whose solution is a text summary result, and an estimated solution to be summarized is directly obtained in solution estimation processing referring to the learning result.

  Although the present invention has been described above with reference to the embodiments, it is obvious that the present invention can be variously modified within the scope of the gist thereof.

  Although the present invention has been described as being implemented as a processing program that is read and executed by a computer, the processing program that implements the present invention includes a portable medium memory, a semiconductor memory, a hard disk, and the like that can be read by a computer. It can be stored in an appropriate recording medium, provided by being recorded on these recording media, or provided by transmission / reception using various communication networks via a communication interface.

It is a figure which shows the structural example of the processing apparatus of this invention in 1st Embodiment. It is a figure which shows the flow of the evaluation customization process in 1st Embodiment. It is a figure which shows the example of the text used as object. It is a figure which shows the example of a summary result. It is a figure which shows the flow of the machine learning process and automatic summarization process in the processing apparatus shown in FIG. It is a figure which shows another structural example of the processing apparatus of this invention in 1st Embodiment. It is a figure which shows the flow of the machine learning process and automatic summarization process in the processing apparatus shown in FIG. It is a figure which shows the structural example of the processing apparatus of this invention in 2nd Embodiment. It is a figure which shows the flow of the evaluation customization process in 2nd Embodiment. It is a figure which shows the example of the text displayed, and the example of the designated range of the user A. It is a figure which shows the example of the text displayed, and the example of the user's B designated range. It is a figure which shows the example of the text displayed, and the example of the designation | designated range of the user C. It is a figure which shows the example of a display of the range designated by the user. It is a figure which shows another structural example of the processing apparatus of this invention in 2nd Embodiment. It is a figure which shows the structural example of the processing apparatus of this invention in 3rd Embodiment. It is a figure which shows the flow of the evaluation customization process in 3rd Embodiment. It is a figure which shows the example of a property information setting screen. It is a figure which shows the example of a property information setting screen. It is a figure which shows another structural example of the processing apparatus of this invention in 3rd Embodiment. It is a figure which shows the structural example of the processing apparatus of this invention in 4th Embodiment. It is a figure which shows the flow of the machine learning process and automatic summarization process in the processing apparatus shown in FIG.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Automatic summary processing apparatus 110 Evaluation customization means 111 Summary display part 112 Evaluation provision part 120 Solution data storage part 121 Solution-feature pair extraction part 122 Machine learning part 123 Learning result data storage part 124 Summary candidate generation part 125 Feature extraction part 126 Summary Candidate-estimated solution pair generation unit 127 Summary candidate-estimated solution pair 128 Summary selection unit 130 Solution data storage unit 131 Feature-solution pair / feature-solution candidate pair extraction unit 132 Machine learning unit 133 Learning result data storage unit 134 Summary candidate generation Unit 135 feature-solution candidate extraction unit 136 summary candidate-estimated solution pair generation unit 137 summary candidate-estimated solution pair 138 summary selection unit 140 evaluation customization unit 141 text display unit 142 summary editing unit 150 evaluation customization unit 151 summary display unit 152 property Information setting unit 160 Solution estimation unit 161 Teikai 2 text 3 summary 4 text summary 5 text 7 user evaluation setting information

Claims (6)

A solution data editing processing device for editing solution data used in a process of automatically summarizing text as document data by a machine learning method,
Text storage means for storing text as document data;
Summary display means for displaying text acquired from the text storage means on a display device, extracting sentence data in a range specified by a user from the text, and displaying the text as a user-specified summary of the text;
Information indicating the characteristics of the summary used as an evaluation of the summary, with emphasis on the fact that the summary includes a short sentence-oriented nature indicating whether a short sentence is emphasized as an abstract, and an expression about quantity Quantitative expression emphasis on whether or not there is an emphasis, method emphasis on whether or not the summary includes expression about the technique, emphasis on the style of the summary Or an item for inputting an evaluation value for each of a plurality of properties including two or more of the properties of emphasizing readability indicating that importance is attached to whether the summary is easy to read. Evaluation giving means for accepting input of an evaluation value of each of the properties for the user-specified summary;
Solution data storage means for storing solution data composed of problems and solutions;
The text and the user-specified summary are used as a problem, an evaluation value input by the user is assigned as a solution to the problem, solution data is generated, a sentence is extracted from the text, and a selection state of any sentence is set as a summary candidate. An important sentence selection process, an important part selection process in which a clause is extracted from the text and the selection state of every phrase is a summary candidate, or a sentence of the text is transformed according to a predetermined transformation rule and the transformed state is taken as a summary candidate Any one of the transformation processes is performed to generate a summary candidate for the text, and the summary candidate consisting of the text and the summary candidate other than the user-specified summary is used as a problem. Is generated as a solution by giving a bad evaluation indicating that is not the user-specified summary, and is input by the user. Solution data edit processing device characterized by comprising an evaluation customization means for outputting to the solution data storage means the solution data and solution data to solution the evil evaluate the solution an evaluation value. The solution data editing apparatus according to claim 1,
The evaluation customization means includes summary editing means for accepting an input of a change of a phrase of the text portion specified by the user and using the changed portion as the user-specified summary. Processing equipment. The solution data editing apparatus according to claim 1,
The evaluation customizing means displays an item for inputting an evaluation value of each of the properties with respect to a summary candidate that is a summary candidate generated by the summary generation process and includes a portion other than the user-specified summary, and each user of the item A solution data editing process, wherein the solution data is generated by accepting an input of the evaluation value of the image and giving a combination of the input evaluation value to the problem that is the text and the displayed summary candidate as a solution apparatus. Solution data editing for editing solution data used by a computer having text storage means, summary display means, evaluation assigning means, evaluation customization means, and solution data storage means for automatically summarizing text as document data by machine learning method A processing method,
The summary display means accesses the text storage means for storing text as document data, acquires the text, displays the text on a display device, and extracts sentence data in a range specified by the user from the text Process to display as a user-specified summary of the text;
Information indicating the characteristics of the summary used as an evaluation of the summary by the evaluation assigning means, and the summary includes a short sentence-oriented nature indicating whether a short sentence is emphasized or an expression about the quantity. Whether the emphasis is on quantity expression that indicates whether or not emphasis is placed on it, whether the emphasis is on the fact that the summaries include expression about the technique, whether or not emphasis is placed on the style of the summary The evaluation value for each of the two or more properties, including two or more of the stylistic properties that indicate or whether the summary is easy to read is important. A process of displaying an item to be input and receiving an input of an evaluation value of each of the properties for the user-specified summary;
The evaluation customizing means sets the text and the user-specified summary as a problem, assigns an evaluation value input by the user to the problem as a solution, generates solution data, extracts a sentence from the text, selects any sentence key sentence selection processing of the status and summary candidate was deformed the deformed according modification rule statement was predetermining the important passage selection process, or the text and selecting the state candidate condensates of any clauses removed clause from the text Performing any one of transformation processes with the state as a summary candidate to generate a summary candidate for the text, and considering the summary candidate consisting of the text and the summary candidate other than the user-specified summary, The solution data is generated by giving a bad evaluation indicating that the summary candidate is not the user-specified summary to the problem. , Solution data editing processing method characterized in that it comprises a process of outputting the solution data and solution data to solution the evil evaluate the solution an evaluation value input by the user on the solution data storage means. In the solution data editing processing method according to claim 4,
The computer comprises summary editing means,
In the processing process executed by the evaluation customizing means, the summary editing means accepts input of a change in the phrase of the text portion designated by the user, and sets the changed portion as the user designated summary. Solution data editing processing method characterized by performing. In the solution data editing processing method according to claim 4,
In the process performed by the evaluation customizing means, items for inputting evaluation values of the respective properties for the summary candidates generated by the summary generation process and consisting of portions other than the user-specified summary are displayed. A process of receiving input of evaluation values of users of each of the items, and generating the solution data by adding a combination of the input evaluation values as a solution to the problem that is the text and the displayed summary candidate Solution data editing processing method characterized by the above.

Images