WO2022113314A1 - Learning method, learning program, and learning device - Google Patents

Learning method, learning program, and learning device Download PDF

Info

Publication number
WO2022113314A1
WO2022113314A1 PCT/JP2020/044396 JP2020044396W WO2022113314A1 WO 2022113314 A1 WO2022113314 A1 WO 2022113314A1 JP 2020044396 W JP2020044396 W JP 2020044396W WO 2022113314 A1 WO2022113314 A1 WO 2022113314A1
Authority
WO
WIPO (PCT)
Prior art keywords
sentence
input
utterance
evaluation
learning
Prior art date
Application number
PCT/JP2020/044396
Other languages
French (fr)
Japanese (ja)
Inventor
航 光田
竜一郎 東中
哲也 杵渕
Original Assignee
日本電信電話株式会社
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 日本電信電話株式会社 filed Critical 日本電信電話株式会社
Priority to PCT/JP2020/044396 priority Critical patent/WO2022113314A1/en
Publication of WO2022113314A1 publication Critical patent/WO2022113314A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/40—Processing or translation of natural language
    • G06F40/55—Rule-based translation
    • G06F40/56—Natural language generation

Definitions

  • the present invention relates to a learning method, a learning program and a learning device.
  • This method targets the problem of generating a supportive utterance for an input utterance (an utterance that states a specific reason for supporting the content of the input utterance).
  • information similar to keywords and categories is acquired as learning knowledge from external knowledge such as SNS and Internet encyclopedias by using information on keywords and categories that frequently appear in input utterances and supportive utterances. ..
  • the method of applying the utterance generation model using deep learning described above is a technique for generating supportive utterances, and is difficult to apply to the generation of counter-argument utterances.
  • For the support utterance it is sufficient to generate a sentence having a content similar to the input utterance, in other words, a sentence having a meaning similar to the input utterance, and expressions such as keywords and categories that frequently appear in the input utterance and the support utterance (for example, "... is fun". Information similar to ”)) may be acquired as prior knowledge. Semantic similarity can be dealt with in a general deep learning-based utterance generation model as used in the method of applying the utterance generation model using deep learning described above.
  • counter-argument utterances are required to generate sentences with different meanings, more specifically sentences with opposite positions, rather than similar content to input utterances. Therefore, in order to handle counter-argument utterances, it is necessary to explicitly handle whether or not the positions of input and output are reversed.
  • the present invention has been made in view of the above, and an object of the present invention is to improve the processing performance of discussion dialogue in a dialogue system.
  • the evaluation expression extraction process collects a plurality of sentences and extracts the evaluation sentences having the specified subject of interest and including the evaluation expression.
  • the pair creation step generates a sentence pair of an input sentence and an output sentence by combining the evaluation sentences based on the polarity of each evaluation of the evaluation sentences extracted in the evaluation expression extraction step.
  • the learning step the generation model is learned based on the sentence pair generated in the sentence pair creation step, and the polarities of the evaluations of the input sentence and the output sentence.
  • the utterance generation step acquires the input data generated in the input step, and the generation model generated in the learning step based on the utterance input sentence and the polarities of the utterance input sentence and the response sentence. Is used to generate and output the response statement.
  • FIG. 1 is a block diagram of an utterance generator.
  • FIG. 2 is a diagram showing an outline of the flow of utterance generation.
  • FIG. 3 is a diagram showing a specific example of a sentence pair created by the pair creation unit.
  • FIG. 4 is a diagram showing an example of learning data.
  • FIG. 5 is a diagram showing an example of a counter-argument utterance generated by the utterance generator.
  • FIG. 6 is a flowchart of the learning process by the utterance generator.
  • FIG. 7 is a flowchart of the utterance generation process by the utterance generation device.
  • FIG. 8 is a diagram showing an example of a computer that executes a learning program.
  • FIG. 1 is a block diagram of an utterance generator. Further, FIG. 2 is a diagram showing an outline of the flow of utterance generation. The configuration of the utterance generation device 1 will be described with reference to FIGS. 1 and 2.
  • the utterance generation device 1 is a learning device such as a server.
  • the utterance generation device 1 is a device that generates and outputs a counter-argument utterance with respect to the input amount. As shown in FIG. 1, the utterance generation device 1 includes an evaluation expression extraction unit 11, a pair creation unit 12, a learning unit 13, an input unit 14, an utterance generation unit 15, and an output unit 16.
  • the evaluation expression extraction unit 11 collects Web texts existing on the Web, as shown in step S1 of FIG. Further, the evaluation expression extraction unit 11 receives the input of the subject of interest, which is the subject of the sentence to be evaluated.
  • the evaluation expression extraction unit 11 performs morphological analysis on the collected Web text.
  • the evaluation expression extraction unit 11 performs focal word extraction for extracting keywords representing the topic of the Web text by using the sentence analyzed by the morphological element. Further, the evaluation expression extraction unit 11 extracts the proper noun using the sentence analyzed by the morphological element. Further, the evaluation expression extraction unit 11 extracts evaluation expressions for extracting evaluation information such as likes, dislikes, and conveniences by using the sentences analyzed by morphological analysis. Further, the evaluation expression extraction unit 11 performs modality extraction for extracting the presence / absence of a negative expression or the like by using the sentence analyzed by the morphological element.
  • the evaluation expression extraction unit 11 uses the sentences analyzed by morphological analysis to estimate the dialogue action in which the dialogue is estimated, such as whether each sentence is a question or a detailed sentence. Since the above-mentioned evaluation expression extraction is a general language processing task, there is no particular limitation on the language analyzer used to realize the above-mentioned evaluation expression extraction.
  • the evaluation expression extraction unit 11 extracts a sentence including an evaluation expression having a subject of interest after performing each of the above-mentioned analyzes. For example, when the subject of interest is "X", the evaluation expression extraction unit 11 extracts an evaluation sentence that starts with "X is” and includes an evaluation for X.
  • the evaluation expression extraction unit 11 uses the evaluation information, the result of modality extraction, and the like to describe an evaluation sentence including an evaluation expression having the extracted subject of interest as a sentence having a positive evaluation polarity and a negative evaluation sentence with respect to the subject of interest. Divide into polar sentences.
  • the polarity of a sentence's evaluation is information that indicates whether to take a supportive position or an opposite position to the subject of the sentence. Called negative polarity. In the following, the polarity of sentence evaluation is simply referred to as polarity.
  • the evaluation expression extraction unit 11 classifies and registers the evaluation sentence including the extracted subject of interest into a sentence having a positive polarity with respect to the subject of interest and a sentence having a negative polarity, and generates an evaluation classification list.
  • the evaluation expression extraction unit 11 classifies sentences according to the polarity with respect to the subject of interest as described above.
  • the evaluation expression extraction unit 11 classifies a sentence having a positive polarity starting with "X is” and a sentence having a negative polarity starting with "X is”.
  • Sent (X, +) n indicates that X is a sentence starting with "X is”
  • + indicates that it is a positive sentence
  • n is assigned a serial number to each sentence. Represents an identification number.
  • Sent (X, +) n ⁇ indicates that it is a negative sentence, and the other symbols are the same as in the case of Sent (X, +) n .
  • the evaluation expression extraction unit 11 can appropriately extract a sentence whose subject of discussion is the subject of interest by acquiring an evaluation sentence starting with the subject of interest as a language pattern.
  • the evaluation expression extraction unit 11 outputs the generated evaluation classification list to the pair creation unit 12.
  • the pair creation unit 12 receives the input of the evaluation classification list from the evaluation expression extraction unit 11. Next, the pair creation unit 12 uses the acquired evaluation classification list to combine evaluation sentences having the same polarity as support sentence pairs to generate a sentence pair in which one is an input side sentence and the other is an output side sentence. do. That is, the pair creation unit 12 combines a positive polarity evaluation sentence and a positive polarity evaluation sentence to form a sentence pair, and further combines a negative polarity evaluation sentence and a negative polarity evaluation sentence to form a sentence pair. And.
  • This sentence pair is a combination of mutually supportive sentences.
  • the statement on the input side may be simply referred to as the input side
  • the statement on the output side may be simply referred to as the output side.
  • the pair creation unit 12 has Sent (X, +) 1 as the input side and Sent (X, +) 2 as the output side, and is a supportive sentence pair. Further, the pair creation unit 12 has Sent (X, ⁇ ) 1 as the input side and Sent (X, ⁇ ) 2 as the output side, and is a supportive sentence pair.
  • the pair creation unit 12 uses the acquired evaluation classification list to combine evaluation sentences of different polarities as a counter-argument sentence pair to generate a sentence pair with one as the input side and the other as the output side. That is, the pair creation unit 12 uses the input side as an evaluation sentence having a positive polarity and combines them as an evaluation sentence having a negative polarity on the output side to form a sentence pair. Further, the pair creation unit 12 combines the input side as an evaluation sentence having a negative polarity and the output side as an evaluation sentence having a positive polarity to form a sentence pair. This sentence pair is a combination of sentences that are mutually counter-arguing.
  • the pair creation unit 12 sets the input side to Sent (X, +) 1 and the output side to Sent (X,-) 1 to form a sentence pair of counterarguments. .. Further, in the pair creation unit 12, the input side is Sent (X, +) 2 and the output side is Sent (X, ⁇ ) 2 , and the sentence pair is a counter-argument sentence.
  • FIG. 3 is a diagram showing a specific example of a sentence pair created by the pair creation unit.
  • FIG. 3 is an example of a sentence pair of an anti-thesis when the subject of interest is “ramen”.
  • the polarity of the data on the input side is positive, and the polarity on the output side is negative.
  • the pair creation unit 12 can generate each sentence pair shown in FIG. 3 when ramen is the subject of interest. As shown in FIG. 3, by using the polarity and the language pattern, the pair creation unit 12 can appropriately acquire a sentence pair that seems to be a counterargument.
  • the learning unit 13 acquires the sentence pair generated by the pair creation unit 12. Then, the learning unit 13 learns the utterance generation model using the acquired sentence pair. The learning unit 13 inserts the polarity on the input side and the polarity on the output side as tokens at the end of the input sentence as a polarity guide indicating whether each sentence has a positive polarity or a negative polarity. Then, input / output learning is performed using an input statement in which the polarity on the input side and the polarity on the output side are inserted as tokens. For example, FIG. 4 is a diagram showing an example of learning data. FIG.
  • FIG. 4 shows a pair of counter-argument sentences in which the subject of interest is ramen, in which the sentence on the input side has a positive polarity and the sentence on the output side has a negative polarity.
  • the arrow in FIG. 4 represents the conversion from the input sentence to the output sentence.
  • the positive and negative signs before the arrow indicate the polarity on the input side and the polarity on the output side.
  • the learning unit 13 learns using, for example, a generative model using a method that uses BERT (Bidirectional Encoder Representations from Transformers), which is a pre-learning method, for encoding and decoding. As a result, as shown in step S4 of FIG. 2, the learning unit 13 generates a trained generative model.
  • BERT Bidirectional Encoder Representations from Transformers
  • the method used for learning is not limited to BERT, and the learning unit 13 may use another utterance generation model learning algorithm.
  • the learning unit 13 performs fine tuning using the data of the support utterance and the data of the counter-argument utterance collected manually in addition to the learning by BERT to improve the accuracy of the generated model.
  • the learning unit 13 does not have to perform fine tuning, but it is possible to generate more accurate supportive utterances and counter-argument utterances by performing fine tuning.
  • the learning unit 13 outputs the trained generation model to the utterance generation unit 15.
  • the input unit 14 receives the input of the spoken sentence to be the target of the dialogue.
  • the sentence generated by voice recognition of the words spoken by a person may be acquired, or the sentence corresponding to the utterance may be manually input by the user.
  • this input sentence is referred to as an utterance input sentence.
  • the input unit 14 receives an instruction from the position of response such as whether to output a support sentence or request a counterargument sentence as a response sentence to the utterance input sentence.
  • the input unit 14 may have acquired the instruction of the position of this response in advance, or may receive the input together with the utterance input sentence.
  • the input unit 14 determines the polarity of the utterance input sentence with respect to the subject. For example, the input unit 14 estimates the polarity of the utterance input sentence with respect to the subject by the same evaluation polarity determination process as the evaluation expression extraction unit 11. Further, the input unit 14 determines the polarity on the output side as the same polarity as the utterance input sentence if the response position is designated as support, and as the opposite polarity if the response position is specified as counterargument. Then, as shown in step S5 of FIG. 2, the input unit 14 inserts the polarity on the input side and the polarity on the output side as a token at the end of the utterance input sentence as a polarity guide to generate input data. Then, the input unit 14 outputs the input data in which the token is inserted at the end of the input sentence to the utterance generation unit 15 having the trained generation model.
  • the utterance generation unit 15 acquires the trained generation model from the learning unit 13. After that, the utterance generation unit 15 receives the input of the input data in which the token is inserted at the end of the utterance input sentence from the input unit 14. Then, as shown in step S6 of FIG. 2, the utterance generation unit 15 generates a supportive utterance or a counter-argument utterance for the utterance input sentence input using the learning model acquired according to the token. If the response position is supportive, the utterance generation unit 15 generates a counter-utterance to the utterance input sentence.
  • the utterance generation unit 15 generates a sentence positive for the subject if the utterance input sentence is a positive sentence for the subject, and the utterance input sentence is a negative sentence for the subject. For example, it produces a negative sentence for the subject. If the response position is a counterargument, the utterance generation unit 15 generates a counterargument utterance to the utterance input sentence. More specifically, the utterance generation unit 15 generates a negative sentence for the subject if the utterance input sentence is a positive sentence for the subject, and the utterance generation unit 15 may be a negative sentence for the subject. For example, it produces a positive sentence for the subject. After that, the utterance generation unit 15 outputs a response sentence to the generated utterance input sentence to the output unit 16.
  • the output unit 16 acquires a response sentence, which is a supportive utterance or a counter-argument utterance to the utterance input sentence, from the utterance generation unit 15. Then, the output unit 16 outputs the acquired response statement as shown in step S7 of FIG.
  • FIG. 5 is a diagram showing an example of a counter-argument utterance generated by the utterance generator.
  • the utterance generation device 1 when a sentence that baseball is fun is input, the utterance generation device 1 generates and outputs a counter-argument that the rules are difficult. In this case, the utterance generation device 1 outputs a negative sentence for baseball as a counter-utterance to a positive sentence for baseball. Further, the utterance generator 1 generates and outputs a counter-argument that the economy is good when a sentence that the stock is not profitable is input. In this case, the utterance generator 1 outputs a positive sentence for the stock as a counter-utterance to the negative sentence for the stock.
  • the utterance generation device 1 generates and outputs a counter-argument that practice is required when a sentence that golf is fun is input.
  • the utterance generator 1 outputs a negative sentence for golf as a counter-utterance to a positive sentence for golf. It can be said that each counter-argument utterance is an appropriate counter-argument to the input sentence.
  • FIG. 6 is a flowchart of the learning process by the utterance generator.
  • the evaluation expression extraction unit 11 collects Web texts existing on the Web (step S11). Further, the evaluation expression extraction unit 11 receives the input of the subject of interest.
  • the evaluation expression extraction unit 11 extracts a sentence having a designated subject of interest and including an evaluation expression from the collected Web text (step S12).
  • the evaluation expression extraction unit 11 classifies the extracted sentences into positive sentences and negative sentences with respect to the subject of interest, and creates an evaluation classification list (step S13). After that, the evaluation expression extraction unit 11 outputs the generated evaluation classification list to the pair creation unit 12.
  • the pair creation unit 12 receives the input of the evaluation classification list from the evaluation expression extraction unit 11. Next, the pair creation unit 12 creates an instruction sentence pair and a counter-argument sentence pair from the sentences registered in the evaluation classification list (step S14). After that, the pair creation unit 12 outputs the generated sentence pair to the learning unit 13 and the learning unit 13.
  • the learning unit 13 receives the input of the sentence pair from the pair creation unit 12. Next, the learning unit 13 inserts tokens representing the polarities of the input side and the output side of the sentence pair at the end of the input sentence (step S15).
  • the learning unit 13 learns the generation model using BERT for the sentence pair in which tokens indicating the polarities of the input side and the output side of the sentence pair are inserted at the end of the input sentence. Further, the learning unit 13 performs fine-tuning on the trained generative model using the data of the support utterance and the data of the counter-argument utterance collected manually, and performs detailed learning (step S16). After that, the learning unit 13 outputs the trained generation model to the utterance generation unit 15. The learning unit 13 acquires and holds a generative model that has already been played from the learning unit 13. This completes the learning process.
  • FIG. 7 is a flowchart of the utterance generation process by the utterance generation device.
  • the input unit 14 receives the input of the utterance input sentence to be the target of the dialogue (step S21). Further, the input unit 14 acquires an instruction of the position of the response to the utterance input sentence.
  • the input unit 14 performs an evaluation process for the subject of the utterance input sentence, and specifies the polarity on the input side. Further, the input unit 14 specifies the polarity on the output side according to the instruction from the position of the response (step S22).
  • the input unit 14 inserts tokens representing the polarities of the specified input side and output side at the end of the utterance input sentence to generate input data (step S23). Then, the input unit 14 outputs the input data in which the tokens representing the polarities of the input side and the output side are inserted at the end of the utterance input sentence to the utterance generation unit 15.
  • the utterance generation unit 15 receives input of input data from the input unit 14. Then, the utterance generation unit 15 generates a response sentence which is a supportive utterance or a counter-speech to the utterance input sentence according to the polarity of the output side by using the trained generation model (step S24). After that, the utterance generation unit 15 outputs the generated response sentence to the output unit 16.
  • the output unit 16 receives an input of a response sentence which is a supportive utterance or a counter-argument utterance to the utterance input sentence from the utterance generation unit 15. Then, the output unit 16 outputs the acquired response statement (step S25).
  • the utterance generator 1 extracts a sentence having a subject of interest and including an evaluation expression from the Web, and generates a sentence pair of an instruction and a sentence pair of a counterargument depending on whether or not the polarities of the evaluations are the same. do. Then, the utterance generation device 1 learns the generation model using the polarized sentence pairs. Further, when the utterance generation device 1 receives the utterance input sentence to be the target of the dialogue, the utterance generation device 1 determines the polarity of the utterance input sentence and also determines the polarity of the output side.
  • the utterance generation device 1 After that, the utterance generation device 1 generates and outputs a supportive utterance or a counter-argument utterance which is a response sentence to the utterance input sentence by using the polarities of the input side and the output side together with the utterance input sentence.
  • the utterance generation device 1 is a generation model capable of robustly generating instructional utterances and counterargument utterances for any proposition by learning using polarized support and counterargument sentence pairs. It is possible to build. In addition to learning, by giving the input utterance position and the output utterance position as input, it is possible to suppress the exchange of instructions and counterarguments in the output, and it is possible to have an appropriate dialogue. Will be. The processing performance of the discussion dialogue in the dialogue system can be improved, and a smooth discussion dialogue can be constructed.
  • each component of each of the illustrated devices is a functional concept, and does not necessarily have to be physically configured as shown in the figure. That is, the specific forms of distribution and integration of each device are not limited to those shown in the figure, and all or part of them may be functionally or physically dispersed or physically distributed in arbitrary units according to various loads and usage conditions. Can be integrated and configured. Further, each processing function performed by each device is realized by a CPU (Central Processing Unit) and a program that is analyzed and executed by the CPU, or hardware by wired logic. Can be realized as.
  • CPU Central Processing Unit
  • the utterance generator 1 can be implemented by installing a learning program that executes the above information processing as package software or online software on a desired computer.
  • the information processing device can function as the utterance generation device 1.
  • the information processing device referred to here includes a desktop type or notebook type personal computer.
  • information processing devices include smartphones, mobile communication terminals such as mobile phones and PHS (Personal Handy-phone System), and slate terminals such as PDAs (Personal Digital Assistants). Is done.
  • the utterance generation device 1 can be implemented as a management server device in which the terminal device used by the user is a client and the service related to the above management process is provided to the client.
  • the management server device is implemented as a server device that receives a config input request as an input and provides a management service for inputting a config.
  • the management server device may be implemented as a Web server, or may be implemented as a cloud that provides services related to the above management processing by outsourcing.
  • FIG. 8 is a diagram showing an example of a computer that executes a learning program.
  • the computer 1000 has, for example, a memory 1010 and a CPU 1020.
  • the computer 1000 also has a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. Each of these parts is connected by a bus 1080.
  • the memory 1010 includes a ROM (Read Only Memory) 1011 and a RAM (Random Access Memory) 1012.
  • the ROM 1011 stores, for example, a boot program such as a BIOS (BASIC Input Output System).
  • BIOS BASIC Input Output System
  • the hard disk drive interface 1030 is connected to the hard disk drive 1090.
  • the disk drive interface 1040 is connected to the disk drive 1100.
  • a removable storage medium such as a magnetic disk or an optical disk is inserted into the disk drive 1100.
  • the serial port interface 1050 is connected to, for example, a mouse 1110 and a keyboard 1120.
  • the video adapter 1060 is connected to, for example, the display 1130.
  • the hard disk drive 1090 stores, for example, the OS 1091, the application program 1092, the program module 1093, and the program data 1094. That is, the learning program that defines each process of the utterance generation device 1 having the same function as the utterance generation device 1 is implemented as a program module 1093 in which a code that can be executed by a computer is described.
  • the program module 1093 is stored in, for example, the hard disk drive 1090.
  • the program module 1093 for executing the same processing as the functional configuration in the utterance generation device 1 is stored in the hard disk drive 1090.
  • the hard disk drive 1090 may be replaced by an SSD (Solid State Drive).
  • the setting data used in the processing of the above-described embodiment is stored as program data 1094 in, for example, a memory 1010 or a hard disk drive 1090. Then, the CPU 1020 reads the program module 1093 and the program data 1094 stored in the memory 1010 and the hard disk drive 1090 into the RAM 1012 as needed, and executes the process of the above-described embodiment.
  • the program module 1093 and the program data 1094 are not limited to those stored in the hard disk drive 1090, but may be stored in, for example, a removable storage medium and read by the CPU 1020 via the disk drive 1100 or the like. Alternatively, the program module 1093 and the program data 1094 may be stored in another computer connected via a network (LAN (Local Area Network), WAN (Wide Area Network), etc.). Then, the program module 1093 and the program data 1094 may be read from another computer by the CPU 1020 via the network interface 1070.
  • LAN Local Area Network
  • WAN Wide Area Network

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Machine Translation (AREA)

Abstract

In the present invention, in an evaluation expression extraction step, a plurality of sentences are collected and evaluation sentences which have a specified subject of interest and include an evaluation expression are extracted. In a pair creation step, evaluation sentences are combined to generate a sentence pair comprising an input sentence and an output sentence on the basis of the polarity of an evaluation of each of the evaluation sentences extracted in the evaluation expression extraction step. In a learning step, a generation model is learned on the basis of the polarity of the evaluations of the input sentences and output sentences and the sentence pairs generated in the sentence pair creation step. In an input step, information about the standpoint of a response sentence with respect to an utterance input sentence representing an utterance composing a dialog and the utterance input sentence is acquired, the polarity of the evaluations of the utterance input sentence and the response sentence is identified, and input data is generated. In an utterance generation step, the input data generated in the input step is acquired, and the generation model generated in the learning step is used to generate and output a response sentence on the basis of the utterance input sentence and the polarities of the utterance input sentence and the response sentence.

Description

学習方法、学習プログラム及び学習装置Learning methods, learning programs and learning devices
　本発明は、学習方法、学習プログラム及び学習装置に関する。 The present invention relates to a learning method, a learning program and a learning device.
　対話システムにおいて、人間はコンピュータと対話を行い、種々の情報を取得や要望の充足を実現する。また、所定のタスクを達成するだけではなく、日常会話を行う対話システムも存在し、そのようなシステムによって人間は精神的な安定を得たり、承認欲を満たしたり、信頼関係を築いたりする。 In a dialogue system, humans interact with a computer to acquire various information and satisfy requests. There are also dialogue systems that not only accomplish certain tasks, but also engage in daily conversations, which allow humans to gain mental stability, satisfy their desire for approval, and build relationships of trust.
　一方、タスク達成や日常会話ではなく、議論をコンピュータによって実現するための研究も進められている。議論は人間の価値判断を変えたり、思考を整理したりする働きがあり、人間にとって重要な役割を果たす。例えば、意見をノードとするグラフデータを用いて、ユーザ発話をノードにマッピングし、マッピングされたノードと接続関係にあるノードをシステム発話としてユーザに返すことで議論を行う技術が提案されている。グラフデータは、例えば、「永住するなら田舎よりも都会がよい」といった予め設定された議論のテーマに基づき、人手で作成される。人手により作成された議論のデータを用いることで、特定の話題についての議論が可能となる。 On the other hand, research is also underway to realize discussions using computers rather than task achievement and daily conversation. Disputation has the function of changing human value judgments and organizing thoughts, and plays an important role for humans. For example, a technique has been proposed in which user utterances are mapped to nodes using graph data with opinions as nodes, and nodes having a connection relationship with the mapped nodes are returned to the user as system utterances for discussion. Graph data is manually created based on a preset discussion theme, such as "If you live permanently, the city is better than the countryside." By using the data of the discussion created by hand, it is possible to discuss a specific topic.
　このようなグラフデータを用いて議論を行う技術で提案されているような対話システムでは、特定の話題（クローズドドメイン）について深い議論が可能である一方で、予め設定された特定の議論テーマを逸脱するユーザ発話には適切に応答することが困難である。この問題を解決するために、任意の話題について議論のためのグラフデータを予め作成しておくアプローチが考えられるが、議論のテーマは無数に存在することから、カバレッジが低くなるため現実的ではない。 In a dialogue system such as that proposed in a technique for discussing using graph data like this, while it is possible to have a deep discussion on a specific topic (closed domain), it deviates from a specific preset discussion theme. It is difficult to respond appropriately to the user's speech. In order to solve this problem, an approach to create graph data for discussion on any topic in advance can be considered, but it is not realistic because there are innumerable discussion themes and the coverage is low. ..
　この問題に対応するため、ディープラーニング（Deep　Learning）を用いた発話生成モデルを適用する手法が提案されている。この手法では、入力発話に対する支持発話（入力発話の内容を支持する具体的な理由を述べる発話）を生成する問題が対象にされる。具体的な手法として、入力発話や支持発話に頻出するキーワードやカテゴリの情報を用いて、ＳＮＳやインターネット百科事典などの外部知識から、キーワードやカテゴリに類似する情報が学習用の知識として獲得される。 In order to deal with this problem, a method of applying an utterance generation model using deep learning has been proposed. This method targets the problem of generating a supportive utterance for an input utterance (an utterance that states a specific reason for supporting the content of the input utterance). As a specific method, information similar to keywords and categories is acquired as learning knowledge from external knowledge such as SNS and Internet encyclopedias by using information on keywords and categories that frequently appear in input utterances and supportive utterances. ..
　しかしながら、上述したディープラーニングを用いた発話生成モデルを適用する手法では、支持発話の生成を対象とした技術であり、反論発話の生成に適用することが困難である。支持発話は、入力発話と類似する内容の文、言い換えれば入力発話と近い意味を持つ文を生成すればよく、入力発話や支持発話に頻出するキーワードやカテゴリなどの表現（例えば、「～は楽しい」）に類似する情報を事前知識として獲得すればよい。意味の類似は、上述したディープラーニングを用いた発話生成モデルを適用する手法で利用されるような一般的なディープラーニングベースの発話生成モデルで扱うことが可能である。しかし、反論発話では、入力発話と類似する内容ではなく、異なる意味の文、より具体的には立場が逆の文を生成することが要求される。このため、反論発話を取り扱うためには、入力と出力とで立場が逆になっているか否かについて明示的に扱うことが求められる。 However, the method of applying the utterance generation model using deep learning described above is a technique for generating supportive utterances, and is difficult to apply to the generation of counter-argument utterances. For the support utterance, it is sufficient to generate a sentence having a content similar to the input utterance, in other words, a sentence having a meaning similar to the input utterance, and expressions such as keywords and categories that frequently appear in the input utterance and the support utterance (for example, "... is fun". Information similar to ")) may be acquired as prior knowledge. Semantic similarity can be dealt with in a general deep learning-based utterance generation model as used in the method of applying the utterance generation model using deep learning described above. However, counter-argument utterances are required to generate sentences with different meanings, more specifically sentences with opposite positions, rather than similar content to input utterances. Therefore, in order to handle counter-argument utterances, it is necessary to explicitly handle whether or not the positions of input and output are reversed.
　人手で収集したデータを用いた実験では、一般的な発話生成モデルを用いた反論発話生成器において、誤って支持が生成されるという事象が発生した。誤りとして、例えば「海は楽しい」に対する反論として「魚が美味しい」という出力がなされる場合がある。誤りの事例を基に人手で収集したデータを調査したところ、反論では「Ｘは良い」に対してＸの部分Ａに言及し、「Ａが悪い」と述べる事例や、ＸとＹとを比較して「Ｙの方が良い」と述べる事例が存在した。 In an experiment using data collected manually, an event occurred in which a support was erroneously generated in a counter-argument utterance generator using a general utterance generation model. As an error, for example, as a counterargument to "the sea is fun", "fish is delicious" may be output. When we investigated the data collected manually based on the case of error, in the counterargument, we mentioned the part A of X against "X is good" and compared the case of saying "A is bad" and X and Y. Then, there was a case where "Y is better".
　一般的な発話生成モデルでは、立場が逆になる表現は、内部で類似した情報として扱われるため、必ずしも適切に逆の立場が出力できるとは限らない。学習に使用できるデータを無数に収集すれば、これらの対応関係を全て学習できる可能性があるが、議論で扱われる話題は多岐に渡るため、現実的ではない。したがって、反論発話を取り扱うためには、立場が逆になっている文を学習に使用できる程度で的確に判定し、それらを捉えるような情報をモデルに入力して学習を行う手法を用いることが好ましい。 In a general utterance generation model, expressions with opposite positions are treated as similar information internally, so it is not always possible to output the opposite positions appropriately. If we collect innumerable data that can be used for learning, it may be possible to learn all of these correspondences, but this is not realistic because the topics covered in the discussion are diverse. Therefore, in order to handle counter-argument utterances, it is necessary to use a method of accurately judging sentences whose positions are reversed to the extent that they can be used for learning, and inputting information that captures them into the model for learning. preferable.
　ただし、Ｗｅｂテキストの中に含まれる文には多くのノイズが含まれるため、支持や反論を自動的に推定することは難しい。そこで、従来の知識獲得手法で代表的なキーワードベースの推定方法を用いることが考えられる。この方法は、例えば、文頭に「したがって」とあれば、対象の文と直前の文は支持の関係にあると推定し、文頭に「しかし」とあれば、対象の文と直前の文は反論の関係にあると推定する方法である。しかし、この方法の場合、獲得できる文ペアの規模が十分であるか否かが明らかではない。そのため、単に代表的なキーワードベースの推定方法を用いた場合、反論発話を取り扱う適切な対話システムを構築することは難しい。 However, it is difficult to automatically estimate support and counterarguments because the sentences contained in the Web text contain a lot of noise. Therefore, it is conceivable to use a typical keyword-based estimation method in the conventional knowledge acquisition method. This method presumes that, for example, if "therefore" is at the beginning of a sentence, the target sentence and the immediately preceding sentence are in a supportive relationship, and if there is "but" at the beginning of the sentence, the target sentence and the immediately preceding sentence are countered. It is a method of presuming that there is a relationship of. However, in the case of this method, it is not clear whether or not the size of the sentence pair that can be acquired is sufficient. Therefore, it is difficult to construct an appropriate dialogue system that handles counter-argument utterances when simply using a typical keyword-based estimation method.
　本発明は、上記に鑑みてなされたものであって、対話システムにおける議論対話の処理性能を向上させることを目的とする。 The present invention has been made in view of the above, and an object of the present invention is to improve the processing performance of discussion dialogue in a dialogue system.
　上述した課題を解決し、目的を達成するために、評価表現抽出工程は、複数の文を収集し、指定された着目主語を有し且つ評価表現を含む評価文を抽出する。ペア作成工程は、前記評価表現抽出工程において抽出された前記評価文のそれぞれの評価の極性を基に、前記評価文を組み合わせて入力文と出力文の文ペアを生成する。学習工程は、前記文ペア作成工程において生成された前記文ペア、並びに、前記入力文及び前記出力文の前記評価の極性を基に、生成モデルの学習を行う。入力工程は、対話の対象となる発話を表す発話入力文及び前記発話入力文に対する応答文の立場の情報を取得し、前記発話入力文及び前記応答文の前記評価の極性を特定して入力データを生成する。発話生成工程は、前記入力工程において生成された入力データを取得して、前記発話入力文、並びに、前記発話入力文及び前記応答文の極性を基に、前記学習工程において生成された前記生成モデルを用いて、前記応答文を生成して出力する。 In order to solve the above-mentioned problems and achieve the purpose, the evaluation expression extraction process collects a plurality of sentences and extracts the evaluation sentences having the specified subject of interest and including the evaluation expression. The pair creation step generates a sentence pair of an input sentence and an output sentence by combining the evaluation sentences based on the polarity of each evaluation of the evaluation sentences extracted in the evaluation expression extraction step. In the learning step, the generation model is learned based on the sentence pair generated in the sentence pair creation step, and the polarities of the evaluations of the input sentence and the output sentence. In the input process, information on the position of the utterance input sentence representing the utterance to be dialogued and the response sentence to the utterance input sentence is acquired, and the polarity of the evaluation of the utterance input sentence and the response sentence is specified and the input data is specified. To generate. The utterance generation step acquires the input data generated in the input step, and the generation model generated in the learning step based on the utterance input sentence and the polarities of the utterance input sentence and the response sentence. Is used to generate and output the response statement.
　本発明によれば、対話システムにおける議論対話の処理性能を向上させることができる。 According to the present invention, it is possible to improve the processing performance of the discussion dialogue in the dialogue system.
図１は、発話生成装置のブロック図である。FIG. 1 is a block diagram of an utterance generator. 図２は、発話生成の流れの概要を表す図である。FIG. 2 is a diagram showing an outline of the flow of utterance generation. 図３は、ペア作成部により作成される文ペアの具体例を表す図である。FIG. 3 is a diagram showing a specific example of a sentence pair created by the pair creation unit. 図４は、学習データの一例を示す図である。FIG. 4 is a diagram showing an example of learning data. 図５は、発話生成装置により生成される反論発話の一例を示す図である。FIG. 5 is a diagram showing an example of a counter-argument utterance generated by the utterance generator. 図６は、発話生成装置による学習処理のフローチャートである。FIG. 6 is a flowchart of the learning process by the utterance generator. 図７は、発話生成装置による発話生成処理のフローチャートである。FIG. 7 is a flowchart of the utterance generation process by the utterance generation device. 図８は、学習プログラムを実行するコンピュータの一例を示す図である。FIG. 8 is a diagram showing an example of a computer that executes a learning program.
　以下に、本願の開示する学習方法、学習プログラム及び学習装置の一実施形態を図面に基づいて詳細に説明する。なお、以下の実施形態により本願の開示する学習方法、学習プログラム及び学習装置が限定されるものではない。 Hereinafter, one embodiment of the learning method, learning program, and learning device disclosed in the present application will be described in detail with reference to the drawings. The learning method, learning program and learning device disclosed in the present application are not limited by the following embodiments.
［発話生成装置の構成］
　図１は、発話生成装置のブロック図である。また、図２は、発話生成の流れの概要を表す図である。図１及び図２を参照して、発話生成装置１の構成について説明する。 [Configuration of utterance generator]
FIG. 1 is a block diagram of an utterance generator. Further, FIG. 2 is a diagram showing an outline of the flow of utterance generation. The configuration of the utterance generation device 1 will be described with reference to FIGS. 1 and 2.
　発話生成装置１は、サーバなどの学習装置である。発話生成装置１は、入力された分に対して反論発話を生成し出力する装置である。図１に示すように、発話生成装置１は、評価表現抽出部１１、ペア作成部１２、学習部１３、入力部１４、発話生成部１５及び出力部１６を有する。 The utterance generation device 1 is a learning device such as a server. The utterance generation device 1 is a device that generates and outputs a counter-argument utterance with respect to the input amount. As shown in FIG. 1, the utterance generation device 1 includes an evaluation expression extraction unit 11, a pair creation unit 12, a learning unit 13, an input unit 14, an utterance generation unit 15, and an output unit 16.
　評価表現抽出部１１は、図２のステップＳ１に示すように、Ｗｅｂ上に存在するＷｅｂテキストを収集する。また、評価表現抽出部１１は、評価対象とする文の主語となる着目主語の入力を受ける。 The evaluation expression extraction unit 11 collects Web texts existing on the Web, as shown in step S1 of FIG. Further, the evaluation expression extraction unit 11 receives the input of the subject of interest, which is the subject of the sentence to be evaluated.
　次に、評価表現抽出部１１は、収集したＷｅｂテキストに対して形態素解析を行う。次に、評価表現抽出部１１は、形態素解析した文を用いて、Ｗｅｂテキストの話題を表すキーワードを抽出する焦点語抽出を行う。また、評価表現抽出部１１は、形態素解析した文を用いて固有名詞抽出を行う。また、評価表現抽出部１１は、形態素解析した文を用いて、好き、嫌い、便利などといった評価情報を抽出する評価表現抽出を行う。また、評価表現抽出部１１は、形態素解析した文を用いて、否定表現の有無などを抽出するモダリティ抽出を行う。また、評価表現抽出部１１は、形態素解析した文を用いて、各文が質問であるか詳述であるかといったどのような行為における対話であるかを推定する対話行為推定を行う。上述した評価表現抽出は一般的な言語処理タスクであるため、上述した評価表現抽出の実現のために用いる言語解析機に特に制限はない。 Next, the evaluation expression extraction unit 11 performs morphological analysis on the collected Web text. Next, the evaluation expression extraction unit 11 performs focal word extraction for extracting keywords representing the topic of the Web text by using the sentence analyzed by the morphological element. Further, the evaluation expression extraction unit 11 extracts the proper noun using the sentence analyzed by the morphological element. Further, the evaluation expression extraction unit 11 extracts evaluation expressions for extracting evaluation information such as likes, dislikes, and conveniences by using the sentences analyzed by morphological analysis. Further, the evaluation expression extraction unit 11 performs modality extraction for extracting the presence / absence of a negative expression or the like by using the sentence analyzed by the morphological element. Further, the evaluation expression extraction unit 11 uses the sentences analyzed by morphological analysis to estimate the dialogue action in which the dialogue is estimated, such as whether each sentence is a question or a detailed sentence. Since the above-mentioned evaluation expression extraction is a general language processing task, there is no particular limitation on the language analyzer used to realize the above-mentioned evaluation expression extraction.
　評価表現抽出部１１は、上述した各解析を行ったうえで、着目主語を有する評価表現を含む文を抽出する。例えば、着目主語が「Ｘ」である場合、評価表現抽出部１１は、「Ｘは」で始まる文で且つＸに対しての評価を含む文である評価文を抽出する。 The evaluation expression extraction unit 11 extracts a sentence including an evaluation expression having a subject of interest after performing each of the above-mentioned analyzes. For example, when the subject of interest is "X", the evaluation expression extraction unit 11 extracts an evaluation sentence that starts with "X is" and includes an evaluation for X.
　次に、評価表現抽出部１１は、評価情報やモダリティ抽出の結果などを用いて、抽出した着目主語を有する評価表現を含む評価文を着目主語に対する評価の極性がポジティブな極性の文とネガティブな極性の文とに分ける。文の評価の極性とは、文の主語に対して支持する立場をとるか反対の立場をとるかを表す情報であり、支持する立場であればポジティブな極性と呼び、反対の立場であればネガティブな極性と呼ぶ。以下では、文の評価の極性を単に極性と呼ぶ。そして、評価表現抽出部１１は、抽出した着目主語を含む評価文を、着目主語に対するポジティブな極性を含む文とネガティブな極性の文とに分類して登録した評価分類リストを生成する。 Next, the evaluation expression extraction unit 11 uses the evaluation information, the result of modality extraction, and the like to describe an evaluation sentence including an evaluation expression having the extracted subject of interest as a sentence having a positive evaluation polarity and a negative evaluation sentence with respect to the subject of interest. Divide into polar sentences. The polarity of a sentence's evaluation is information that indicates whether to take a supportive position or an opposite position to the subject of the sentence. Called negative polarity. In the following, the polarity of sentence evaluation is simply referred to as polarity. Then, the evaluation expression extraction unit 11 classifies and registers the evaluation sentence including the extracted subject of interest into a sentence having a positive polarity with respect to the subject of interest and a sentence having a negative polarity, and generates an evaluation classification list.
　ここで、立場が逆になっている文を学習に使用できる程度で的確に判定し、それらを捉えるような情報を生成モデルに入力して学習を行うためには、出力される意見が着目主語に対するポジティブな意見かネガティブな意見かを捉えることが重要である。そこで、本実施例に係る評価表現抽出部１１は、上述したように着目主語に対する極性に応じて文の分類を行う。 Here, in order to accurately judge sentences whose positions are reversed to the extent that they can be used for learning, and to input information that captures them into the generative model for learning, the output opinion is the subject of interest. It is important to capture positive or negative opinions about. Therefore, the evaluation expression extraction unit 11 according to this embodiment classifies sentences according to the polarity with respect to the subject of interest as described above.
　例えば、図２のステップＳ２に示すように、評価表現抽出部１１は、「Ｘは」で始まるポジティブな極性の文と、「Ｘは」で始まるネガティブな極性の文とに分類する。図２において、Ｓｅｎｔ（Ｘ，＋）ｎは、Ｘが「Ｘは」で始まる文であることを表し、＋がポジティブな文であることを表し、ｎは各文に連番で振られた識別番号を表す。また、Ｓｅｎｔ（Ｘ，＋）ｎにおける、－がネガティブな文であることを表し、他の記号は、Ｓｅｎｔ（Ｘ，＋）ｎの場合と同様である。 For example, as shown in step S2 of FIG. 2, the evaluation expression extraction unit 11 classifies a sentence having a positive polarity starting with "X is" and a sentence having a negative polarity starting with "X is". In FIG. 2, Sent (X, +) n indicates that X is a sentence starting with "X is", + indicates that it is a positive sentence, and n is assigned a serial number to each sentence. Represents an identification number. Further, in Sent (X, +) n , − indicates that it is a negative sentence, and the other symbols are the same as in the case of Sent (X, +) n .
　評価表現抽出部１１は、言語パターンとして着目主語で始まる評価文を獲得することで、議論の対象が着目主語である文を適切に抽出することができる。評価表現抽出部１１は、生成した評価分類リストをペア作成部１２へ出力する。 The evaluation expression extraction unit 11 can appropriately extract a sentence whose subject of discussion is the subject of interest by acquiring an evaluation sentence starting with the subject of interest as a language pattern. The evaluation expression extraction unit 11 outputs the generated evaluation classification list to the pair creation unit 12.
　ペア作成部１２は、評価分類リストの入力を評価表現抽出部１１から受ける。次に、ペア作成部１２は、取得した評価分類リストを用いて、支持の文ペアとして同じ極性の評価文を組み合わせて一方を入力側の文とし他方を出力側の文とした文ペアを生成する。すなわち、ペア作成部１２は、ポジティブな極性の評価文とポジティブな極性の評価文とを組み合わせて文ペアとし、さらに、ネガティブな極性の評価文とネガティブな極性の評価文とを組み合わせて文ペアとする。この文ペアは、相互に支持となる文の組み合わせである。以下では、入力側の文を単に入力側と呼び、出力側の文を単に出力側と呼ぶ場合がある。 The pair creation unit 12 receives the input of the evaluation classification list from the evaluation expression extraction unit 11. Next, the pair creation unit 12 uses the acquired evaluation classification list to combine evaluation sentences having the same polarity as support sentence pairs to generate a sentence pair in which one is an input side sentence and the other is an output side sentence. do. That is, the pair creation unit 12 combines a positive polarity evaluation sentence and a positive polarity evaluation sentence to form a sentence pair, and further combines a negative polarity evaluation sentence and a negative polarity evaluation sentence to form a sentence pair. And. This sentence pair is a combination of mutually supportive sentences. In the following, the statement on the input side may be simply referred to as the input side, and the statement on the output side may be simply referred to as the output side.
　例えば、図２のステップＳ３に示すように、ペア作成部１２は、入力側をＳｅｎｔ（Ｘ，＋）１とし、出力側をＳｅｎｔ（Ｘ，＋）２として、支持の文ペアとする。また、ペア作成部１２は、入力側をＳｅｎｔ（Ｘ，－）１とし、出力側をＳｅｎｔ（Ｘ，－）２として、支持の文ペアとする。 For example, as shown in step S3 of FIG. 2, the pair creation unit 12 has Sent (X, +) 1 as the input side and Sent (X, +) 2 as the output side, and is a supportive sentence pair. Further, the pair creation unit 12 has Sent (X, −) 1 as the input side and Sent (X, −) 2 as the output side, and is a supportive sentence pair.
　また、ペア作成部１２は、取得した評価分類リストを用いて、反論の文ペアとして異なる極性の評価文を組み合わせて一方を入力側とし他方を出力側とした文ペアを生成する。すなわち、ペア作成部１２は、入力側をポジティブな極性の評価文とし、出力側のネガティブな極性の評価文として組み合わせて文ペアとする。また、ペア作成部１２は、入力側をネガティブな極性の評価文とし、出力側をポジティブな極性の評価文として組み合わせて文ペアとする。この文ペアは、相互に反論となる文の組み合わせである。 Further, the pair creation unit 12 uses the acquired evaluation classification list to combine evaluation sentences of different polarities as a counter-argument sentence pair to generate a sentence pair with one as the input side and the other as the output side. That is, the pair creation unit 12 uses the input side as an evaluation sentence having a positive polarity and combines them as an evaluation sentence having a negative polarity on the output side to form a sentence pair. Further, the pair creation unit 12 combines the input side as an evaluation sentence having a negative polarity and the output side as an evaluation sentence having a positive polarity to form a sentence pair. This sentence pair is a combination of sentences that are mutually counter-arguing.
　例えば、図２のステップＳ３に示すように、ペア作成部１２は、入力側をＳｅｎｔ（Ｘ，＋）１とし、出力側をＳｅｎｔ（Ｘ，－）１として、反論の文の文ペアとする。また、ペア作成部１２は、入力側をＳｅｎｔ（Ｘ，＋）２とし、出力側をＳｅｎｔ（Ｘ，－）２として、反論の文の文ペアとする。 For example, as shown in step S3 of FIG. 2, the pair creation unit 12 sets the input side to Sent (X, +) 1 and the output side to Sent (X,-) 1 to form a sentence pair of counterarguments. .. Further, in the pair creation unit 12, the input side is Sent (X, +) 2 and the output side is Sent (X, −) 2 , and the sentence pair is a counter-argument sentence.
　図３は、ペア作成部により作成される文ペアの具体例を表す図である。図３は、着目主語が「ラーメン」の場合の反論文の文ペアの一例である。この場合、入力側のデータの極性がポジティブであり、出力側の極性がネガティブである。ペア作成部１２は、ラーメンを着目主語とした場合に、図３に示される各文ペアを生成することができる。図３に示すように、極性と言語パターンを利用することで、ペア作成部１２は、反論らしい文ペアを適切に獲得できる。 FIG. 3 is a diagram showing a specific example of a sentence pair created by the pair creation unit. FIG. 3 is an example of a sentence pair of an anti-thesis when the subject of interest is “ramen”. In this case, the polarity of the data on the input side is positive, and the polarity on the output side is negative. The pair creation unit 12 can generate each sentence pair shown in FIG. 3 when ramen is the subject of interest. As shown in FIG. 3, by using the polarity and the language pattern, the pair creation unit 12 can appropriately acquire a sentence pair that seems to be a counterargument.
　学習部１３は、ペア作成部１２が生成した文ペアを取得する。そして、学習部１３は、取得した文ペアを用いて、発話生成モデルの学習を行う。学習部１３は、各文がポジティブの極性を有する文かネガティブの極性を有する文かを表す極性ガイドとして、入力側の極性と出力側の極性とを、入力文の末尾にトークンとして挿入する。そして、入力側の極性と出力側の極性がトークンとして挿入された入力文を用いて、入出力の学習を行う。例えば、図４は、学習データの一例を示す図である。図４は、着目主語をラーメンとする、反論の文のペアであり、入力側の文がポジティブな極性を有し、出力側の文がネガティブな極性を有する。図４における矢印は、入力された文から出力される文への変換を表す。また、矢印の前の正負の符号は、入力側の極性と出力側の極性とを表す。 The learning unit 13 acquires the sentence pair generated by the pair creation unit 12. Then, the learning unit 13 learns the utterance generation model using the acquired sentence pair. The learning unit 13 inserts the polarity on the input side and the polarity on the output side as tokens at the end of the input sentence as a polarity guide indicating whether each sentence has a positive polarity or a negative polarity. Then, input / output learning is performed using an input statement in which the polarity on the input side and the polarity on the output side are inserted as tokens. For example, FIG. 4 is a diagram showing an example of learning data. FIG. 4 shows a pair of counter-argument sentences in which the subject of interest is ramen, in which the sentence on the input side has a positive polarity and the sentence on the output side has a negative polarity. The arrow in FIG. 4 represents the conversion from the input sentence to the output sentence. The positive and negative signs before the arrow indicate the polarity on the input side and the polarity on the output side.
　学習部１３は、例えば、事前学習手法であるＢＥＲＴ（Bidirectional　Encoder　Representations　from　Transformers）をエンコード及びデコードに利用する手法を用いた生成モデルを使用して学習を行う。これにより、図２のステップＳ４に示すように、学習部１３は、学習済みの生成モデルを生成する。ただし、学習に用いる手法はＢＥＲＴに限らず、学習部１３は、他の発話生成モデル学習アルゴリズムを用いてもよい。 The learning unit 13 learns using, for example, a generative model using a method that uses BERT (Bidirectional Encoder Representations from Transformers), which is a pre-learning method, for encoding and decoding. As a result, as shown in step S4 of FIG. 2, the learning unit 13 generates a trained generative model. However, the method used for learning is not limited to BERT, and the learning unit 13 may use another utterance generation model learning algorithm.
　さらに、本実施形態に係る学習部１３は、ＢＥＲＴによる学習に加えて、人手で収集した支持発話のデータ及び反論発話のデータを用いてファインチューニングを行い、生成モデルの精度を上げる。学習部１３は、ファインチューニングを行わなくてもよいが、ファインチューニングを行った方が正確な支持発話及び反論発話の生成が可能である。学習部１３は、学習済みの生成モデルを発話生成部１５へ出力する。 Further, the learning unit 13 according to the present embodiment performs fine tuning using the data of the support utterance and the data of the counter-argument utterance collected manually in addition to the learning by BERT to improve the accuracy of the generated model. The learning unit 13 does not have to perform fine tuning, but it is possible to generate more accurate supportive utterances and counter-argument utterances by performing fine tuning. The learning unit 13 outputs the trained generation model to the utterance generation unit 15.
　入力部１４は、対話の対象とする発話された文の入力を受ける。この文の入力は、人が発話した言葉を音声認識して生成された文を取得してもよいし、発話にあたる文を利用者が手作業で入力してもよい。以下では、この入力された文を発話入力文と呼ぶ。また、入力部１４は、発話入力文に対する応答文として支持の文を出力するか反論の文を要求するかといった応答の立場の指示を受ける。入力部１４は、この応答の立場の指示を予め取得していてもよいし、発話入力文とともに入力を受けてもよい。 The input unit 14 receives the input of the spoken sentence to be the target of the dialogue. As for the input of this sentence, the sentence generated by voice recognition of the words spoken by a person may be acquired, or the sentence corresponding to the utterance may be manually input by the user. In the following, this input sentence is referred to as an utterance input sentence. Further, the input unit 14 receives an instruction from the position of response such as whether to output a support sentence or request a counterargument sentence as a response sentence to the utterance input sentence. The input unit 14 may have acquired the instruction of the position of this response in advance, or may receive the input together with the utterance input sentence.
　そして、入力部１４は、発話入力文の主語に対する極性を判定する。例えば、入力部１４は、評価表現抽出部１１と同様の評価の極性の判定処理により発話入力文の主語に対する極性の推定を行う。さらに、入力部１４は、応答の立場が支持と指定されていれば発話入力文と同じ極性とし、反論と指定されていれば反対の極性として出力側の極性を決定する。そして、入力部１４は、図２のステップＳ５に示すように、入力側の極性と出力側の極性とを極性ガイドとして発話入力文の末尾にトークンとして挿入して入力データを生成する。そして、入力部１４は、発話入力文の末尾に入力文の末尾にトークンが挿入された入力データを学習済みの生成モデルを有する発話生成部１５へ出力する。 Then, the input unit 14 determines the polarity of the utterance input sentence with respect to the subject. For example, the input unit 14 estimates the polarity of the utterance input sentence with respect to the subject by the same evaluation polarity determination process as the evaluation expression extraction unit 11. Further, the input unit 14 determines the polarity on the output side as the same polarity as the utterance input sentence if the response position is designated as support, and as the opposite polarity if the response position is specified as counterargument. Then, as shown in step S5 of FIG. 2, the input unit 14 inserts the polarity on the input side and the polarity on the output side as a token at the end of the utterance input sentence as a polarity guide to generate input data. Then, the input unit 14 outputs the input data in which the token is inserted at the end of the input sentence to the utterance generation unit 15 having the trained generation model.
　発話生成部１５は、学習済みの生成モデルを学習部１３から取得する。その後、発話生成部１５は、発話入力文の末尾にトークンが挿入された入力データの入力を入力部１４から受ける。そして、発話生成部１５は、図２のステップＳ6に示すように、トークンにしたがって取得した学習モデルを用いて入力された発話入力文に対する支持発話又は反論発話を生成する。応答の立場が支持であれば、発話生成部１５は、発話入力文に対する反論発話を生成する。より具体的には、発話生成部１５は、発話入力文が主語に対してポジティブな文であれば主語に対してポジティブな文を生成し、発話入力文が主語に対してネガティブな文であれば主語に対してネガティブな文を生成する。また、応答の立場が反論であれば、発話生成部１５は、発話入力文に対する反論発話を生成する。より具体的には、発話生成部１５は、発話入力文が主語に対してポジティブな文であれば主語に対してネガティブな文を生成し、発話入力文が主語に対してネガティブな文であれば主語に対してポジティブな文を生成する。その後、発話生成部１５は、生成した発話入力文に対する応答文を出力部１６へ出力する。 The utterance generation unit 15 acquires the trained generation model from the learning unit 13. After that, the utterance generation unit 15 receives the input of the input data in which the token is inserted at the end of the utterance input sentence from the input unit 14. Then, as shown in step S6 of FIG. 2, the utterance generation unit 15 generates a supportive utterance or a counter-argument utterance for the utterance input sentence input using the learning model acquired according to the token. If the response position is supportive, the utterance generation unit 15 generates a counter-utterance to the utterance input sentence. More specifically, the utterance generation unit 15 generates a sentence positive for the subject if the utterance input sentence is a positive sentence for the subject, and the utterance input sentence is a negative sentence for the subject. For example, it produces a negative sentence for the subject. If the response position is a counterargument, the utterance generation unit 15 generates a counterargument utterance to the utterance input sentence. More specifically, the utterance generation unit 15 generates a negative sentence for the subject if the utterance input sentence is a positive sentence for the subject, and the utterance generation unit 15 may be a negative sentence for the subject. For example, it produces a positive sentence for the subject. After that, the utterance generation unit 15 outputs a response sentence to the generated utterance input sentence to the output unit 16.
　出力部１６は、発話入力文に対する支持発話又は反論発話である応答文を発話生成部１５から取得する。そして、出力部１６は、図２のステップＳ７に示すように、取得した応答文を出力する。 The output unit 16 acquires a response sentence, which is a supportive utterance or a counter-argument utterance to the utterance input sentence, from the utterance generation unit 15. Then, the output unit 16 outputs the acquired response statement as shown in step S7 of FIG.
　図５は、発話生成装置により生成される反論発話の一例を示す図である。例えば、野球は楽しいという文が入力された場合、発話生成装置１は、ルールが難しいという反論発話を生成して出力する。この場合、発話生成装置１は、野球に対するポジティブな文に対して、野球に対するネガティブな文を反論発話として出力している。また、発話生成装置１は、株は儲からないという文が入力された場合、景気は良いという反論発話を生成して出力する。この場合、発話生成装置１は、株に対するネガティブな文に対して、株に対するポジティブな文を反論発話として出力している。また、発話生成装置１は、ゴルフは楽しいという文が入力された場合、練習が必要という反論発話を生成して出力する。この場合、発話生成装置１は、ゴルフに対するポジティブな文に対して、ゴルフに対するネガティブな文を反論発話として出力している。いずれの反論発話も、入力された文に対して適切な反論となっているといえる。 FIG. 5 is a diagram showing an example of a counter-argument utterance generated by the utterance generator. For example, when a sentence that baseball is fun is input, the utterance generation device 1 generates and outputs a counter-argument that the rules are difficult. In this case, the utterance generation device 1 outputs a negative sentence for baseball as a counter-utterance to a positive sentence for baseball. Further, the utterance generator 1 generates and outputs a counter-argument that the economy is good when a sentence that the stock is not profitable is input. In this case, the utterance generator 1 outputs a positive sentence for the stock as a counter-utterance to the negative sentence for the stock. Further, the utterance generation device 1 generates and outputs a counter-argument that practice is required when a sentence that golf is fun is input. In this case, the utterance generator 1 outputs a negative sentence for golf as a counter-utterance to a positive sentence for golf. It can be said that each counter-argument utterance is an appropriate counter-argument to the input sentence.
　［生成モデル学習処理及び発話生成処理］
　次に、図６を参照して、発話生成装置１による学習処理の流れについて説明する。図６は、発話生成装置による学習処理のフローチャートである。 [Generative model learning process and utterance generation process]
Next, with reference to FIG. 6, the flow of the learning process by the utterance generation device 1 will be described. FIG. 6 is a flowchart of the learning process by the utterance generator.
　評価表現抽出部１１は、Ｗｅｂ上に存在するＷｅｂテキストを収集する（ステップＳ１１）。また、評価表現抽出部１１は、着目主語の入力を受ける。 The evaluation expression extraction unit 11 collects Web texts existing on the Web (step S11). Further, the evaluation expression extraction unit 11 receives the input of the subject of interest.
　次に、評価表現抽出部１１は、指定された着目主語を有し且つ評価表現を含む文を、収集したＷｅｂテキストから抽出する（ステップＳ１２）。 Next, the evaluation expression extraction unit 11 extracts a sentence having a designated subject of interest and including an evaluation expression from the collected Web text (step S12).
　次に、評価表現抽出部１１は、抽出した文を着目主語に対してポジティブな文とネガティブな文に分類して評価分類リストを作成する（ステップＳ１３）。その後、評価表現抽出部１１は、生成した評価分類リストをペア作成部１２へ出力する。 Next, the evaluation expression extraction unit 11 classifies the extracted sentences into positive sentences and negative sentences with respect to the subject of interest, and creates an evaluation classification list (step S13). After that, the evaluation expression extraction unit 11 outputs the generated evaluation classification list to the pair creation unit 12.
　ペア作成部１２は、評価分類リストの入力を評価表現抽出部１１から受ける。次に、ペア作成部１２は、評価分類リストに登録された文から指示の文ペアと反論の文ペアとを作成する（ステップＳ１４）。その後、ペア作成部１２は、生成した文ペアを学習部１３及へ出力する。 The pair creation unit 12 receives the input of the evaluation classification list from the evaluation expression extraction unit 11. Next, the pair creation unit 12 creates an instruction sentence pair and a counter-argument sentence pair from the sentences registered in the evaluation classification list (step S14). After that, the pair creation unit 12 outputs the generated sentence pair to the learning unit 13 and the learning unit 13.
　学習部１３は、文ペアの入力をペア作成部１２から受ける。次に、学習部１３は、文ペアの入力側及び出力側の極性を表すトークンを入力文の末尾に挿入する（ステップＳ１５）。 The learning unit 13 receives the input of the sentence pair from the pair creation unit 12. Next, the learning unit 13 inserts tokens representing the polarities of the input side and the output side of the sentence pair at the end of the input sentence (step S15).
　次に、学習部１３は、入力文の末尾に文ペアの入力側及び出力側の極性を表すトークンが挿入された文ペアに対してＢＥＲＴを用いて生成モデルの学習を行う。さらに、学習部１３は、人手で収集した支持発話のデータ及び反論発話のデータを用いて学習済みの生成モデルに対してfine-tuningを施し詳細な学習を行う（ステップＳ１６）。その後、学習部１３は、学習済みの生成モデルを発話生成部１５へ出力する。学習部１３は、楽手済みの生成モデルを学習部１３から取得して保持する。これにより学習処理が完了する。 Next, the learning unit 13 learns the generation model using BERT for the sentence pair in which tokens indicating the polarities of the input side and the output side of the sentence pair are inserted at the end of the input sentence. Further, the learning unit 13 performs fine-tuning on the trained generative model using the data of the support utterance and the data of the counter-argument utterance collected manually, and performs detailed learning (step S16). After that, the learning unit 13 outputs the trained generation model to the utterance generation unit 15. The learning unit 13 acquires and holds a generative model that has already been played from the learning unit 13. This completes the learning process.
　次に、図７を参照して、発話生成装置１による発話生成処理の流れについて説明する。図７は、発話生成装置による発話生成処理のフローチャートである。 Next, with reference to FIG. 7, the flow of the utterance generation process by the utterance generation device 1 will be described. FIG. 7 is a flowchart of the utterance generation process by the utterance generation device.
　入力部１４は、対話の対象となる発話入力文の入力を受ける（ステップＳ２１）。さらに、入力部１４は、発話入力文に対する応答の立場の指示を取得する。 The input unit 14 receives the input of the utterance input sentence to be the target of the dialogue (step S21). Further, the input unit 14 acquires an instruction of the position of the response to the utterance input sentence.
　次に、入力部１４は、発話入力文に対して主語に対する評価の推定処理を行い、入力側の極性を特定する。さらに、入力部１４は、応答の立場の指示にしたがって出力側の極性を特定する（ステップＳ２２）。 Next, the input unit 14 performs an evaluation process for the subject of the utterance input sentence, and specifies the polarity on the input side. Further, the input unit 14 specifies the polarity on the output side according to the instruction from the position of the response (step S22).
　次に、入力部１４は、特定した入力側及び出力側の極性を表すトークンを発話入力文の末尾に挿入して入力データを生成する（ステップＳ２３）。そして、入力部１４は、発話入力文の末尾に入力側及び出力側の極性を表すトークンが挿入された入力データを発話生成部１５へ出力する。 Next, the input unit 14 inserts tokens representing the polarities of the specified input side and output side at the end of the utterance input sentence to generate input data (step S23). Then, the input unit 14 outputs the input data in which the tokens representing the polarities of the input side and the output side are inserted at the end of the utterance input sentence to the utterance generation unit 15.
　発話生成部１５は、入力データの入力を入力部１４から受ける。そして、発話生成部１５は、学習済みの生成モデルを用いて出力側の極性に応じた発話入力文に対する支持発話又は反論発話である応答文を生成する（ステップＳ２４）。その後、発話生成部１５は、生成した応答文を出力部１６へ出力する。 The utterance generation unit 15 receives input of input data from the input unit 14. Then, the utterance generation unit 15 generates a response sentence which is a supportive utterance or a counter-speech to the utterance input sentence according to the polarity of the output side by using the trained generation model (step S24). After that, the utterance generation unit 15 outputs the generated response sentence to the output unit 16.
　出力部１６は、発話入力文に対する支持発話又は反論発話である応答文の入力を発話生成部１５から受ける。そして、出力部１６は、取得した応答文を出力する（ステップＳ２５）。 The output unit 16 receives an input of a response sentence which is a supportive utterance or a counter-argument utterance to the utterance input sentence from the utterance generation unit 15. Then, the output unit 16 outputs the acquired response statement (step S25).
［生成モデル学習処理及び発話生成処理による効果］
　以上に説明したように、発話生成装置１は、Ｗｅｂから着目主語を有し且つ評価表現を含む文を抽出し、評価の極性が同じか否かにより指示の文ペア及び反論の文ペアを生成する。そして、発話生成装置１は、極性付きの文ペアを用いて生成モデルの学習を行う。さらに、発話生成装置１は、対話の対象となる発話入力文を受けると、その発話入力文の極性を判定するとともに出力側の極性を決定する。その後、発話生成装置１は、発話入力文とともに入力側及び出力側の極性を用いて、発話入力文に対する応答文である支持発話又は反論発話を生成して出力する。 [Effects of generative model learning process and utterance generation process]
As described above, the utterance generator 1 extracts a sentence having a subject of interest and including an evaluation expression from the Web, and generates a sentence pair of an instruction and a sentence pair of a counterargument depending on whether or not the polarities of the evaluations are the same. do. Then, the utterance generation device 1 learns the generation model using the polarized sentence pairs. Further, when the utterance generation device 1 receives the utterance input sentence to be the target of the dialogue, the utterance generation device 1 determines the polarity of the utterance input sentence and also determines the polarity of the output side. After that, the utterance generation device 1 generates and outputs a supportive utterance or a counter-argument utterance which is a response sentence to the utterance input sentence by using the polarities of the input side and the output side together with the utterance input sentence.
　このように、発話生成装置１は、極性が付加された支持及び反論の文ペアを用いて学習を行うことで、任意の命題に対して頑健に指示発話及び反論発話を生成可能な生成モデルの構築することが可能である。また、学習だけでなく、入力された発話の立場及び出力される発話の立場を入力として与えることで、出力における指示と反論の入れ替わりを抑制することが可能となり、適切な対話を行うことが可能となる。対話システムにおける議論対話の処理性能を向上させることができ、スムースな議論対話を構築することが可能となる。 In this way, the utterance generation device 1 is a generation model capable of robustly generating instructional utterances and counterargument utterances for any proposition by learning using polarized support and counterargument sentence pairs. It is possible to build. In addition to learning, by giving the input utterance position and the output utterance position as input, it is possible to suppress the exchange of instructions and counterarguments in the output, and it is possible to have an appropriate dialogue. Will be. The processing performance of the discussion dialogue in the dialogue system can be improved, and a smooth discussion dialogue can be constructed.
［システム構成等］
　また、図示した各装置の各構成要素は機能概念的なものであり、必ずしも物理的に図示のように構成されていることを要しない。すなわち、各装置の分散及び統合の具体的形態は図示のものに限られず、その全部又は一部を、各種の負荷や使用状況等に応じて、任意の単位で機能的又は物理的に分散又は統合して構成することができる。さらに、各装置にて行われる各処理機能は、その全部又は任意の一部が、ＣＰＵ（Central　Processing　Unit）及び当該ＣＰＵにて解析実行されるプログラムにて実現され、あるいは、ワイヤードロジックによるハードウェアとして実現され得る。 [System configuration, etc.]
Further, each component of each of the illustrated devices is a functional concept, and does not necessarily have to be physically configured as shown in the figure. That is, the specific forms of distribution and integration of each device are not limited to those shown in the figure, and all or part of them may be functionally or physically dispersed or physically distributed in arbitrary units according to various loads and usage conditions. Can be integrated and configured. Further, each processing function performed by each device is realized by a CPU (Central Processing Unit) and a program that is analyzed and executed by the CPU, or hardware by wired logic. Can be realized as.
　また、本実施形態において説明した各処理のうち、自動的に行われるものとして説明した処理の全部又は一部を手動的に行うこともでき、あるいは、手動的に行われるものとして説明した処理の全部又は一部を公知の方法で自動的に行うこともできる。この他、上記文書中や図面中で示した処理手順、制御手順、具体的名称、各種のデータやパラメータを含む情報については、特記する場合を除いて任意に変更することができる。 Further, among the processes described in the present embodiment, all or part of the processes described as being automatically performed can be manually performed, or the processes described as being manually performed can be performed. All or part of it can be done automatically by a known method. In addition, the processing procedure, control procedure, specific name, and information including various data and parameters shown in the above document and drawings can be arbitrarily changed unless otherwise specified.
［プログラム］
　一実施形態として、発話生成装置１は、パッケージソフトウェアやオンラインソフトウェアとして上記の情報処理を実行する学習プログラムを所望のコンピュータにインストールさせることによって実装できる。例えば、上記の学習プログラムを情報処理装置に実行させることにより、情報処理装置を発話生成装置１として機能させることができる。ここで言う情報処理装置には、デスクトップ型又はノート型のパーソナルコンピュータが含まれる。また、その他にも、情報処理装置にはスマートフォン、携帯電話機やＰＨＳ（Personal　Handy-phone　System）等の移動体通信端末、さらには、ＰＤＡ（Personal　Digital　Assistant）等のスレート端末等がその範疇に含まれる。 [program]
As one embodiment, the utterance generator 1 can be implemented by installing a learning program that executes the above information processing as package software or online software on a desired computer. For example, by causing the information processing device to execute the above learning program, the information processing device can function as the utterance generation device 1. The information processing device referred to here includes a desktop type or notebook type personal computer. In addition, information processing devices include smartphones, mobile communication terminals such as mobile phones and PHS (Personal Handy-phone System), and slate terminals such as PDAs (Personal Digital Assistants). Is done.
　また、発話生成装置１は、ユーザが使用する端末装置をクライアントとし、当該クライアントに上記の管理処理に関するサービスを提供する管理サーバ装置として実装することもできる。例えば、管理サーバ装置は、コンフィグ投入要求を入力とし、コンフィグ投入を行う管理サービスを提供するサーバ装置として実装される。この場合、管理サーバ装置は、Webサーバとして実装することとしてもよいし、アウトソーシングによって上記の管理処理に関するサービスを提供するクラウドとして実装することとしてもかまわない。 Further, the utterance generation device 1 can be implemented as a management server device in which the terminal device used by the user is a client and the service related to the above management process is provided to the client. For example, the management server device is implemented as a server device that receives a config input request as an input and provides a management service for inputting a config. In this case, the management server device may be implemented as a Web server, or may be implemented as a cloud that provides services related to the above management processing by outsourcing.
　図８は、学習プログラムを実行するコンピュータの一例を示す図である。コンピュータ１０００は、例えば、メモリ１０１０、ＣＰＵ１０２０を有する。また、コンピュータ１０００は、ハードディスクドライブインタフェース１０３０、ディスクドライブインタフェース１０４０、シリアルポートインタフェース１０５０、ビデオアダプタ１０６０、ネットワークインタフェース１０７０を有する。これらの各部は、バス１０８０によって接続される。 FIG. 8 is a diagram showing an example of a computer that executes a learning program. The computer 1000 has, for example, a memory 1010 and a CPU 1020. The computer 1000 also has a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. Each of these parts is connected by a bus 1080.
　メモリ１０１０は、ＲＯＭ（Read　Only　Memory）１０１１及びＲＡＭ（Random　Access　Memory）１０１２を含む。ＲＯＭ１０１１は、例えば、ＢＩＯＳ（BASIC　Input　Output　System）等のブートプログラムを記憶する。ハードディスクドライブインタフェース１０３０は、ハードディスクドライブ１０９０に接続される。ディスクドライブインタフェース１０４０は、ディスクドライブ１１００に接続される。例えば磁気ディスクや光ディスク等の着脱可能な記憶媒体が、ディスクドライブ１１００に挿入される。シリアルポートインタフェース１０５０は、例えばマウス１１１０、キーボード１１２０に接続される。ビデオアダプタ１０６０は、例えばディスプレイ１１３０に接続される。 The memory 1010 includes a ROM (Read Only Memory) 1011 and a RAM (Random Access Memory) 1012. The ROM 1011 stores, for example, a boot program such as a BIOS (BASIC Input Output System). The hard disk drive interface 1030 is connected to the hard disk drive 1090. The disk drive interface 1040 is connected to the disk drive 1100. For example, a removable storage medium such as a magnetic disk or an optical disk is inserted into the disk drive 1100. The serial port interface 1050 is connected to, for example, a mouse 1110 and a keyboard 1120. The video adapter 1060 is connected to, for example, the display 1130.
　ハードディスクドライブ１０９０は、例えば、ＯＳ１０９１、アプリケーションプログラム１０９２、プログラムモジュール１０９３、プログラムデータ１０９４を記憶する。すなわち、発話生成装置１と同等の機能を持つ発話生成装置１の各処理を規定する学習プログラムは、コンピュータにより実行可能なコードが記述されたプログラムモジュール１０９３として実装される。プログラムモジュール１０９３は、例えばハードディスクドライブ１０９０に記憶される。例えば、発話生成装置１における機能構成と同様の処理を実行するためのプログラムモジュール１０９３が、ハードディスクドライブ１０９０に記憶される。なお、ハードディスクドライブ１０９０は、ＳＳＤ（Solid　State　Drive）により代替されてもよい。 The hard disk drive 1090 stores, for example, the OS 1091, the application program 1092, the program module 1093, and the program data 1094. That is, the learning program that defines each process of the utterance generation device 1 having the same function as the utterance generation device 1 is implemented as a program module 1093 in which a code that can be executed by a computer is described. The program module 1093 is stored in, for example, the hard disk drive 1090. For example, the program module 1093 for executing the same processing as the functional configuration in the utterance generation device 1 is stored in the hard disk drive 1090. The hard disk drive 1090 may be replaced by an SSD (Solid State Drive).
　また、上述した実施形態の処理で用いられる設定データは、プログラムデータ１０９４として、例えばメモリ１０１０やハードディスクドライブ１０９０に記憶される。そして、CPU１０２０は、メモリ１０１０やハードディスクドライブ１０９０に記憶されたプログラムモジュール１０９３やプログラムデータ１０９４を必要に応じてRAM１０１２に読み出して、上述した実施形態の処理を実行する。 Further, the setting data used in the processing of the above-described embodiment is stored as program data 1094 in, for example, a memory 1010 or a hard disk drive 1090. Then, the CPU 1020 reads the program module 1093 and the program data 1094 stored in the memory 1010 and the hard disk drive 1090 into the RAM 1012 as needed, and executes the process of the above-described embodiment.
　なお、プログラムモジュール１０９３やプログラムデータ１０９４は、ハードディスクドライブ１０９０に記憶される場合に限らず、例えば着脱可能な記憶媒体に記憶され、ディスクドライブ１１００等を介してＣＰＵ１０２０によって読み出されてもよい。あるいは、プログラムモジュール１０９３及びプログラムデータ１０９４は、ネットワーク（ＬＡＮ（Local　Area　Network）、ＷＡＮ（Wide　Area　Network）等）を介して接続された他のコンピュータに記憶されてもよい。そして、プログラムモジュール１０９３及びプログラムデータ１０９４は、他のコンピュータから、ネットワークインタフェース１０７０を介してＣＰＵ１０２０によって読み出されてもよい。 The program module 1093 and the program data 1094 are not limited to those stored in the hard disk drive 1090, but may be stored in, for example, a removable storage medium and read by the CPU 1020 via the disk drive 1100 or the like. Alternatively, the program module 1093 and the program data 1094 may be stored in another computer connected via a network (LAN (Local Area Network), WAN (Wide Area Network), etc.). Then, the program module 1093 and the program data 1094 may be read from another computer by the CPU 1020 via the network interface 1070.
　１　発話生成装置
　１１　評価表現抽出部
　１２　ペア作成部
　１３　学習部
　１４　入力部
　１５　発話生成部
　１６　出力部 1 Utterance generator 11 Evaluation expression extraction unit 12 Pair creation unit 13 Learning unit 14 Input unit 15 Utterance generation unit 16 Output unit

Claims (7)

  1. 　複数の文を収集し、指定された着目主語を有し且つ評価表現を含む評価文を抽出する評価表現抽出工程と、
    　前記評価表現抽出工程において抽出された前記評価文のそれぞれの評価の極性を基に、前記評価文を組み合わせて入力文と出力文の文ペアを生成するペア作成工程と、
    　前記ペア作成工程において生成された前記文ペア、並びに、前記入力文及び前記出力文の前記評価の極性を基に、生成モデルの学習を行う学習工程と、
    　対話の対象となる発話を表す発話入力文及び前記発話入力文に対する応答文の立場の情報を取得し、前記発話入力文及び前記応答文の前記評価の極性を特定して入力データを生成する入力工程と、
    　前記入力工程において生成された入力データを取得して、前記発話入力文、並びに、前記発話入力文及び前記応答文の極性を基に、前記学習工程において生成された前記生成モデルを用いて、前記応答文を生成して出力する発話生成工程と
    　を含んだことを特徴とする学習方法。     An evaluation expression extraction process that collects multiple sentences and extracts an evaluation sentence that has a specified subject of interest and includes an evaluation expression.
    A pair creation step of combining the evaluation sentences to generate a sentence pair of an input sentence and an output sentence based on the polarity of each evaluation of the evaluation sentence extracted in the evaluation expression extraction step.
    A learning step of learning a generation model based on the sentence pair generated in the pair creation step, and the polarities of the evaluations of the input sentence and the output sentence.
    Input that acquires the position information of the utterance input sentence representing the utterance to be the dialogue and the response sentence to the utterance input sentence, specifies the polarity of the evaluation of the utterance input sentence and the response sentence, and generates input data. Process and
    The input data generated in the input step is acquired, and the generation model generated in the learning step is used based on the utterance input sentence and the polarities of the utterance input sentence and the response sentence. A learning method characterized by including an utterance generation process that generates and outputs a response sentence.
  2. 　前記評価表現抽出工程は、前記評価文のそれぞれの前記評価の極性の推定を行うことを特徴とする請求項１に記載の学習方法。 The learning method according to claim 1, wherein the evaluation expression extraction step estimates the polarity of each evaluation of the evaluation sentence.
  3. 　前記ペア作成工程は、前記評価の極性が同じ前記評価文を組み合わせた支持の文ペア及び前記評価の極性が反対の前記評価文を組み合わせた反論の文ペアを生成することを特徴とする請求項１又は２に記載の学習方法。 The claim is characterized in that the pair creation step generates a supporting sentence pair in which the evaluation sentences having the same evaluation polarity and a counter-argument sentence pair in which the evaluation sentences having the opposite evaluation polarities are combined. The learning method according to 1 or 2.
  4. 　前記学習工程は、前記入力文及び前記出力文の前記評価の極性を表す情報を前記入力文に付加して学習を行うことを特徴とする請求項１～３のいずれか一つに記載の学習方法。 The learning according to any one of claims 1 to 3, wherein the learning step performs learning by adding information indicating the polarity of the evaluation of the input sentence and the output sentence to the input sentence. Method.
  5. 　前記入力工程は、前記発話入力文及び前記応答文の前記評価の極性を表す情報を前記発話入力文に付加して前記入力データを生成し、
    　前記発話生成工程は、前記発話入力文及び前記応答文の前記評価の極性を表す情報を付加された前記発話入力文に対して前記生成モデルを用いて前記応答文を生成する
    　ことを特徴とする請求項１～３のいずれか一つに記載の学習方法。     In the input step, information indicating the polarity of the evaluation of the utterance input sentence and the response sentence is added to the utterance input sentence to generate the input data.
    The utterance generation step is characterized in that the response sentence is generated by using the generation model for the utterance input sentence to which information indicating the polarity of the evaluation of the utterance input sentence and the response sentence is added. The learning method according to any one of claims 1 to 3.
  6. 　コンピュータに、請求項１～５に記載の方法を実行させるための学習プログラム。 A learning program for causing a computer to execute the method according to claims 1 to 5.
  7. 　複数の文を収集し、指定された着目主語を有し且つ評価表現を含む評価文を抽出する評価表現抽出部と、
    　前記評価表現抽出部により抽出された前記評価文のそれぞれの評価の極性を基に、前記評価文を組み合わせて入力文と出力文の文ペアを生成するペア作成部と、
    　前記ペア作成部により生成された前記文ペア、並びに、前記入力文及び前記出力文の前記評価の極性を基に、生成モデルの学習を行う学習部と、
    　対話の対象となる発話を表す発話入力文及び前記発話入力文に対する応答文の立場の情報を取得し、前記発話入力文及び前記応答文の前記評価の極性を特定して入力データを生成する入力部と、
    　前記入力部から前記入力データの入力を受けて、前記発話入力文、並びに、前記発話入力文及び前記応答文の極性を基に、前記学習部により生成された前記生成モデルを用いて、前記応答文を生成して出力する発話生成部と
    　を備えたことを特徴とする学習装置。     An evaluation expression extraction unit that collects multiple sentences and extracts evaluation sentences that have a specified subject of interest and include evaluation expressions.
    A pair creation unit that generates a sentence pair of an input sentence and an output sentence by combining the evaluation sentences based on the polarity of each evaluation of the evaluation sentence extracted by the evaluation expression extraction unit.
    A learning unit that learns a generation model based on the sentence pair generated by the pair creation unit, and the polarities of the evaluations of the input sentence and the output sentence.
    Input that acquires the position information of the utterance input sentence representing the utterance to be the dialogue and the response sentence to the utterance input sentence, specifies the polarity of the evaluation of the utterance input sentence and the response sentence, and generates input data. Department and
    The response is received from the input unit using the generation model generated by the learning unit based on the input of the input data, the utterance input sentence, and the polarities of the utterance input sentence and the response sentence. A learning device characterized by having an utterance generator that generates and outputs sentences.
PCT/JP2020/044396 2020-11-27 2020-11-27 Learning method, learning program, and learning device WO2022113314A1 (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/JP2020/044396 WO2022113314A1 (en) 2020-11-27 2020-11-27 Learning method, learning program, and learning device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/JP2020/044396 WO2022113314A1 (en) 2020-11-27 2020-11-27 Learning method, learning program, and learning device

Publications (1)

Publication Number Publication Date
WO2022113314A1 true WO2022113314A1 (en) 2022-06-02

Family

ID=81755450

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2020/044396 WO2022113314A1 (en) 2020-11-27 2020-11-27 Learning method, learning program, and learning device

Country Status (1)

Country Link
WO (1) WO2022113314A1 (en)

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2017119060A1 (en) * 2016-01-05 2017-07-13 株式会社日立製作所 Information presentation system
JP2020106905A (en) * 2018-12-26 2020-07-09 日本電信電話株式会社 Speech sentence generation model learning device, speech sentence collection device, speech sentence generation model learning method, speech sentence collection method, and program

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2017119060A1 (en) * 2016-01-05 2017-07-13 株式会社日立製作所 Information presentation system
JP2020106905A (en) * 2018-12-26 2020-07-09 日本電信電話株式会社 Speech sentence generation model learning device, speech sentence collection device, speech sentence generation model learning method, speech sentence collection method, and program

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
FURUMAI, KAZUAKI ET AL.: "Examination of an automatic generation method of supporting and opposing opinion using language model in debate system", LECTURE PROCEEDINGS CD-ROM [ CD-ROM ] OF THE 2019 SPRING RESEARCH CONFERENCE OF THE ACOUSTICAL SOCIETY OF JAPAN, 19 February 2019 (2019-02-19), pages 957 - 960 *

Similar Documents

Publication Publication Date Title
Yu et al. Strategy and policy learning for non-task-oriented conversational systems
US10936664B2 (en) Dialogue system and computer program therefor
CN110457432B (en) Interview scoring method, interview scoring device, interview scoring equipment and interview scoring storage medium
Qaffas Improvement of Chatbots semantics using wit. ai and word sequence kernel: Education Chatbot as a case study
CN111694940B (en) User report generation method and terminal equipment
US11636272B2 (en) Hybrid natural language understanding
CN110335608B (en) Voiceprint verification method, voiceprint verification device, voiceprint verification equipment and storage medium
Su et al. Deep learning for conversational AI
CN110019691A (en) Conversation message treating method and apparatus
CN112667791A (en) Latent event prediction method, device, equipment and storage medium
JP2022183122A (en) Computer-implemented method, computer system and computer program (cognitive learning for generating scripts that simulate live-agent actions in synchronous conferencing)
Guo et al. Fast sequence generation with multi-agent reinforcement learning
WO2022113314A1 (en) Learning method, learning program, and learning device
JP2016045769A (en) Dialog system evaluation method, dialog system evaluation device, and program
Wang et al. Emily: Developing An Emotion-affective Open-Domain Chatbot with Knowledge Graph-based Persona
WO2023129352A1 (en) Using token level context to generate ssml tags
JP7111017B2 (en) Paralinguistic information estimation model learning device, paralinguistic information estimation device, and program
Ohta et al. Response type selection for chat-like spoken dialog systems based on LSTM and multi-task learning
Dharmatti et al. Interview Practice-Voice-Based Chatbot
Takahashi et al. Natural Language Dialog System Considering Speaker’s Emotion Calculated from Acoustic Features
Abdrabuh AI-Synthesized Speech: Generation and Detection
WO2023119675A1 (en) Estimation method, estimation device, and estimation program
Mannekote et al. Agreement Tracking for Multi-Issue Negotiation Dialogues
Rahimi Linguistic Entrainment in Multi-Party Spoken Dialogues
WO2023234128A1 (en) Conversation management device, conversation management method, and conversation management system

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20963578

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20963578

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: JP