JP2024009724A - Dialogue reply candidate proposal system and dialogue reply candidate proposal method - Google Patents

Abstract

To provide a dialogue reply candidate proposal system for generating an appropriate reply candidate sentence for a new question sentence in a dialogue.SOLUTION: A dialogue reply candidate proposal system comprises a processor and a sub-storage device 33. The sub-storage device 33 stores a question-reply database 21 in which past question sentences and past replay sentences for the past question sentences are associated with each other and a learning question-replay database 22 in which the sentences in the question replay database 21 have been coupled with sentences in dialogues time series information. When a new question sentence is input from a questioner, the processor calculates a similarity with question information for each past question sentence stored in the learning question-reply database 22, further calculates a summed similarity by summing similarities in a dialogue, extracts a past question sentence similar to the new question sentence on the basis of each similarity, and presents it as a reply candidate sentence.SELECTED DRAWING: Figure 2

Description

The present invention relates to a dialog answer candidate proposing system and a dialog answer candidate proposing method that generate answer candidate sentences that are answer text candidates for new question sentences in a dialog.

BACKGROUND ART In order to prepare answer texts for new question texts from users obtained via the Internet or public communication networks, there is a technique for generating answer candidate texts that serve as candidates. For example, in Patent Document 1, a question string based on a user's question is decomposed into a plurality of morphemes (words), and based on the obtained plurality of morphemes, an answer string is created from a group of accumulated past answer strings. A technique is disclosed for selecting and outputting.

Patent Document 2 discloses that by performing syntactic, semantic, and thematic analysis on a prepared FAQ, even if the content of the customer inquiry is incomplete, it can be guided through dialogue based on natural feedback. A technology has been disclosed that interprets the intent and meaning of an inquiry, complements and narrows down the content, and presents an answer. Furthermore, in Patent Document 3, each past dialog sentence that has been accumulated is labeled with "question, information provision, etc.", and based on the labels, each word by the questioner as a dialog answer candidate included in the dialog data is labeled. For each word by a dialog answer candidate answer, it is estimated whether the word is an expression of a question, and a dialog answer candidate question is generated as a response to the dialog answer candidate answer's word that is estimated to be an expression of a question. A technique has been disclosed for reconstructing question sentences and answer sentences by extracting words uttered by a person from dialog answer candidate dialog data.

JP 2018-181033 Publication JP 2019-3319 Publication WO2021/234773 publication

In the question answering support device of Patent Document 1, the user's question text is insufficient in terms of content, and dialogue is required such as asking additional questions to obtain complementary information and presenting options to narrow down the content. Insufficient assumptions about the case. Specifically, if the technique of Patent Document 1 is applied to an input sentence as shown in FIG. 14, each of the question sentences Q1 to Q4 will be learned independently. In other words, in response to a question similar to Q2, ``It's in.'', students will learn to suggest an answer such as ``Is the antenna picto standing?'' This can be said to be learning that does not take into account the dialogues such as ``I can't connect to my cell phone...'' and ``Is it turned on?'' that led to the answer ``?''. If you train each sentence without including its surrounding context, the model may end up proposing an answer such as "Is the antenna pictograph standing?" in response to a question with a completely unrelated context. There will be negative effects.

The technology described in Patent Document 2 is a technology that assumes multiple dialogues, and generates appropriate feedback sentences from Frequently Asked Questions (FAQ), so appropriate exchanges can be expected. However, there is a problem in that the FAQ needs to be generated manually, and the generation cost is high. The same applies to well-known rule-based algorithms (methods in which rules for responding are manually determined in advance) that are often employed in chatbots and the like.

To solve this problem, the technology described in Patent Document 3 combines a series of sentences written by the user into one sentence from past dialogues and recombines a series of sentences written by respondents into one sentence. Efforts are being made to reduce production costs. For example, when Patent Document 3 is applied to an input sentence as shown in FIG. 14, question sentences and answer sentences as shown in FIG. 15 are extracted. However, as you can see from "It's in." and "Three bars are standing." in this example, the user only responds to the minimum necessary during dialogue, excluding important words such as the subject. Since this is common, there is a risk that the subject in Figure 15 may be omitted from the question text that the AI is trained to learn, which could have a negative impact on the model.

SUMMARY OF THE INVENTION Therefore, an object of the present invention is to provide a dialog answer candidate proposing system and a dialog answer candidate proposing method that output suitable answer candidate sentences for new question sentences in a dialog at a low cost.

Representative features of the invention disclosed in this application are as follows.
According to one feature of the present invention, the present invention includes a processor, a storage device, and a transmitting/receiving device that communicates with a user terminal used by a questioner and an operator terminal used by a questioner, and the transmitting/receiving device communicates with a user terminal used by a questioner and an operator terminal used by a questioner. In an interactive answer candidate proposal system that generates answer candidate sentences that serve as sentence candidates and displays them on the operator terminal, past question sentences input by the questioner and past answer sentences for the past question sentences are A question-and-answer database in which dialogue time-series information is associated and recorded, and a learning question-and-answer database in which learning question sentences and learning answer sentences generated based on the information stored in the question and answer database are recorded. , an answer candidate sentence generation unit that extracts past question sentences or learning question sentences similar to the new question sentence from the question and answer database or the learning question and answer database when a new question sentence is input by the questioner; A learning question and answer data generation unit is provided that generates learning question and answer data using past question sentences similar to the new question sentence or past answer sentences associated with the learning question sentence. The learning question and answer data generation unit calculates the degree of similarity between each of the past question sentences and the learning question sentences stored in the question and answer database and the learning question and answer database with the new question sentences, and calculates the degree of similarity between each dialogue. By calculating the total similarity, learning question and answer data is generated based on the similarity with the new question sentence or the total similarity of each dialogue.

According to another feature of the present invention, the storage device includes a question and answer database, a learning question and answer database, an interrogative word table that stores interrogative words expressing questions, and a question word table that stores interrogative words expressing questions, and a A synonym table that stores words that correspond to the same term, and a negative/affirmative word table are stored. The learning question and answer database is generated based on rules that use text analysis to extract the words stored in the synonym table from past question texts and answer texts, and generate the text analysis results in advance. be. When a new question sentence is input from the questioner, the answer candidate sentence generation unit searches the learning question and answer database and displays similar questions to the new question sentence and answer candidates for the similar questions on the operator terminal. This will assist in answering the questioner's dialogue. The construction of the learning question and answer database is executed in batch processing by the learning question and answer data generation program.

According to the present invention, past question sentences similar to question information are extracted from a question and answer database, and past question sentences or question sentence groups similar to the extracted question information and past answer sentences or A learning question-and-answer database is constructed by extracting answer sentences from the question-and-answer database and creating answer candidate sentences, so when responding to a question, the operator can use the learning question-and-answer database to provide a proposed answer to the questioner. can be obtained, and suitable answer candidate sentences for new question sentences can be efficiently provided to the questioner.

1 is a block diagram showing a hardware configuration of a dialog answer candidate proposal system 1 according to an embodiment of the present invention. FIG. 2 is a data configuration diagram showing an example of contents stored in the secondary storage device 33 of FIG. 1. FIG. 3 is a diagram showing an example of data stored in the question and answer database 21 of FIG. 2. FIG. 3 is a diagram showing an example of data stored in the learning question and answer database 22 of FIG. 2. FIG. 3 is a diagram showing an example of data stored in the question word table 23 of FIG. 2. FIG. 3 is a diagram showing an example of data stored in the synonym table 24 of FIG. 2. FIG. 3 is a diagram showing an example of data stored in the affirmative/negative word table 25 of FIG. 2. FIG. It is a flowchart (1) which shows the generation procedure of the learning question data in a present Example. It is a flowchart which shows the generation procedure of the learning question data in a present Example (Part 2). FIG. 3 is a diagram showing an example of preprocessing and combination in the generation of learning question data according to the present embodiment. 3 is a diagram showing an example of a question and answer display screen 1200 displayed on the operator terminal 3. FIG. 2 is a flowchart showing the procedure of an answer candidate sentence generation process of the dialog answer candidate suggestion system 1. FIG. 14 is an explanatory diagram showing an example of an answer candidate sentence display screen 1400 displayed on the operator terminal 3. FIG. FIG. 2 is a diagram showing an example of an input sentence for explaining the problems of the prior art and the features of the present invention. It is a figure which shows the example of a question text and an answer text for explaining the problem of a prior art.

Embodiments of the present invention will be described below with reference to the drawings. The present invention can also be implemented in various other forms. Unless specifically limited, each component may be singular or plural. Although the expression "table" is used in the description as an example of various types of information, various types of information may be expressed using data structures other than these. For example, it may be "XX list", "XX queue", or "XX information". When describing identification information, expressions such as "identification information", "identifier", "name", "ID", and "number" are used, but these expressions can be replaced with each other.

A program may be installed on a computer from a program source. The program source may be, for example, a program distribution server or a computer-readable storage medium. When the program source is a program distribution server, the program distribution server includes a processor and a storage resource for storing the program to be distributed, and the processor of the program distribution server may distribute the program to be distributed to other computers. Furthermore, in the embodiments, two or more programs may be implemented as one program, or one program may be implemented as two or more programs.

<System configuration>
FIG. 1 is a block diagram showing the system configuration of a dialogue answer candidate proposal system 1 according to an embodiment of the present invention. The dialog answer candidate proposal system 1 can be realized by a computer that executes a program and performs processing, and includes a processor 31, storage resources (main storage 32, secondary storage 33), input device 34, and output device 35. These are interconnected by a data bus 37. The data bus 7 of the dialogue answer candidate proposal system 1 is further provided with a network interface 36 for connecting to the network 5.

The processor 31 implements specific functions by reading data and programs stored in the secondary storage device 33 into the main storage device 32 and executing them. The processor 31 is, for example, a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit), and the type, operating frequency, number of cores, and number of processors used are arbitrary. Note that the subject of processing may be not only the processor 31 but also a controller, device, system, computer, or node having a processor. The main body of processing performed by executing the program may be an arithmetic unit, and may include a dedicated circuit that performs specific processing. Here, the dedicated circuits include, for example, FPGA (Field Programmable Gate Array), ASIC (Application Specific Integrated Circuit), and CPLD (Complex Programmable Logic D). evice) etc.

The processor 31 executes various programs installed in the main storage device 32 or the secondary storage device 33. Here, the function of the learning question and answer data generation unit 11 is realized by executing the learning question and answer data generation program 26 (see FIG. 2), and the answer candidate sentence generation program 27 (see FIG. 2) is executed. By doing this, the functions of the answer candidate sentence generation section 12 are realized by software. Further, the functions of the question and answer execution unit 13, such as providing a chat environment, are realized by software using the question and answer execution program 28 (see FIG. 2). Although the functions realized by the processor 31 are not limited to these three, illustrations and explanations of other functions are omitted.

As the network 5, a known global network such as the Internet can be used, for example. However, the network 5 to be used is not limited to the Internet, and may be a public communication network provided by a communication carrier, a local area network (LAN), or other known networks. The network interface (I/F) 36 is a device that enables the dialogue answer candidate proposal system 1 and the network 5 to be connected. Data is transmitted and received between each user terminal 2 owned by the user and the operator terminal 3 used by a plurality of operators (respondents) working at a help desk or the like.

The dialog answer candidate proposal system 1 is further provided with an input device 34 that accepts operations such as a mouse or a keyboard, and an output device 35 that outputs information such as a display device. operably configured by. Note that it is also possible to configure one of the operator terminals 3 to serve both the functions of the input device 34 and the output device 35.

The main storage device 32 has a volatile storage element such as a RAM, and temporarily stores programs executed by the processor 31 and data. The secondary storage device 33 is a device that has a non-volatile storage element such as an HDD (Hard Disk Drive) or an SSD (Solid State Drive), and stores programs, data, and the like. The configuration (storage resources) of the secondary storage device 33 is not limited to the above-mentioned configuration, and external storage devices, network servers, cloud servers, etc. may be used in addition to or in place of the secondary storage device 33. It may also be configured to store data, programs, etc. in a distributed manner.

The user terminal 2 is an information device operated by a user who asks a question, and a question sentence for the dialogue answer candidate proposal system 1 is input from the user terminal 2. The user terminal 2 is an electronic device that includes at least an input device (not shown) that receives input from a user, an output device (not shown) that displays information such as a display or touch panel, and a network interface (not shown). For example, a known information terminal such as a personal computer, tablet, smartphone, or mobile phone can be used. The user terminal 2 is capable of bidirectional communication with the dialog answer candidate proposal system 1 and the operator terminal 3 via the network 5 .

The operator terminal 3 is a device used by an operator who interacts with a user who operates the user terminal 2, and is installed at a place where the operator works, such as a help desk. The operator terminal 3 includes at least an input device (not shown) that receives input operations from an operator, an output device (not shown) that displays information such as a display or a touch panel, and a network interface (not shown). . As the operator terminal 3, a known electronic device such as a personal computer, a tablet, or a smartphone can be used. Note that some of the data stored in the sub-storage device 33 inside the operator terminal 3, for example, copies of the question-and-answer database 21 and the learning question-and-answer database 22, are copied to the internal storage device in advance so that the user can The question answering process may be performed by the processor of the operator terminal. In this way, some of the functions executed by the processor 31 of the dialogue answer candidate proposal system 1 may be configured to be executed by the operator terminal 3 to speed up the processing.

When a new question text input by the user is transmitted from the user terminal 2 to the dialogue answer candidate proposal system 1, the processor 31 selects past sentences that are candidates for answer sentences to the new question text based on the received new question text. A question text or question text group and past answer texts or answer text groups associated therewith are generated and displayed on the output device of the operator terminal 3 via the network 5. The operator provides an appropriate answer to the user who asked the question while looking at various information including the dialog answer candidates provided by the dialog answer candidate suggestion system 1. The answer to the question text determined by the operator is transmitted to the user terminal 2 via the dialog answer candidate proposal system 1. In other words, the processor 31 systemically mediates the exchange of information between the user who asks the question and the operator in charge of answering the question, and also uses past question texts or question text groups related to the questions asked by the user. , a group of past answer texts or question texts associated with it is generated and displayed on the operator terminal 3, thereby supporting the answer work by the operator. Since the operator can create an answer to the user and send it back to the user while referring to various information displayed on the operator terminal 3, the operator can quickly and accurately answer the answer.

The learning question and answer data generation unit 11 realizes a function of generating learning question and answer data using past questions and answer texts stored in a question and answer database 21 (described later in FIG. 2). This function is realized by executing a computer program, the details of which will be described later with reference to FIG. When a new question text is input from the user, the answer candidate text generation unit 12 generates an optimal answer text while referring to past question texts or question text groups and past answer texts or question text groups that are associated with the past question texts or question text groups. Generate answer candidates. The question-and-answer execution unit 13 provides a series of question-and-answer exchange environments in which questions are received from the user, sent to the operator terminal 3, and answers input from the operator terminal 3 are returned to the user terminal 2. do. The question-and-answer execution unit 13 uses known methods such as a chatbot, a question-and-answer using e-mail, a question-and-answer format answering environment via a chat screen, and a dedicated question-and-answering environment using a web screen. These can be realized by the processor 31 executing a computer program (question answering execution program 28 described later in FIG. 2). The question and answer execution unit 13 transmits the answer sentence candidates generated by the answer candidate sentence generation unit 12 to the operator terminal 3 via the network 5.

FIG. 2 is a data configuration diagram showing an example of the contents stored in the secondary storage device 33 of FIG. 1. A question and answer database 21 and a learning question and answer database 22 are constructed in the secondary storage device 33. Also stored are an interrogative word table 23, a synonym table 24, and an affirmation/negative word table 25, which are used when constructing or using these databases. Furthermore, a learning question-and-answer program 26, an answer candidate sentence generation program 27, and a question-and-answer execution program 28 executed by the processor 31 shown in FIG. 1 are stored. Although various other programs, data, etc. are stored in the secondary storage device 33, a description thereof will be omitted.

The question and answer database 21 is a data group in which past question texts, past answer texts to past question texts, and past question text vectors are stored in association with each other (details will be described later with reference to FIG. 3). The learning question and answer database 22 is a database that stores a data group generated by the learning question and answer data generation unit 11 using sentences in the question and answer database 21 (details will be described later with reference to FIG. 4).

The interrogative word table 23 is a database that stores interrogative words expressing a question and request words expressing a request (details will be described later with reference to FIG. 5). The synonym table 24 is a database that stores synonyms (details will be described later with reference to FIG. 6). The affirmative/negative word table 25 is a database stored to determine whether a sentence is an affirmative sentence or a negative sentence (details will be described later with reference to FIG. 7).

A learning question and answer data generation program 26, an answer candidate sentence generation program 27, and a question and answer execution program 28 are installed. The learning question and answer data generation unit 11 shown in FIG. 1 is realized by the processor 31 reading and executing the learning question and answer data generation program 26 stored in the secondary storage device 33. Further, the answer candidate sentence generation section 12 shown in FIG. 1 is realized by the processor 31 reading out and executing the answer candidate sentence generation program 27 stored in the sub-storage device 33. The question and answer execution program 28 is a program that realizes an environment in which a questioner can exchange questions and answers using the user terminal 2 and the operator terminal 3.

<Various data structures>
FIG. 3 is a diagram showing an example of the question and answer database 21. In the question and answer database 21, a question and answer ID 501 is an ID that identifies a plurality of stored past question texts 502. The past answer text 503 is a text that was answered to the past question text 502, and the Next ID 504 (details will be described later) is the question answer ID 501 of the next question following the past question text 502. In the same record 511, a sentence vector 505, a user evaluation 506, and a training data generation progress 507 are further stored in association with the past question sentence 502.

In this way, the question and answer database 21 stores past question texts 502 and past answer texts 503 for the past question texts 502 in association with each other. NextID 504 is a number for saving the connection of dialogues; "0" means that the dialogue has ended, and "- (NULL)" means that the answer has been completed in one exchange. By using the question and answer ID 501 and assigning the Next ID 504, "sequence information at the time of dialogue" can be confirmed. The sentence vector 505 is generated for all past question sentences stored in the question and answer database 21 every time a new pair of past question sentences and past answer sentences is stored in the question and answer database 21. You may. Alternatively, the text vector 505 may be regenerated at a preset timing, such as every time a predetermined number of pairs of past question texts and past answer texts are stored in the question and answer database 21. User evaluation 506 is saved to associate the user's evaluation of the answer with the interaction result, and "-(NULL)" means "no user evaluation." The learning data generation progress 507 is saved to determine how far the learning question and answer data generation program has been executed. The program is not subject to training data generation, and "x" means that data generation by the program has not been completed.

FIG. 4 is a diagram showing an example of the learning question and answer database 22. The learning question and answer database 22 shown in FIG. 4 stores learning question sentences 602 and learning answer sentences 603 generated by executing the learning question and answer data generation program 26 (see FIG. 2). The question answer No. 601 is an ID for identifying the learning question text 602. The creation source ID 604 is the ID 501 (see FIG. 3) of the past question text 502 (see FIG. 3) from which the study question text 602 was constructed. The sentence vector 605 is a numerical representation of a natural language sentence. A known method can be used to calculate the sentence vector 605, and the characteristics of the sentence are expressed using a set of numbers. The user evaluation 606 stores the user's evaluation of the learning answer text 603 with respect to the learning question text 602, and is the evaluation input by the user at the end of the question. Here, it is input that the user who asked the question included in the learning question text 602 as a result of the learning answer text 603 evaluated the final answer of the learning answer text 603 as appropriate (“Good”). . In this manner, the learning question and answer database 22 stores the learning question text 602 and the learning answer text 603 for the learning question text 602 in association with each other.

FIG. 5 is a diagram showing an example of the interrogative word table 23. In the interrogative word table 23 shown in FIG. 5, interrogative words 702 used to analyze a question from a user are stored in advance in a list format. The interrogative word ID 701 is an ID for identifying the stored interrogative word 702. The interrogative word 702 stores interrogative words expressing doubts. The interrogative word type 703 indicates which type of interrogative word each of the interrogative words 702 is. In FIG. 5, records of "WHEN" and "WHY" (interrogative word ID="1", "2") are shown as examples of the interrogative word 702 and the type 703. Other examples of interrogative word 702: type 703 include "WHO", "WHAT", "WHERE", "HOW", "WHEN", "WHICH", and "Who". Examples include ": WHO", "WHAT", "HOW", "HOW", "YES/NO", "YES/NO".

FIG. 6 is a diagram showing an example of the synonym table 24. In the synonym table 24 shown in FIG. 6, a synonym ID 801 is an ID for identifying a synonym 802. The synonym 802 is a word representing a stored synonym pair. FIG. 6 shows "smartphone" and its synonym "smartphone" as an example of the synonym 802. In this way, in the text analysis process in the dialog answer candidate proposal system 1, when the user inputs "smartphone", it is treated as a term synonymous with "smartphone". A large number of synonyms 802 are stored in the synonym table 24.

FIG. 7 is a diagram showing an example of the affirmative/negative word table 25. In the affirmative/negative word table 25 shown in FIG. 7, a list of positive/negative words 902 is stored. The affirmative/negative word ID 901 is an ID for identifying each positive/negative word 902. The affirmative/negative words 902 are a list of words that are considered to often have important meanings regarding the intention contained in the sentence. The affirmative/negative word type 903 is information indicating whether the positive/negative word 902 means affirmative or negative. In FIG. 7, "yes" and "no" are stored as examples of affirmative/negative words 902, and their types 902 indicate "yes: affirmative" and "no: negative." Note that a large number of data other than the example shown in FIG. 7 are stored in the affirmative/negative word 902 and the affirmative/negative word type 903.

<Similar sentence extraction method>
In the process of generating answer candidate sentences based on a new question sentence input by the user, the dialog answer candidate proposal system 1 vectorizes the sentence and generates a model using a known technique such as the tfidf method (word frequency inverse sentence frequency method). Then, the cosine similarity between the vectors is calculated. Below, an outline of a text vectorization method, a model generation method, and a similar text extraction method that calculates the cosine similarity between texts and extracts similar texts similar to the target text will be explained below. In text vectorization and calculation of cosine similarity, a database in which a plurality of texts are stored (in this embodiment, the question and answer database 21 and the learning question and answer database 22) is used.

In the similar sentence extraction method, which calculates the cosine similarity between sentences and extracts similar sentences that are similar to the target sentence, as explained below, each sentence stored in the database and the target sentence are Then, a vector is calculated using the tfidf method or the like, and a cosine similarity is calculated.

When using the tfidf method, first, all sentences stored in the database and the target sentence are morphologically analyzed, and the sentences are broken down into words (morphemes). Next, duplicate words are deleted from the plurality of words obtained by decomposition, and a word vector having each word as a component is generated. Next, a tfidf vector is calculated for each sentence stored in the database and the target sentence. The tfidf vector is a vector whose components are tfidf values for words that are components of the word vector.

To give an example of a word vector and a tfidf vector, a word vector generated by morphologically analyzing the sentence "Smartphones are light." is, for example, (smartphone, wa, light, .). The tfidf vector for this is, for example, (tfidf value of "smartphone", tfidf value of "ha", tfidf value of "light", tfidf value of ".").

Next, the cosine similarity (cosine value for the angle between the two tfidf vectors) between each of the tfidf vectors of the sentences stored in the database and the tfidf vector of the target sentence is calculated. The cosine similarity of the two tfidf vectors A and B is cosine similarity=A·B/(|A||B|). A sentence with a larger value of cosine similarity with the target sentence (a sentence with a higher cosine similarity) is considered to have a higher degree of similarity.

Among the sentences included in the database, a predetermined percentage (or a predetermined number) of sentences from the top in terms of cosine similarity are determined to be similar sentences with high similarity. Here, instead of the cosine similarity, the inner product of the tfidf vector of the text stored in the database and the tfidf vector of the target text may be used.

The above explanation is an outline of the tfidf method, and the algorithm of the tfidf method when using the tfidf method can be changed as appropriate from the method explained above. In addition, instead of the tfidf method, which is a similar sentence extraction method that calculates the cosine similarity between sentences and extracts similar sentences that are similar to the target sentence, a method such as the BERT (Bidirectional Encoder Representations from Transformers) method can be used. Similar sentences may be extracted using other methods of calculating sentence similarity. Unlike the tfidf method, the BERT method generates a model that can understand the context, and therefore is more suitable for a program that takes dialogue into account in the present invention. As an example of the use of the BERT method, first, a small amount of data is labeled as to whether the question sentences in the held database are synonymous or not. Next, there is a method of generating a model for determining synonymous sentences by performing fine tuning using a known Japanese Wikipedia pre-trained model and prepared label data.

<Processing procedure>
Next, the processing procedure of the dialog answer candidate suggestion system 1 will be explained. 8 and 9 are flowcharts showing an example of the procedure of the learning question and answer data generation process of the dialog answer candidate proposal system 1. The learning question and answer data generation process in this system is executed by the system administrator or the dialog answer candidate proposal system administrator as a batch process, and the learning question and answer data generation process is performed at regular or irregular intervals by the system administrator or the dialog answer candidate proposal system administrator. The question and answer data generation program 26 (see FIG. 2) is activated, and the processing shown in FIGS. 8 and 9 is started.

When execution of the learning question and answer data generation program 26 (see FIG. 2) is specified, the processor 31 of the dialog answer candidate proposal system 1 reads the learning question and answer data generation program 26 (see FIG. 2) from the secondary storage device 33. Read the file and start executing the program. In FIG. 8, a processor 31 performs variable initialization work when executing a program. Here, a variable i for counting ID is cleared to zero (step S101), and then the variable i is counted up by 1 (step S102).

Next, the processor 31 checks whether the data of the ID stored in the variable i of the question and answer database 21 has been confirmed by this program by checking whether the learning data generation progress is "x" (step S103). . If it is determined that the information has not been confirmed (step S103: NO), the process advances to step S104, and if it is determined that the information has been confirmed (step S103: YES), the process returns to step S102. As a result, the processor 31 repeats the processing of steps S102 and S103 until unconfirmed data is found in this program in order from ID=1.

Next, the processor 31 analyzes the question text and answer text of data equal to or higher than the ID corresponding to the variable i according to preprocessing A (step S104). Preprocessing A refers to preprocessing for combining question data and answer data, and performs analysis necessary for combining sentences. Each analysis method of preprocessing A may be a known technique. For example, syntactic analysis, semantic analysis, thematic analysis, etc. described in Patent Document 2 may be used.

FIG. 10 is an explanatory diagram showing an example of the learning question and answer data generation process. First, as shown in the data column 1000, pairs of question sentences and answer sentences to be combined are read out from the question and answer database 21. Next, these question sentences and answer sentences are analyzed into the form shown in the analysis result column 1002 by performing syntactic analysis, semantic analysis, thematic analysis, etc. in preprocessing A1101. After this, rule A1103 is applied to the content shown in the analysis result column 1102 to determine whether or not to perform the combination. Furthermore, a rule B 1104 for performing the combination is applied to the objects determined to be combined according to the rule A. Table 1105 shows an example of the results combined by rule B1104.

In the syntactic analysis, the question sentences and answer sentences are analyzed using morphological analysis, etc., and the sentences are analyzed into sentences such as ``Oshiru ⇒ verb'' and ``ga ⇒ particle.'' In semantic analysis, based on the results of syntactic analysis, the semantic attributes are grammatically analyzed and the meaning is analyzed, such as ``Shika ⇒ Auxiliary verb: question.'' At this time, by using the interrogative word table 23, it is possible to predict the unknown part necessary for the combination, such as "which⇒joint word: unknown part:WHAT". Furthermore, by using the synonym table 24, it is possible to recognize words that are synonymous during analysis, such as "smartphone⇒smartphone". Furthermore, by using the affirmative/negative word table 24, it is possible to predict whether the answer to YES/NO is affirmative or negative, such as "Yes⇒affirmative". Thematic analysis is defined in Patent Document 2 and is a means of identifying important parts in a sentence, and the sentence is analyzed such as "mobile phone ⇒ noun MAIN: WHAT" and "use ⇒ verb MAIN: HOW."

Returning to FIG. 8 again, in step S105, the processor 31 determines whether the data of the ID corresponding to the variable i in the question and answer database 21 is a multiple interaction or a single interaction, and the Next ID is "-". Check depending on whether If it is determined that the process is completed with one exchange (step S105: YES), a "-" indicating that the learning data is not to be generated is stored in the learning data generation progress of the data of the ID corresponding to the variable i, After incrementing the variable i, the process returns to step S104 (step S106). If it is determined in step S105 that there is a plurality of dialogues (step S105: NO), the process advances to step S107. As a result, the processor 31 repeats the processing of steps S105 and S106 until it finds dialogue data for a plurality of times.

In step S107, the processor 31 checks whether the dialogue text from the ID corresponding to the variable i in the question and answer database 21 to the ID (corresponding Next ID is 0) until the end of the dialogue applies to rule A. If it is determined that this applies (step S107: YES), the process advances to step S108 (after step S108, the process returns to step S104). In step S108, "-" is stored in the learning data generation progress 507 (see FIG. 3) for ID=i, and the variable i is updated to the next ID indicating the end of the dialogue (Next ID 504 in FIG. 3 is "0"). If it is determined in step S107 that this does not apply, the process advances to step S109. As a result, the processor 31 repeats the processing of steps S105 and S107 until a dialogue sentence that applies to rule A is found. Note that rule A is a rule for determining whether a series of dialogues is suitable for being combined as learning question and answer data, as shown in FIG. 10. This is applied to the results of preprocessing A1101. Examples of rule A are: ``There are two or more unknown parts in one answer by the operator,'' ``One answer by the operator consists of two or more sentences,'' and ``Unknown parts in the operator's answer. Even though the question element is not included (it is an answer, not a question), the interaction with the user continues (unresolved).''

Proceeding to FIG. 9, the processor 31 analyzes the results of the analysis by preprocessing A on the question text from the ID corresponding to the variable i to the ID (corresponding Next ID is 0) until the end of the dialogue from the question and answer database 21. It is checked whether there is a sentence that includes "unknown part" or "question" (step S109). If it is determined that such a sentence exists (step S109: YES), a "-" indicating that the learning data is not subject to generation is added to the learning data generation progress of the ID data corresponding to the timing of the question. The ID corresponding to the timing of the question is stored (step 110). If it is determined in step S109 that there is no sentence that includes "unknown part" or "question" (step S109: NO), the process advances to step S111. By processing S109 and S110, it is possible to branch when the user asks a question at a timing other than the start of the dialogue.

Next, the processor 31 combines the answer sentence with the ID corresponding to the variable i and the question sentence with the Next ID of that ID according to rule B (step S111). Rule B is a rule for combining a series of dialogues as learning question and answer data, and is applied to the results of preprocessing A1101 as shown in FIG. In FIG. 10, the analyzed question text and answer text shown in the analysis result column 1002 are combined using rule B in the form shown in table 1105. For example, if the ``HOW type'' and ``Unknown part type'' in the analysis result are ``verb'' and ``WHAT,'' respectively, as per the ``combination rule'' column, ``Mobile phones use...'' Combine like this. In this way, by defining each type of analysis result as a rule, it becomes possible to combine sentences with an appropriate syntax.

Next, the processor 31 updates the variable i to the combined ID (step S112). Next, it is checked whether the NextID 504 of the variable i is "0", which means the end of the dialogue (step S113). If it is determined that the dialogue has ended (step S113: YES), the process advances to step S114, and if it is determined that the dialogue has not ended (step S113: NO), the process returns to step S111. As a result, the processes of steps S111 and S112 are repeated until the series of dialogues is completed.

Next, the processor 31 stores the question sentence at the beginning of the dialogue and the combined sentences generated in steps S111 to S113 as learning question data, and the answer data at the end of the dialogue as learning answer data, in the learning question and answer database 22. (Step S112). Next, the processor 31 checks whether an ID equal to the next variable i+1 exists (step S115). If it exists (step S115: YES), the process returns to step S105, and if it is determined that it does not exist (step S113: NO), the process shown in FIGS. 8 and 9 ends. As a result, it is possible to perform the process of creating a learning question and answer sentence for data that has not been subjected to the data generation process. In this embodiment, the processing in FIGS. 9 and 10 is performed as a batch process, and is performed at a timing that is not linked to the input operation from the questioner. The configuration may be such that the processing in FIGS. 9 and 10 is executed in real time when the input operation and the answer by the operator are completed.

FIG. 11 is an explanatory diagram showing an example of a question and answer display screen displayed on the operator terminal 3. A question and answer display screen 1200 shown in FIG. 11 displays a dialog past question text display field 1201, an answer text display field 1202, a new question text display field 1203, a send button 1204, and an answer button 1305. The dialog past question text display column 1201 is a column that displays question texts in a series of dialogs. The answer text display column 1102 is a column for displaying the answer text given by the operator. The new question text display field 1203 is a field for displaying new question texts. Here, an example is shown in which only three display columns for question sentences and answer sentences are displayed, but as the number of sentences increases, many corresponding display columns will be displayed (on the display screen 1200). (If they cannot be displayed simultaneously, they can be scrolled.)

When the operator operating the operator terminal 3 presses (clicks, etc.) inside the dialog past question text display field 1201 or the new question text display field 1203, the operator terminal 3 accepts input from the operator, The operator can edit the question text in the dialog past question text display field 1201 or the new question text display field 1203. Through this editing, the operator terminal 3 can transmit a new question text that has been edited by the operator, such as correcting typographical errors, to the dialog answer candidate proposal system 1. As a result, the processor 31 regards the edited question text as a new question text and generates an answer candidate text. Thereby, the processor 31 can generate more suitable answer candidate sentences.

When the operator presses the send button 1204, the operator terminal 3 sends the dialogue answer candidate suggestion system 1 the question text in the dialog past question text display field 1201 and the new question text in the new question text display field 1203. Generation start information including answer candidate sentence generation start information, which is information instructing the dialog answer candidate proposal system 1 to generate an answer candidate sentence, is transmitted. Note that the send button 1204 is displayed in the form of an icon, and "pressing" here means selecting the icon. If the operator terminal 3 has a touch-type input panel, the "send button 1204" is displayed in the form of an icon. ``is pressed'' is synonymous with ``the icon is touched''. The question text included in the generation start information is the question text in the past question text display field 1201 and the new question text display field 1203 in the dialogue at the time when the operator presses the send button 1204. Therefore, if the operator edits the question text in the dialogue past question text display field 1201 or the new question text display field 1203 before pressing the send button 1204, the edited question text will be used as the question text in the start information. include.

When the operator presses the answer button 1205, the screen displayed on the operator terminal 3 is configured to switch from a question and answer display screen to a screen where the operator inputs an answer to a new question. As an example of a screen for inputting an answer text to a new question text, a known input method may be used, so a description thereof will be omitted here.

When the processor 31 receives the generation start information from the operator terminal 3 through the network I/F 36 (thereby, a new question sentence included in the generation start information is input), the answer candidate sentence generation unit 12 executes the process. Start answer candidate sentence generation processing. The answer sentences generated by the answer candidate sentence generation process are sent by the question answer execution program 28 (see FIG. 2) to the user terminal 2 used by the questioner, and further questions from the questioner are accepted. .

Next, the answer candidate sentence generation process will be explained using FIG. 12. FIG. 12 is a flowchart showing an example of the answer candidate sentence generation process of the dialog answer candidate suggestion system 1. (See FIG. 2). First, the processor 31 stores the past question sentences in the dialogue and the new question sentences (step S201). Next, the processor 31 generates a sentence vector from a model generated by the similar sentence extraction method for the sentence that is the sum of the past question sentences in the dialogue and the new question sentences (step S202).

Next, the processor 31 uses the question and answer database 21 and the learning question and answer database 22 to calculate the degree of similarity using a similar sentence extraction method, and extracts past question sentences that are similar to the question information. past question texts,
Then, the past answer sentences associated therewith are saved as answer candidate sentences (step S203). Here, the sentences whose learning data generation progress in the question and answer database 21 is "〇" are sentences that have not been combined with the sentences stored in the study question and answer database 22, and are therefore excluded from similarity calculation. At this time, it is also possible to use the user evaluations stored in the database and perform processing to increase the degree of similarity if the evaluation is good.

Next, of the similarities calculated in step S203, the processor 31 calculates a total similarity for the dialogue sentences for which the learning data generation progress of the question and answer database 21 is "-" and "x" (step S203). S204). As a result, an index for the entire dialogue is calculated for dialogue data that is not combined as learning question and answer data. As an example of calculating the total similarity, there is a method of weighting sentences that include "unknown parts," "questions," and "requests" among the question sentences in the dialogue, and then adding up the similarities. This can be expected to increase the degree of similarity of similar questions in line with the user's intentions.

Next, the processor 31 uses the similarity calculated in steps S203 and S204 to extract a predetermined percentage (for example, 20%) or a predetermined number (for example, 3) of past question texts from the top based on the high similarity. do. Then, past answer texts associated with the extracted past question texts are extracted from the question and answer database 21, and the extracted past question texts and answer texts are saved as answer candidate texts (S205). Further, when the total similarity is high in similarity, all of the corresponding dialogues may be used as answer candidate sentences, or the question sentence with the highest similarity among them and the corresponding answer candidate sentences may be used. Further, when the similarity of the learning question data is high, the learning question data and the corresponding learning answer data may be used as the answer candidate sentences, or the original dialogue data may be used as the answer candidate sentences. Furthermore, in order to avoid displaying similar answer candidate sentences in an overlapping manner, the degree of similarity between answer sentences with the highest similarities may be calculated, and if the degree of similarity exceeds a predetermined degree, one of the answer sentences may not be displayed. Then, the answer candidate sentences and the answer candidate sentence display screen information are output to the output device (network I/F 36), and the answer candidate sentences and the answer candidate sentence display screen information are output to the output device (network I/F 36). have it sent to

The processor 31 waits for a predetermined time (step S206). Next, the processor 31 checks whether a question text has been input by the same user (step S207). If it is determined that the information has been input (step S207: YES), the process proceeds to step S201, and if it is determined that the information has not been input (step S207: NO), the process is completed. As a result, the processor 31 repeats the process and proposes answer candidate sentences as long as the dialogue continues.

FIG. 13 is an explanatory diagram showing an example of an answer candidate sentence display screen displayed on the operator terminal 3. The answer candidate sentence display screen 1400 shown in FIG. 12 includes answer candidate sentence (past question sentences) columns 1401, 1402, 1403, and 1404, and answer candidate sentence (past answer sentences) columns 1405, 1406, 1407, and 1408. . Furthermore, in this example, one dialogue sentence is displayed as an answer candidate sentence, but the display format is not limited to this, as described with reference to FIG. 12.

By displaying the answer candidate sentences on the answer candidate sentence display screen on the operator terminal 3, the operator can generate an answer sentence for the new question sentence by referring to the displayed answer candidate sentences. This allows the operator to more easily generate answer texts. In addition, the energy required for the operator to generate the answer texts and the amount of carbon dioxide emissions generated can be reduced, and global warming can be suppressed.

In this embodiment, the dialog answer candidate proposal system 1 generates answer candidate sentences based not only on new question sentences but also on past question sentences and new question sentences stored in the question and answer database 21 and the learning question and answer database 22. generate. As a result, the dialogue answer candidate proposal system 1 generates and outputs suitable answer candidate sentences that are more in line with the intent of the question in the new question sentence than when generating answer candidate sentences based only on the new question sentence. can.

Further, answer candidate sentences are generated using past question sentences and past answer sentences stored in the question and answer database 21 and the learning question and answer database 22. Thereby, the dialog answer candidate proposal system 1 can more easily generate answer candidate sentences.

In addition, the dialog answer candidate proposal system 1 outputs answer candidate sentences to the network I/F 36 (transmission/reception device), and sends the answer candidate sentences to the network I/F 36 (transmission/reception device) via the network 5 to the operator terminal. 3 to send. Thereby, the operator operating the operator terminal 3 can easily read the answer candidate sentences.

Although the present invention has been described above based on examples, the present invention is not limited to the above-mentioned examples, and various changes can be made without departing from the spirit thereof. For example, the embodiments described above have been described in detail to explain the present invention in an easy-to-understand manner, and the present invention is not necessarily limited to having all the configurations described. Furthermore, other configurations may be added, deleted, or replaced with some of the configurations of the embodiments.

1 Dialogue answer candidate suggestion system 2 User terminal 3 Operator terminal 5 Network 11 Study question and answer data generation section 12 Answer candidate sentence generation section 13 Question and answer execution section 21 Question and answer database 22 Study question and answer database 23 Interrogative word table 24 Synonyms Table 25 Affirmative and negative word table 26 Learning question and answer data generation program 27 Answer candidate sentence generation program 28 Question and answer execution program 31 Processor 32 Main storage 33 Secondary storage 34 Input device 35 Output device 36 Network interface 37 Data bus

Claims (14)

It has a processor, a storage device, and a transmitting/receiving device that communicates with a user terminal used by the questioner and an operator terminal used by the answerer, and it generates answer candidate sentences that are candidates for answer sentences to the question sentences from the questioner. A dialogue answer candidate suggestion system that generates and displays on the operator terminal,
a question-and-answer database in which past question texts input by the questioner are associated with past answer texts to the past question texts, and dialogue time series information is added and recorded;
a learning question and answer database that records learning question sentences and learning answer sentences generated based on information stored in the question and answer database;
an answer candidate sentence generation unit that extracts past question sentences or learning question sentences similar to the new question sentence from the question and answer database or the learning question and answer database when a new question sentence is input by the questioner;
a learning question and answer data generation unit that generates learning question and answer data using past question sentences similar to the extracted new question sentences or past answer sentences associated with the learning question sentences;
A dialog answer candidate suggestion system characterized by having the following. The learning question and answer data generation unit calculates the degree of similarity with the new question text for each of the past question text and the learning question text stored in the question and answer database and the learning question and answer database. According to claim 1, the learning question and answer data is generated based on the similarity with the new question text or the total similarity of each dialogue by calculating the total similarity of each dialogue. Dialogue answer candidate suggestion system described. storing in the storage device an interrogative word table that stores interrogative words expressing questions;
The learning question and answer data generation unit extracts the interrogative words stored in the interrogative word table from the past question sentences and answer sentences using sentence analysis, and generates a rule in which the result of the sentence analysis is generated in advance. 3. The dialog answer candidate proposal system according to claim 2, wherein the learning question text is generated based on the following. storing in the storage device a synonym table that stores words that are synonymous when analyzing the question text;
The learning question and answer database extracts the words stored in the synonym table from the past question sentences and answer sentences using sentence analysis, and extracts the words stored in the synonym table from the past question sentences and answer sentences, and extracts the words stored in the synonym table from the past question sentences and answer sentences, based on rules generated in advance as a result of the sentence analysis. The dialog answer candidate suggestion system according to claim 2, wherein the dialog answer candidate suggestion system is generated. The dialog answer candidate proposal according to claim 4, wherein each of the answer candidate sentences stored in the learning question and answer database is generated using a plurality of the past question sentences having similar contents. system. The transmitting/receiving device includes a network interface that allows the dialogue answer candidate suggestion system to be connected to a network,
The dialog answer candidate proposal system according to claim 2, wherein the processor receives input from the user terminal and the operator terminal, and outputs to the user terminal and the operator terminal via the network interface. . When the new question sentence from the questioner is input, the answer candidate sentence generation unit searches the learning question and answer database,
7. The dialogue answer candidate proposal system according to claim 6, wherein the processor causes an operator terminal to display a similar question to the new question text and answer candidates to the similar question. The processor executes in real time or in batch processing a learning question and answer data generation program that associates past question sentences stored in the question and answer database with past answer sentences for the past question sentences, 8. The dialogue answer candidate proposal system according to claim 7, wherein the learning question and answer database is constructed. a processor, a storage device, a transmitting/receiving device that communicates with a user terminal used by the questioner and an operator terminal used by the answerer, and past question texts from the questioner and past answers to the past question texts. a question-and-answer database that associates and stores
A method for proposing dialogue answer candidates in a dialogue answer candidate proposal system, comprising: a learning question and answer database that records learning question sentences and learning answer sentences generated based on information stored in the question and answer database; ,
Create a learning question and answer database using past question sentences or past answer sentences that are associated with learning question sentences,
When a new question sentence is input by the questioner,
Calculating the degree of similarity with the new question sentence for each of the past question sentence and the learning question sentence, and calculating the total similarity of each dialogue,
Past question sentences similar to the extracted new question sentences or past answer sentences associated with the learning question sentences are extracted from the question answer database or the learning question answer database and used as answer candidate sentences for the learning purpose. A method for proposing dialogue answer candidates, characterized by adding them to a question and answer database. storing in the storage device an interrogative word table that stores interrogative words expressing questions;
The processor extracts the interrogative words stored in the interrogative word table from the past question sentences and answer sentences using sentence analysis, and selects the answer candidates based on a rule generated in advance from the result of the sentence analysis. 10. The dialogue answer candidate proposal method according to claim 9, further comprising the step of creating a sentence. storing in the storage device a synonym table that stores words that are synonymous when analyzing the question text;
The processor extracts the words stored in the synonym table from the past question sentences and answer sentences using sentence analysis, and extracts the words from the past question sentences and answer sentences based on the result of the sentence analysis based on a rule generated in advance. 11. The dialogue answer candidate proposal method according to claim 10, wherein: The transmitting/receiving device includes a network interface that allows the dialogue answer candidate proposal system to be connected to a network,
The dialog answer candidate proposal according to claim 11, wherein the processor performs input from the user terminal and the operator terminal and output to the user terminal and the operator terminal via the network interface. Method. When the new question text from the questioner is input, the processor searches for similar questions from the learning question and answer database;
12. The dialog answer candidate proposing method according to claim 11, further comprising displaying the similar questions extracted by the search and answer candidates for the similar questions on the operator terminal. The processor performs in advance a learning question-and-answer data generation process that associates past question sentences stored in the question-and-answer database with past answer sentences for the past question sentences in a batch process. 14. The dialog answer candidate proposal method according to claim 13, wherein the learning question and answer database is enabled to be used when the new question text from the questioner is input.

Images