JP7629844B2 - ANSWER CANDIDATE PROPOSAL SYSTEM AND ANSWER CANDIDATE PROPOSAL METHOD - Google Patents

Description

The present invention relates to an answer candidate suggestion system and an answer candidate suggestion method that generate answer candidate sentences that are candidates for answers to new question sentences.

There is a technology that generates candidate answer sentences that serve as answers to new questions from users obtained via the Internet or public communication networks. For example, Patent Document 1 discloses a technology that breaks down a question string into multiple morphemes (words) and selects and outputs an answer string from a group of accumulated past answer strings based on the obtained multiple morphemes.

Patent Document 2 also discloses a technology that performs morphological analysis on a question sentence, breaks it down into multiple morphemes, performs a search process using a search query generated from the multiple morphemes obtained, extracts a set of candidate answer sentences from the search results, and ranks the candidate answer sentences included in the extracted set of candidate answer sentences.

JP 2018-181033 A JP 2013-254420 A

However, when a new question text contains many words unrelated to the intent of the question, the multiple morphemes obtained by morphological analysis of the question text contain many words unrelated to the intent of the question. In this case, the technology described in Patent Document 1 selects and outputs an answer string from a group of past answer strings based on the multiple morphemes containing many words unrelated to the intent of the question, which are obtained by morphological analysis of the question string. Therefore, the answer string to be output is an answer string related to many words unrelated to the intent of the question. Therefore, with the technology described in Patent Document 1, there is a risk that the answer string to be output will be an inappropriate answer string that does not match the intent of the user's question.

In the above case, the technology described in Patent Document 2 generates a search query from multiple morphemes containing many words unrelated to the intent of the question, which are obtained by morphological analysis of the question sentence, and performs a search process using the search query. The search results are influenced by words unrelated to the intent of the question that are contained in many morphemes included in the search query. For this reason, the contents of the answer candidate sentences extracted from the search results may have a low relationship to the intent of the question. Therefore, with the technology described in Patent Document 2, the answer candidate sentences may be inappropriate.

The object of the present invention is to provide an answer candidate suggestion system and an answer candidate suggestion method that output suitable answer candidate sentences for new question sentences.

In order to achieve the above object, one aspect of the answer candidate suggestion system of the present invention is an answer candidate suggestion system that generates answer candidate sentences that are candidates for answer sentences to a new question sentence, the system comprising a processor and a storage device, the storage device stores a question and answer database that associates and stores past question sentences with past answer sentences to the past question sentences, the processor generates a group of item candidate words based on the past question sentences and the new question sentence stored in the question and answer database when the new question sentence is input, and when an item word group selected by the user from the generated item candidate word group is input, the processor generates question information including the item word group based on the item word group and the new question sentence, calculates a similarity between each of the past question sentences stored in the question and answer database and the question information, extracts past question sentences similar to the question information from the question and answer database based on the similarity with the question information, and extracts from the question and answer database past answer sentences associated with the extracted past question sentences similar to the question information to set them as first answer candidate sentences.

An aspect of the answer candidate suggestion method of the answer candidate suggestion system of the present invention is an answer candidate suggestion method in an answer candidate suggestion system that generates answer candidate sentences that are candidates for answer sentences to a new question sentence, in which a storage device of the answer candidate suggestion system stores a question and answer database that associates and stores past question sentences with past answer sentences to the past question sentences, and when the new question sentence is input, a group of item candidate words is generated based on the past question sentence and the new question sentence stored in the question and answer database, and when an item word group selected by a user from the generated item candidate word group is input, question information including the item word group is generated based on the item word group and the new question sentence, and a similarity between each of the past question sentences stored in the question and answer database and the question information is calculated, and past question sentences similar to the question information are extracted from the question and answer database based on the similarity with the question information, and past answer sentences associated with the extracted past question sentences similar to the question information are extracted from the question and answer database to be used as first answer candidate sentences.

The present invention makes it possible to output suitable answer candidate sentences for new question sentences.

FIG. 2 is a diagram illustrating an example of a functional block diagram of the answer candidate suggestion system in the embodiment. 1 is a block diagram showing an example of a hardware configuration of an answer candidate suggestion system according to an embodiment. FIG. 2 is a diagram illustrating an example of a question and answer database. FIG. 13 is a diagram showing an example of a question word request word list. FIG. 13 is a diagram showing an example of a personal information word list. FIG. 13 is a diagram showing an example of a supplementary word list. FIG. 13 is a diagram showing an example of an item candidate word table. 11 is a flowchart illustrating an example of a process for generating a group of candidate item words according to an embodiment. FIG. 13 is an explanatory diagram showing an example of a new question sentence input screen displayed on a user terminal. FIG. 13 is an explanatory diagram showing an example of an item word group selection screen displayed on a user terminal. FIG. 13 is an explanatory diagram showing an example of an answer candidate generation and selection screen displayed on an operator terminal. 11 is a flowchart showing an example of an answer candidate sentence generation process of the answer candidate suggestion system 1. FIG. 13 is an explanatory diagram showing an example of an answer candidate display screen displayed on an operator terminal.

The following describes embodiments of the present invention with reference to the drawings. The embodiments are illustrative for explaining the present invention, and some parts have been omitted or simplified as appropriate for clarity of explanation. The present invention can also be implemented in various other forms. Unless otherwise specified, each component may be singular or plural.

The position, size, shape, range, etc. of each component shown in the drawings may not represent the actual position, size, shape, range, etc., in order to facilitate understanding of the invention. Therefore, the present invention is not necessarily limited to the position, size, shape, range, etc. disclosed in the drawings.

As examples of various types of information, expressions such as "table," "list," and "queue" may be used, but the various types of information may be expressed in other data structures. For example, various types of information such as "XX table," "XX list," and "XX queue" may be expressed as "XX information." When explaining identification information, expressions such as "identification information," "identifier," "name," "ID," and "number" are used, but these are interchangeable.

When there are multiple components with the same or similar functions, they may be described using the same reference numerals with different subscripts. Also, when there is no need to distinguish between these multiple components, the subscripts may be omitted.

In the embodiments, the processing performed by executing a program may be described. Here, the computer executes the program using a processor (e.g., CPU, GPU), and performs the processing defined by the program using storage resources (e.g., memory) and interface devices (e.g., communication ports). Therefore, the subject of the processing performed by executing the program may be the processor. Similarly, the subject of the processing performed by executing the program may be a controller, device, system, computer, or node having a processor. The subject of the processing performed by executing the program may be a calculation unit, and may include a dedicated circuit that performs specific processing. Here, the dedicated circuit is, for example, an FPGA (Field Programmable Gate Array), an ASIC (Application Specific Integrated Circuit), or a CPLD (Complex Programmable Logic Device).

The program may be installed on the 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 may include a processor and a storage resource that stores the program to be distributed, and the processor of the program distribution server may distribute the program to be distributed to other computers. In addition, in the embodiments, two or more programs may be realized as one program, and one program may be realized as two or more programs.

In the embodiment, when a new question sentence and a group of item words are input, the answer candidate suggestion system 1 generates answer candidate sentences (first answer candidate sentence and second answer candidate sentence described below) that are candidates for the answer sentence to the new question sentence based on the new question sentence and the group of item words.
The item word group is words related to the new question sentence, selected from the item candidate word group.
The item candidate word group is a plurality of words (item candidate words) related to a new question sentence that are generated based on past question sentences and new question sentences stored in the question and answer database 21 (described later). Note that the expression "-- word group" means at least one "-- word."

<System Configuration>
1 is a diagram showing an example of a functional block diagram of an answer candidate suggestion system 1 in an embodiment. As shown in FIG. 1, the answer candidate suggestion system 1 is connected to a user terminal 2, an operator terminal 3, and a web search engine 4 via a network NW.

The user terminal 2 is operated by a user who inputs a question sentence. The user terminal 2 has an input device that accepts input from the user, and an output device such as a display or touch panel that displays information. The user terminal 2 can send and receive information to and from the answer candidate suggestion system 1 and the operator terminal 3 via the network NW. The user terminal 2 can also display information received from the answer candidate suggestion system 1 and the operator terminal 3. The user terminal 2 can then transmit information input by the user to the answer candidate suggestion system 1 and the operator terminal 3.

The operator terminal 3 is operated by an operator and includes an input device that accepts input from the operator, and an output device such as a display or touch panel that displays information. The operator terminal 3 can send and receive information to and from the answer candidate suggestion system 1 and the user terminal 2 via the network NW. The operator terminal 3 is installed in a help desk that uses the answer candidate suggestion system 1, and may also be owned by, for example, a contractor of the help desk. For example, an electronic device such as a personal computer is used as the operator terminal 3.

When the web search engine 4 receives at least one word via the network NW, it returns search results including information about websites related to the received word. The website information included in the search results includes a summary of the website and the URL. Here, the summary is a summary sentence (e.g., about 100 characters) of each website generated by a web search engine, etc., and is sometimes called a snippet.

The network NW may be a wired network or a wireless network. The network NW may also be a global network such as the Internet, or a local area network (LAN).

The answer candidate suggestion system 1 includes an item candidate word group generation unit 11 and an answer candidate sentence generation unit 12. The answer candidate suggestion system 1 also stores a question and answer database 21, an interrogative word/request word list 22, a personal information word list 23, a supplementary word list 24, and an item candidate word table 25.

The item candidate word group generation unit 11, the details of which will be described later using the flowchart in FIG. 8, generates an item candidate word group related to a new question sentence when the new question sentence input by the user of the user terminal 2 is input to the answer candidate suggestion system 1. Furthermore, the item candidate word group generation unit 11 outputs the generated item candidate word group to a network I/F 36 (transmitting/receiving device) described later, and causes the network I/F 36 to transmit the item candidate word group to the user terminal 2 via the network NW.

The answer candidate sentence generation unit 12, the details of which will be described later using the flowchart in FIG. 12, generates answer candidate sentences (first answer candidate sentence, second answer candidate sentence) when a group of item words and a new question sentence are input to the answer candidate suggestion system 1. The answer candidate sentence generation unit 12 then outputs the generated answer candidate sentences to a network I/F 36 (transmitting/receiving device) described later, and causes the network I/F 36 to transmit the answer candidate sentences to the operator terminal 3 via the network NW.

The question and answer database 21, the details of which will be described later with reference to FIG. 3, stores past question sentences, past answer sentences to the past question sentences, and tfidf vectors of the past question sentences in association with each other.

The interrogative word/request word list 22, details of which will be described later with reference to FIG. 4, is a database that stores interrogative words that express doubts and request words that express requests.

The personal information word list 23 is a database that stores multiple personal information words that represent personal information, details of which will be described later using Figure 5.

The supplementary word list 24 is a database that stores supplementary words, the details of which will be described later using FIG. 6.

The item candidate word table 25, details of which will be described later with reference to FIG. 7, is a database that stores words (particularly verbs) contained in past question sentences in association with item candidate word groups.

Figure 2 is a block diagram showing an example of the hardware configuration of the answer candidate suggestion system 1. As shown in Figure 2, the answer candidate suggestion system 1 has a processor 31, a main memory device 32, a sub-memory device 33, an input device 34, an output device 35, a network I/F 36, and a bus 37 that connects these. The answer candidate suggestion system 1 can be realized by a general information processing device such as a PC or a server computer.

The processor 31 reads data and programs stored in the secondary storage device 33 into the main storage device 32 and executes the processing defined by the programs.

The main memory device 32 has volatile memory elements such as RAM, and stores the programs and data executed by the processor 31.

The secondary storage device 33 is a device that has non-volatile memory elements, such as a hard disk drive (HDD) or a solid state drive (SSD), and stores programs, data, etc. The secondary storage device 33 stores the above-mentioned question and answer database 21, interrogative word and request word list 22, personal information word list 23, supplementary word list 24, and item candidate word table 25.

In addition, the sub-storage device 33 has installed therein an item candidate word group generation program 11a and an answer candidate sentence generation program 12a. The item candidate word group generation unit 11 and the answer candidate sentence generation unit 12 described above with reference to FIG. 1 are realized by the processor 31 reading the item candidate word group generation program 11a and the answer candidate sentence generation program 12a stored in the sub-storage device 33 into the main storage device 32 and executing them.

The input device 34 is a device that accepts user operations such as a keyboard or mouse, and acquires information input by user operations. The output device 35 is a device that outputs information such as a display, and presents information to the user by displaying it on a screen, for example.

The network I/F 36 is an interface for transmitting and receiving data via the network NW with devices such as the user terminal 2, the operator terminal 3, and the web search engine 4. In other words, the network I/F 36 is a transmitting/receiving device capable of transmitting and receiving information to and from the user terminal 2, the operator terminal 3, and the web search engine 4 via the network NW. The answer candidate suggestion system 1 can use the network I/F 36 to transmit and receive data with devices such as the user terminal 2, the operator terminal 3, and the web search engine 4 that are connected to the network NW.

The user terminal 2 and the operator terminal 3 can be configured using hardware resources similar to those of the answer candidate suggestion system 1.

<Various data structures>
3 is a diagram showing an example of the question and answer database 21. In the question and answer database 21 shown in FIG. 3, a question and answer ID 301 is an ID for identifying a past question sentence 302. An answer sentence 303 and a tfidf vector 304 (details of which will be described later) are associated with the past question sentence 302. In this manner, the question and answer database 21 stores the past question sentence 302 and the past answer sentence 303 to the past question sentence 302 in association with each other. The tfidf vector 304 may be generated for all past question sentences stored in the question and answer database 21 every time a new pair of a past question sentence and a past answer sentence is stored in the question and answer database 21. In addition, the tfidf vector 304 may be regenerated at a preset timing, for example, every time a predetermined number of pairs of a past question sentence and a past answer sentence are stored in the question and answer database 21.

Fig. 4 is a diagram showing an example of the interrogative word request list 22. In the interrogative word request list 22 shown in Fig. 4, the interrogative word request ID 401 is an ID that identifies the interrogative word request 402. The interrogative word request 402 is an interrogative word that expresses a doubt or a request word that expresses a request. In Fig. 4, "shouka" and "kudasai" are shown as examples of the interrogative word request 402. Other examples of the interrogative word request 402 include "suka", "noka", "nka", "nano", "dare", "nani", "what", "where", "when", "how many", "how much", "how", "why", "ika", "do", "who", "who", "who", "what", "which", "what", "what kind", "any", "want", "want", "want", "want", "want", "request", and "?".

Figure 5 is a diagram showing an example of personal information word list 23. In personal information word list 23 shown in Figure 5, personal information word ID 501 is an ID that identifies personal information word 502. Personal information word 502 is a plurality of personal information words that represent personal information. Figure 5 shows, as examples of personal information word 502, "****-****" as a postal code and "****-****-****" as a telephone number. Other examples of personal information word 502 include "****" as a postal code, "********" as a telephone number, "card number", "date of birth", "email address", "name", and "address". Note that in the above, "*" represents a single numeric character.

Figure 6 is a diagram showing an example of the supplementary word list 24. In the supplementary word list 24 shown in Figure 6, a supplementary word ID 601 is an ID that identifies a supplementary word 602. A supplementary word 602 is a word that is likely to have an important meaning with respect to the intent of a question contained in a new question sentence. Furthermore, the supplementary word 602 is a word (morpheme) that is likely to be calculated as having a low importance and not extracted as an important word in the "(A) important word extraction method for calculating the importance of words contained in a sentence and extracting important words contained in the sentence" of the tf-idf method described below. Figure 6 shows "not" as an example of a supplementary word 602.

7 is a diagram showing an example of the item candidate word table 25. In the item candidate word table 25 shown in FIG. 7, the item candidate word ID 701 is an ID for identifying the word 702. The item candidate word group 703 is associated with the word 702. The item candidate word table 25 is generated, for example, based on past question sentences stored in the question and answer database 21 as follows. First, for each past question sentence stored in the question and answer database 21, an interrogative request sentence containing at least one interrogative word and request word stored in the interrogative word request word list 22 is extracted. Next, each of the extracted interrogative request sentences is subjected to morphological analysis, and the verb contained in the interrogative request sentence is set as the word 702 in the item candidate word table 25, and at least one noun contained in the interrogative request sentence is set as the item candidate word group 703 and stored in the item candidate word table 25. In this way, the item candidate word group 703 in the item candidate word table 25 is generated based on past question sentences stored in the question and answer database 21.

<tf-idf method>
The answer candidate suggestion system 1 calculates the importance and cosine similarity of the tf-idf method (word frequency inverse document frequency method) in the process of generating a group of item candidate words, a first answer candidate sentence, and a second answer candidate sentence based on a new question sentence. The importance is the importance of a word contained in a sentence. On the other hand, the cosine similarity is the similarity between sentences. Below, an overview of (A) an important word extraction method that calculates the importance of a word contained in a sentence and extracts important words contained in the sentence, and (B) a similar document extraction method that calculates the cosine similarity between sentences and extracts similar sentences similar to the target sentence in the tf-idf method will be described below. In the calculation of the importance in (A) and the calculation of the cosine similarity in (B), a database in which a plurality of sentences is stored (the question and answer database 21 in this embodiment) is used.

(A) In a method for extracting important words that calculates the importance of words contained in a sentence and extracts important words contained in the sentence, the importance is calculated for all words in the target sentence.

The importance of a word (referred to as the tfidf value) is the product of tf and idf. First, the sentence is subjected to morphological analysis to break the sentence down into morphemes (words). Then, tf is calculated. If the total number of words in the sentence is N, and the number of times the word for which importance is to be calculated appears in the sentence is n, then tf can be expressed, for example, as tf = n/N. tf represents the number of times a word appears in a sentence. In addition, with tf, the more times a word appears in a sentence, n, the more important it is considered to be.

Next, idf is calculated. Let D be the number of sentences stored in the database, and d be the number of sentences containing the word for which importance is to be calculated. idf is expressed, for example, as idf = -log(d/D). This can also be expressed as idf = log(D/d). The smaller the number d of sentences containing the word for which importance is to be calculated, the larger idf = log(D/d). idf indicates how few the number d of sentences containing the word for which importance is to be calculated is among all the sentences stored in the database. With idf, the smaller the number d of sentences containing the target word, the more important the word is considered to be.

The importance of a word is tfidf value = tf * idf = n/N * (-log(d/D)). Then, the tfidf value (importance) is calculated for all words in the sentence. Then, among the words with high tfidf values, a certain percentage (or a certain number) of the top words are determined to be important words.

(B) In a similar document extraction method that calculates the cosine similarity between sentences and extracts sentences that are similar to a target sentence, as described below, a tfidf vector is calculated between each sentence stored in a database and the target sentence, and the cosine similarity is calculated.

First, all sentences stored in the database and the target sentence are subjected to morphological analysis to break the sentences down into words (morphemes). Next, duplicate words are removed from the multiple words obtained from the breakdown, and a word vector with each word as a component is generated. Next, tfidf vectors are calculated for each sentence stored in the database and the target sentence. A tfidf vector is a vector whose components are the tfidf values for the words that are components of the word vector.

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

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

Then, among the sentences included in the database, a certain percentage (or a certain number) of the sentences that rank highest in terms of cosine similarity are determined to be similar sentences with high similarity. Here, instead of cosine similarity, the inner product of the tfidf vector of a sentence stored in the database and the tfidf vector of the target sentence may be used.

The above explanation is an overview of the tf-idf method, and the algorithm of the tf-idf method when using the tf-idf method can be changed as appropriate from the method explained above. Also, instead of the tf-idf method's "(B) similar document extraction method that calculates the cosine similarity between sentences and extracts similar sentences that are similar to the target sentence," other methods that calculate the similarity of sentences, such as the Doc2Vec method, may be used to extract similar sentences.

<Processing Procedure>
Next, a description will be given of the processing procedure of the answer candidate suggestion system 1. A user operates the user terminal 2 to cause the user terminal 2 to access the answer candidate suggestion system 1. When the answer candidate suggestion system 1 is accessed from the user terminal 2, it starts an item candidate word group generation process executed by the item candidate word group generation unit 11. Below, the item candidate word group generation process will be described using FIG. 8 with reference to FIG. 9 and FIG. 10.

Figure 8 is a flowchart showing an example of the process for generating a group of item candidate words in the answer candidate suggestion system 1.

The answer candidate suggestion system 1 transmits new question text input screen information to the user terminal 2 (step S101). The new question text input screen information includes information on the configuration of the new question text input screen and information to display the new question text input screen on the user terminal 2. The new question text input screen is configured to be able to accept input of a new question text and input to transmit the input new question text to the answer candidate suggestion system 1.

9 is an explanatory diagram showing an example of a new question sentence input screen displayed on the user terminal 2. The new question sentence input screen 900 shown in FIG. 9 includes a new question sentence input field 901 and an item selection button 902. The new question sentence input field 901 is a field in which the user inputs a new question sentence. The item selection button 902 is a button for inputting that the input new question sentence is to be sent to the answer candidate proposal system 1. When the user inputs a new question sentence in the new question sentence input field 901 and further presses the item selection button 902, the user terminal 2 transmits the new question sentence input in the new question sentence input field 901 to the answer candidate proposal system 1. In FIG. 9, a new question sentence is input in the new question sentence input field 901, saying, "I can't connect, so please help me. I, Hitachi Hanako, live in C-chome, B-city, A-prefecture, and I discovered this when I was sitting in a chair at home reading a book." When the item selection button 902 is pressed, the input new question sentence is transmitted to the answer candidate proposal system 1.

Returning to FIG. 8, next, the answer candidate suggestion system 1 waits for a predetermined time (step S102).

Next, the answer candidate suggestion system 1 determines whether or not a new question text has been received from the user terminal 2 (step S103). If it is determined that a new question text has been received from the user terminal 2 (step S103: YES), the process proceeds to step S104, and if it is determined that a new question text has not been received from the user terminal 2 (step S103: NO), the process returns to step S102. As a result, the answer candidate suggestion system 1 waits for a new question text by repeating the processes of steps S102 and S103 until a new question text is received from the user terminal 2.

Next, the answer candidate suggestion system 1 saves the new question text received from the user terminal 2 (step S104). Here, when the network I/F 36 (transmitter/receiver) of the answer candidate suggestion system 1 receives (inputs) the new question text from the user terminal 2, the processor 31 receives a notification from the network I/F 36 that the new question text has been received, and stores the received new question text in the main memory device 32. In this manner, the new question text is input to the answer candidate suggestion system 1.

Next, the answer candidate suggestion system 1 uses the question and answer database 21 to calculate the cosine similarity by the tf-idf method, extracts past question sentences similar to the new question sentence, and saves the extracted past question sentences similar to the new question sentence (step S105). Here, the answer candidate suggestion system 1 uses the question and answer database 21 to calculate the cosine similarity by the tf-idf method described above between the question information and each of the past question sentences stored in the question and answer database 21. Then, from the past question sentences stored in the question and answer database 21, a predetermined percentage (e.g., 20%) or a predetermined number (e.g., 3) of sentences are extracted from the top in terms of cosine similarity, and saved as past question sentences similar to the new question sentence.

Next, the answer candidate suggestion system 1 uses the question and answer database 21 to calculate the importance of the above-mentioned tf-idf method, and generates and saves a group of high importance words from past question sentences similar to the new question sentence extracted in step S105 (step S106). Here, the answer candidate suggestion system 1 performs morphological analysis on the past question sentences similar to the new question sentence extracted in step S105 to generate multiple past question sentence morphemes. Past question sentence morphemes are morphemes (words) obtained by morphological analysis of past question sentences similar to the new question sentence. Then, the answer candidate suggestion system 1 calculates the importance of the above-mentioned tf-idf method for each of the multiple past question sentence morphemes using the question and answer database 21. Then, among the multiple past question sentence morphemes, nouns from a predetermined percentage (e.g., 20%) or a predetermined number (e.g., 10) of past question sentence morphemes ranked in the top order of importance by the tf-idf method are saved as a group of high importance words.

Next, the answer candidate suggestion system 1 sets the word group remaining after excluding the words contained in the new question sentence from the high importance word group as an item candidate word group and saves it (step S107). Here, the answer candidate suggestion system 1 performs morphological analysis on the new question sentence to generate new question sentence morphemes (words) and saves them. New question sentence morphemes are morphemes obtained by morphological analysis of the new question sentence. In addition, the new question sentence morphemes are set as "words contained in the new question sentence." Then, the high importance word group is set as an item candidate word group by excluding "words contained in the new question sentence" (new question sentence morphemes). Note that step S107 may be omitted and the item importance word group may be set as the item candidate word group.

Next, the answer candidate suggestion system 1 outputs the item candidate word group and the item word group selection screen information to the network I/F 36 (transmitting/receiving device), causes the network I/F 36 to transmit the item candidate word group and the item word group selection screen information to the user terminal 2 via the network NW, and ends the process (step S108). The item word group selection screen information includes information on the configuration of the item word group selection screen and information to cause the user terminal 2 to display the item word group selection screen. The item word group selection screen, which will be described later with reference to FIG. 10, is configured to be able to display the item candidate word group and to be able to accept input of an item word group selected from the item candidate word group and input to transmit the input item word group and new question sentence to the operator terminal 3.

Figure 10 is an explanatory diagram showing an example of an item word group selection screen displayed on the user terminal 2. The item word group selection screen 1000 shown in Figure 10 includes item word selection buttons 1001-1004 and an item word submission button 1005. The item word selection buttons 1001-1004 are buttons on which item candidate words are drawn. When the item word selection buttons 1001-1004 are pressed by the user, the type of line indicating the frame changes. In the item word selection buttons 1001-1004, a frame drawn with a solid line indicates that the user has selected the item candidate word written in the item word selection button as an item word, and a frame drawn with a dashed line indicates that the user has not selected the item candidate word written in the item word selection button as an item word.

In the example of FIG. 10, the frames of the item word selection buttons 1001 and 1002 are solid lines, and the item word selection button 1001 "smartphone" and the item word selection button 1002 "radio waves" are selected as item words. The item word selection buttons 1003 and 1004 are dashed lines, and the item word selection button 1003 "code" and the item word selection button 1004 "electronic book" are not selected as item words.

The item word submission button 1005 is a button for inputting that the input item word group is to be sent to the answer candidate suggestion system 1. When the user presses the item word selection buttons 1001 to 1004 to select an item word, and then presses the item word submission button 1005, the user terminal 2 transmits the item word (item word group) selected with the item word selection buttons 1001 to 1004, the new question text, and answer candidate generation selection screen information to the operator terminal 3.

The answer candidate generation selection screen information includes information on the configuration of the answer candidate generation selection screen and information on displaying the answer candidate generation selection screen on the operator terminal 3. The answer candidate generation selection screen, which will be described later with reference to FIG. 11, is configured to be able to display a new question sentence and a group of item words, and to be able to accept input of information on whether or not to cause the answer candidate proposal system 1 to generate an answer candidate sentence, and input of web search setting information, which is information on whether or not to collect second answer candidate sentences by web search.

In addition, instead of transmitting the item word group, the new question sentence, and the answer candidate generation selection screen information to the operator terminal 3, the user terminal 2 may transmit the item word group and the new question sentence to the answer candidate suggestion system 1. Here, when the answer candidate suggestion system 1 receives the item word group and the new question sentence from the user terminal 2, it may appropriately set whether or not to acquire a second answer candidate sentence from the WEB (details will be described later), and execute an answer candidate sentence generation process, an example of which is shown in the flowchart of FIG. 12.

In steps S105 to S107 in FIG. 8 described above, the similarity (cosine similarity) between each past question sentence stored in the question and answer database 21 and the new question sentence is calculated, and past question sentences similar to the new question sentence are extracted based on the similarity with the new question sentence (step S105). The importance of each of the past question sentence morphemes generated from the past question sentences similar to the extracted new question sentence is calculated, and past question sentence morphemes with high importance are extracted from the past question sentence morphemes to form a group of item candidate words (steps S106 to S107). As a result, the group of item candidate words becomes words (past question sentence morphemes) with high importance and relatively important meanings in past question sentences similar to the new question sentence.

Furthermore, the answer candidate sentences (first answer candidate sentence, second answer candidate sentence) are generated based on the item candidate words selected from the item candidate word group and the new question sentence. For this reason, it is desirable that the item candidate word group has an important meaning with respect to the question of the new question sentence. In contrast, as described above, the item candidate word group is words that have a relatively important meaning in past question sentences similar to the new question sentence. Therefore, by generating the item candidate word group as described above, the answer candidate suggestion system 1 can generate a more appropriate item candidate word group.

In addition, instead of steps S105 to S107 in FIG. 8, a group of item candidate words may be generated from the new question sentence using the item candidate word table 25 as follows. First, the answer candidate suggestion system 1 extracts an interrogative request sentence that includes at least one interrogative word and request word stored in the interrogative word list 22 from a sentence included in the new question sentence. Next, the answer candidate suggestion system 1 performs morphological analysis on the extracted interrogative request sentence to obtain multiple words (morphemes) included in the interrogative request sentence. Next, the answer candidate suggestion system 1 extracts verbs from the multiple words included in the interrogative request sentence. Next, the answer candidate suggestion system 1 refers to the item candidate word table 25 (see FIG. 7), extracts an item candidate word group corresponding to the extracted verb from the item candidate word table 25, and obtains the item candidate word group.

The item candidate word group stored in the item candidate word table 25 is generated based on past question sentences stored in the question and answer database 21, as described above with reference to FIG. 7. Therefore, even in the method of generating an item candidate word group using the item candidate word table 25 described above, the item candidate word group is generated based on past question sentences and new question sentences stored in the question and answer database 21. This allows the answer candidate suggestion system 1 to generate a more appropriate item candidate word group.

11 is an explanatory diagram showing an example of an answer candidate generation selection screen displayed on the operator terminal 3. The answer candidate generation selection screen 1100 shown in FIG. 11 includes a new question sentence display frame 1101, an item word group display frame 1102, a web search selection button 1103, a send button 1104, and an answer button 1105. The new question sentence display frame 1101 is a frame for displaying a new question sentence. The item word group display frame 1102 is a frame for displaying an item word group. When an operator who operates the operator terminal 3 presses (clicks, etc.) inside the new question sentence display frame 1101, the operator terminal 3 accepts input from the operator and allows the operator to edit the new question sentence in the new question sentence display frame 1101. This allows the operator terminal 3 to transmit the new question sentence edited by the operator, such as correcting typos, to the answer candidate proposal system 1. As a result, the answer candidate proposal system 1 regards the edited new question sentence as a new question sentence and generates an answer candidate sentence. This allows the answer candidate suggestion system 1 to generate a more suitable first answer candidate sentence.

The web search selection button 1103 is a button for inputting web search setting information, which is information on whether or not the answer candidate suggestion system 1 will collect second answer candidate sentences through a web search. The web search selection button 1103 is filled in black as shown in FIG. 11 if the answer candidate suggestion system 1 is to collect second answer candidate sentences through a web search, and is filled in white if the answer candidate suggestion system 1 is not to collect second answer candidate sentences through a web search. Here, whether it is filled in black or white is switched every time the operator presses the web search selection button 1103.

When the operator presses the send button 1104, the operator terminal 3 transmits to the answer candidate suggestion system 1 generation start information including the new question sentence in the new question sentence display frame 1101, the item word group, the web search setting information, and answer candidate sentence generation start information which is information that instructs the answer candidate suggestion system 1 to generate an answer candidate sentence. Here, the new question sentence included in the generation start information is the new question sentence in the new question sentence display frame 1101 at the time the operator presses the send button 1104. Therefore, if the operator edits the new question sentence in the new question sentence display frame 1101 before pressing the send button 1104, the edited new question sentence is included in the start information as the new question sentence. As described above, when the operator edits the new question sentence in the new question sentence display frame 1101, the operator can add modifications such as underlining or bolding to a part of the new question sentence (for example, a part that the operator thinks is important to the question), and further, the character information of the part that the operator added modifications to may be included in the generated disclosure information as important character information. Then, as described below, the answer candidate suggestion system 1 may generate answer candidate sentences (first answer candidate sentence, second answer candidate sentence) using the important character information.

When the answer button 1105 is pressed, the screen displayed on the operator terminal 3 is switched from the answer candidate generation selection screen to a screen where the operator inputs an answer sentence to the new question sentence.

When the answer candidate suggestion system 1 receives generation start information from the operator terminal 3 via the network I/F 36 (which inputs a new question sentence included in the generation start information, and further inputs a group of item words selected by the user from the group of item candidate words), it starts the answer candidate sentence generation process executed by the answer candidate sentence generation unit 12.

Figure 12 is a flowchart showing an example of the answer candidate sentence generation process of the answer candidate suggestion system 1.

The answer candidate suggestion system 1 saves the web search setting information, the new question text, and the group of item words contained in the generation start information received from the operator terminal 3 (step S201).

Next, the answer candidate suggestion system 1 uses the interrogative word request word list 22 to extract an interrogative request sentence from the new question sentence, and generates question information by adding the item word group to the extracted interrogative request sentence (step S202). Here, the answer candidate suggestion system 1 extracts an interrogative request sentence that includes at least one interrogative word or request word stored in the interrogative word request word list 22 from the new question sentence. Then, the answer candidate suggestion system 1 adds the item word group to the end or beginning of the extracted interrogative request sentence to generate question information. For example, if the interrogative request sentence is "I can't connect, please help me." and the item word group is "smartphone" and "radio waves," the question information is, for example, "I can't connect, please help me. Smartphone, radio waves" or "Smartphone, radio waves, I can't connect, please help me." Note that, as described above, if the generation start information includes important character information, the interrogative request sentence, the item word group, and characters included in the important character information (parts that overlap with the interrogative request sentence may be deleted) may be added to generate question information. As a result, the operator can add the characters of the part of the new question sentence that has been modified by, for example, underlining to the question information. This allows the answer candidate suggestion system 1 to generate a more suitable first answer candidate sentence.

Next, the answer candidate suggestion system 1 uses the question and answer database 21 to calculate the cosine similarity by the tf-idf method, extracts past question sentences similar to the question information, and saves the past answer sentences associated with the extracted past question sentences as first answer candidate sentences (step S203). Here, the answer candidate suggestion system 1 uses the question and answer database 21 to calculate the cosine similarity between the question information and each of the past question sentences stored in the question and answer database 21 by the above-mentioned tf-idf method. In addition, among the past question sentences stored in the question and answer database 21, a predetermined percentage (e.g., 20%) or a predetermined number (e.g., 3) of past question sentences are extracted from the top in terms of cosine similarity. Then, past answer sentences associated with the extracted past question sentences are extracted from the question and answer database 21, and the extracted past answer sentences are saved as first answer candidate sentences.

Next, the answer candidate suggestion system 1 determines whether or not to collect second answer candidate sentences through a web search based on the web search setting information saved in step S201 (step S204). If it is determined that second answer candidate sentences are to be collected through a web search (step S204: YES), the process proceeds to step S205, and if it is determined that second answer candidate sentences are not to be collected through a web search (step S204: NO), the process proceeds to step S208. Here, as described above, the web search information is information regarding whether or not to collect second answer candidate sentences through a web search, and the answer candidate suggestion system 1 can determine whether or not to collect second answer candidate sentences through a web search based on the web search information.

Next, the answer candidate suggestion system 1 performs morphological analysis on the question information to obtain multiple question information morphemes, calculates the importance of the tf-idf method for each question information morpheme using the question and answer database 21, extracts multiple question information morphemes with high importance in the tf-idf method from the multiple question information morphemes, and further stores the question information morphemes obtained by excluding the personal information words stored in the personal information word list 23 as a search word group (step S205). Here, the question information morphemes are morphemes obtained by performing morphological analysis on the question information. The method of calculating the importance of the tf-idf method has been described above. The multiple question information morphemes with high importance in the tf-idf method are a predetermined percentage (e.g., 20%) or a predetermined number (e.g., 3) of question information morphemes that are ranked from the top in terms of importance among the question information morphemes.

In addition, in step S205, if a supplementary word (e.g., "not") stored in the supplementary word list 24 is present in the question information, the answer candidate suggestion system 1 may add the supplementary word in the question information to the search word group. This may result in a more desirable second answer candidate being obtained.

In addition, among the words (morphemes) included in the search word group, it is preferable that the words to be inflected are in a form that retains the inflected form (for example, "tsunagara" (connect), "tasuku" (help)), but they may be in a basic form that removes the inflected form information (for example, "tsunagara" (connect), "tasuku" (help)).

Next, the answer candidate suggestion system 1 transmits the search word group to the web search engine, obtains search results related to the search word group returned from the web search engine, and collects summaries of the web sites included in the search results (step S206). Here, the answer candidate suggestion system 1 (processor 31) outputs the search word group to the network I/F 36 (transmitter/receiver) together with a notice that the network I/F 36 (transmitter/receiver) will transmit the search word group to the web search engine 4 so that the network I/F 36 (transmitter/receiver) transmits the search word group to the web search engine. As a result, the network I/F 36 transmits the search word group to the web search engine 4 via the network NW. The web search engine 4, which has received the search word group, returns the search results related to the search word group to the answer candidate suggestion system 1. The search results related to the search word group include summaries of the web sites related to the search word group.

The answer candidate suggestion system 1 may store in advance at least one web page that lists pairs of past question sentences and their answers, and set the stored web page as the target web site to be searched by a web search engine in the process of step S206. This makes it easier to collect summary sentences of web sites, and therefore easier to collect second answer candidate sentences (described below).

In addition, instead of using a web search engine in the process of step S206, a search device may be used that searches for pre-registered and stored data in a specified device (e.g., data pairs of past question sentences and their answers). Here, the search device may, for example, generate and store a summary similar to the summary of the website, and use this summary instead of the summary of the website. This makes it possible to collect summaries more efficiently, and ultimately to collect second candidate answer sentences (described below) more efficiently.

Next, the answer candidate suggestion system 1 ranks the web site summaries obtained in step S206 using a predetermined ranking method, and saves a predetermined number (or a predetermined percentage) of the top ranked web site summaries as second answer candidate sentences (step S207). Here, the ranking method may be, for example, the cosine similarity between each of the web site summaries and the new question sentence by the tf-idf method. The ranking method may also be ranking using the ranking model described in Patent Document 2, or other known techniques. Instead of ranking in step S207, the search summaries of the top search results ranked by the web search engine in step S206 may be used as the second answer candidate sentence. Furthermore, the second answer candidate sentence may include the URL of the web site. In this way, when reading the second answer candidate sentence, the operator of the operator terminal 3 can access the web site using the URL to obtain information about the web site summaries included in the second answer candidate sentence.

Next, the answer candidate suggestion system 1 outputs the answer candidate sentence and the answer candidate sentence display screen information to the output device (network I/F 36), causes the output device (network I/F 36) to transmit the answer candidate sentence and the answer candidate sentence display screen information to the operator terminal 3, and ends the process (step S208). Here, the answer candidate sentence includes the first answer candidate sentence and the second answer candidate sentence. Needless to say, if it is determined in the process of step S204 that the second answer candidate sentence is not to be collected by the web search (step S204: NO), the answer candidate sentence will be only the first answer candidate sentence. In addition, the answer candidate sentence display screen information includes information on the configuration of the answer candidate sentence display screen and information to the effect that the answer candidate sentence display screen is to be displayed on the operator terminal 3. The answer candidate sentence display screen is configured to be able to display the answer candidate sentence.

Figure 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 1300 shown in Figure 13 includes a first frame 1301, first answer candidate sentence fields 1302 and 1303, a second frame 1304, and second answer candidate sentence fields 1305 and 1306. To the right of the first frame 1301, which is written "Past Answer" indicating the first answer candidate sentence, the first answer candidate sentence fields 1302 and 1303 displaying the first answer candidate sentence are shown. Similarly, to the right of the second frame 1304, which is written "Web Search" indicating the second answer candidate sentence, the second answer candidate sentence fields 1305 and 1306 displaying the second answer candidate sentence are shown.

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 to a new question sentence by referring to the displayed answer candidate sentences. This allows the operator to generate an answer sentence more easily. In addition, the amount of energy required by the operator to generate an answer sentence and the amount of carbon dioxide emissions generated can be reduced, thereby curbing global warming.

In this way, in the embodiment, the answer candidate suggestion system 1 generates answer candidate sentences (first answer candidate sentence and second answer candidate sentence) based not only on the new question sentence but also on a group of item words selected by the user from a group of item candidate words generated based on the past question sentence and the new question sentence stored in the question and answer database 21. As a result, the answer candidate suggestion system 1 can generate and output suitable answer candidate sentences that are more in line with the intent of the question of the new question sentence compared to when answer candidate sentences are generated based only on the new question sentence.

In addition, answer candidate sentences (first answer candidate sentences and second answer candidate sentences) are generated using past question sentences and past answer sentences stored in the question and answer database 21. This allows the answer candidate suggestion system 1 to more easily generate answer candidate sentences.

The answer candidate suggestion system 1 also extracts a question/request sentence that includes at least one interrogative word and one request word from the new question sentence, and generates question information by adding a group of item words to the extracted question/request sentence (step S202 in FIG. 12). This allows the answer candidate suggestion system 1 to generate a first answer candidate sentence based on a question/request sentence that is highly related to the intent of the question, excluding parts of the new question sentence that are less related to the intent of the question. Therefore, the answer candidate suggestion system 1 can generate a suitable first answer candidate sentence that is more in line with the intent of the user's question.

In addition, in step S203 of FIG. 12, the similarity between the question information and each of the past question sentences stored in the question and answer database 21, which is calculated by the answer candidate suggestion system 1, is the cosine similarity calculated by the tf-idf method using the question and answer database 21. This allows the answer candidate suggestion system 1 to easily calculate the similarity, and therefore to more easily generate the first answer candidate sentence.

In addition, the answer candidate suggestion system 1 outputs the answer candidate sentences (first answer candidate sentence and second answer candidate sentence) to the network I/F 36 (transmitter/receiver device), and causes the network I/F 36 (transmitter/receiver device) to transmit the answer candidate sentences to the operator terminal 3 via the network NW. This allows the operator operating the operator terminal 3 to easily read the answer candidate sentences (first answer candidate sentence and second answer candidate sentence).

The answer candidate suggestion system 1 also transmits the search words to the web search engine 4, and generates the second answer candidate sentence based on the search results related to the search words returned from the web search engine 4. This allows the answer candidate suggestion system 1 to more easily generate the second answer candidate sentence.

The answer candidate suggestion system 1 also extracts a question request sentence from the new question sentence, generates question information by adding a group of item words to the question request sentence (step S202 in FIG. 12), and generates a search word group from multiple question information morphemes (question information morphemes) of the question information based on the importance of the tf-idf method (step S205 in FIG. 12). As a result, the answer candidate suggestion system 1 can generate a search word group based on the question request sentence that is highly related to the intention of the question, excluding parts of the new question sentence that are less related to the intention of the question, and can thus generate a more appropriate second answer candidate sentence. In addition, the search word group is generated based on the importance of the tf-idf method, so that a search word group that is more in line with the intention of the question can be generated. Therefore, the answer candidate suggestion system 1 can generate a suitable second answer candidate sentence that is more in line with the intention of the user's question.

The answer candidate suggestion system 1 also generates a search word group from multiple question information morphemes (question information morphemes) with high importance in the tf-idf method, excluding personal information words stored in the personal information word list (step S205 in FIG. 12). The search word group is sent to the web search engine 4, which sends search results searched for using the search word group to the answer candidate suggestion system 1. Because personal information words related to the user's privacy are excluded from the search word group, the answer candidate suggestion system 1 can obtain search results while protecting the user's privacy, and can therefore generate a second answer candidate sentence while protecting the user's privacy.

The answer candidate suggestion system 1 also generates second answer candidate sentences based on the summaries of the websites included in the search results related to the search word group (steps S206 and S207 in FIG. 12). This adjusts the length of the second answer candidate sentences to an extent that makes it easy to understand the content of the second answer candidate sentences. This makes it easy for the operator to understand the content of the second answer candidate sentences.

The answer candidate suggestion system 1 also ranks the summaries of the websites using a predetermined ranking method, and sets the summaries of the top-ranked websites as the second answer candidate sentences (step S207 in FIG. 12). This allows the answer candidate suggestion system 1 to generate more appropriate second answer candidate sentences.

The answer candidate suggestion system 1 also calculates the similarity (cosine similarity) between each past question sentence stored in the question and answer database 21 and the new question sentence, and extracts past question sentences similar to the new question sentence (step S105 in FIG. 8). The importance of each of the past question sentence morphemes generated from the past question sentences similar to the extracted new question sentence is calculated, and past question sentence morphemes with high importance are extracted and used as an item candidate word group (steps S106 to S107 in FIG. 8). As a result, the item candidate word group becomes words (past question sentence morphemes) that have high importance and relatively important meanings in past question sentences similar to the new question sentence. In addition, since the item candidate word group is used to generate answer candidate sentences (first answer candidate sentence, second answer candidate sentence), it is desirable that the item candidate word group be words that have important meanings to the question of the new question sentence. Therefore, the answer candidate suggestion system 1 can generate a more appropriate item candidate word group.

In addition, in the item candidate word group generation process, an example of which is shown in the flowchart in FIG. 8, the similarity between the new question sentence and each of the past question sentences stored in the question and answer database 21 is the cosine similarity calculated by the tf-idf method using the question and answer database 21 (step S105 in FIG. 8). In addition, the importance of each of the multiple past question sentence morphemes is the importance calculated by the tf-idf method using the question and answer database 21 (step S106 in FIG. 8). In this way, by using the cosine similarity and importance calculated by the tf-idf method, the answer candidate suggestion system 1 can more easily generate item candidate word groups.

The present invention is not limited to the above-described embodiments, but includes various modifications and equivalent configurations within the spirit of the appended claims. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to those having all of the configurations described. In addition, other configurations may be added to, deleted from, or substituted for part of the configuration of the embodiments.

1: Answer candidate suggestion system 2: User terminal 3: Operator terminal 4: Web search engine 11: Item candidate word group generation unit 11a: Item candidate word group generation program 12: Answer candidate sentence generation unit 12a: Answer candidate sentence generation program 21: Question and answer database 22: Interrogative word and request word list 23: Personal information word list 24: Supplementary word list 25: Item candidate word table 31: Processor 32: Main memory device 33: Sub-memory device 34: Input device 35: Output device 36: Network I/F
37: Bus

Claims (10)

An answer candidate suggestion system that generates answer candidate sentences that are candidates for answer sentences to a new question sentence,
A processor and a storage device,
The storage device includes:
storing a question and answer database that stores past question sentences and past answer sentences to the past question sentences in association with each other;
The processor,
When the new question text is input,
calculating a similarity between the new question sentence and each of the past question sentences stored in the question and answer database;
extracting the past question sentences similar to the new question sentence based on the calculated similarity with the new question sentence;
performing morphological analysis on the extracted past question sentences to generate a plurality of past question sentence morphemes;
Calculate the importance of each of the multiple past question sentence morphemes,
extracting the past question sentence morphemes having high importance from the plurality of past question sentence morphemes, and setting the extracted morphemes as a group of candidate words for an item;
When the item word group selected by the user from the item candidate word group and the new question sentence are input,
generating question information including the item word group based on the item word group and the new question sentence;
calculating a similarity between the question information and each of the past question sentences stored in the question and answer database;
extracting past question sentences similar to the question information from the question and answer database based on the similarity with the question information;
extracting from the question and answer database a past answer sentence associated with a past question sentence similar to the extracted question information, and setting the extracted answer sentence as a first answer candidate sentence;
Answer candidate suggestion system. The answer candidate suggestion system according to claim 1,
The storage device further stores an interrogative word and a request word list for storing interrogative words expressing doubt and request words expressing requests,
The processor,
extracting an interrogative sentence including at least one of the interrogative words and the desired words stored in the interrogative word desired word list from the new question sentence;
generating the question information by adding the item word group to the extracted question/request sentence;
Answer candidate suggestion system. The answer candidate suggestion system according to claim 1,
The similarity to the question information is a cosine similarity calculated by a tf-idf method using the question and answer database.
Answer candidate suggestion system. The answer candidate suggestion system according to claim 1,
Further, a transmitting/receiving device is provided which is connected to the network and capable of transmitting and receiving information via the network,
The processor outputs the first answer candidate sentence to the transmitting/receiving device.
Answer candidate suggestion system. The answer candidate suggestion system according to claim 1,
Further, a transmitting/receiving device is connected to a network connected to a web search engine that, upon receiving at least one word, returns search results including information on a web site related to the received word, and is capable of transmitting and receiving information to and from the web search engine via the network,
The storage device further stores an interrogative word and a request word list for storing interrogative words expressing doubt and request words expressing requests,
The processor,
Extracting an interrogative sentence including at least one of the interrogative words or the desired words stored in the interrogative word desired word list from the new question sentence;
generating the question information by adding the item word group to the question request sentence;
morphologically analyzing the question information to generate a plurality of question information morphemes;
calculating a TF-IDF importance for each of the plurality of question information morphemes using a question and answer database;
generating a search word group from the plurality of question information morphemes based on the importance of each of the plurality of question information morphemes in the TF-IDF method ;
outputting the generated search word group to the transmitting/receiving device so that the transmitting/receiving device transmits the generated search word group to the web search engine;
obtaining a search result relating to the group of search words received by the transmitting/receiving device from the web search engine;
The answer candidate suggestion system generates a second answer candidate sentence based on the search results related to the acquired group of search words. The answer candidate suggestion system according to claim 5 ,
The storage device further stores a personal information word list storing a plurality of personal information words representing personal information;
The processor,
extracting a plurality of question information morphemes having a high importance in the TF-IDF method from the plurality of question information morphemes, and generating the search word group by excluding at least one of the personal information words stored in the personal information word list;
Answer candidate suggestion system. The answer candidate suggestion system according to claim 5,
the search results returned by the web search engine include a summary of the web site;
The processor generates a second answer candidate sentence based on an outline of a website included in a search result related to the group of search words.
Answer candidate suggestion system. The answer candidate suggestion system according to claim 7 ,
the processor ranks abstracts of the websites included in the search results related to the group of search words using a predetermined ranking method, and sets the abstracts of the websites ranked higher as second answer candidate sentences.
Answer candidate suggestion system. The answer candidate suggestion system according to claim 1 ,
the similarity with the new question sentence is a cosine similarity calculated by the tf-idf method using the question and answer database,
the importance of each of the plurality of past question sentence morphemes is an importance calculated by a TF-IDF method using the question and answer database;
Answer candidate suggestion system. An answer candidate suggestion method in an answer candidate suggestion system that generates answer candidate sentences that are candidates for answer sentences to a new question sentence, comprising:
The answer candidate suggestion system includes a processor and a storage device,
The storage device includes:
storing a question and answer database that stores past question sentences and past answer sentences to the past question sentences in association with each other;
The processor,
When the new question text is input,
calculating a similarity between the new question sentence and each of the past question sentences stored in the question and answer database;
extracting the past question sentences similar to the new question sentence based on the calculated similarity with the new question sentence;
performing morphological analysis on the extracted past question sentences to generate a plurality of past question sentence morphemes;
Calculate the importance of each of the multiple past question sentence morphemes,
extracting the past question sentence morphemes having high importance from the plurality of past question sentence morphemes, and setting the extracted morphemes as a group of candidate words for an item;
When the item word group selected by the user from the item candidate word group and the new question sentence are input,
generating question information including the item word group based on the item word group and the new question sentence;
calculating a similarity between the question information and each of the past question sentences stored in the question and answer database;
extracting past question sentences similar to the question information from the question and answer database based on the similarity with the question information;
extracting from the question and answer database a past answer sentence associated with a past question sentence similar to the extracted question information, and setting the extracted answer sentence as a first answer candidate sentence;
How to suggest answer candidates.

Images