Text conversion method and text conversion system

Abstract

To enhance accuracy of answers given by generation AI to questions by appropriately rewriting input texts when querying the generation AI.SOLUTION: A text conversion method identifies a word combination or phrase combination to be converted in an input text based on relation information indicating a relation between word combinations or relation information indicating a relation between phrase combinations included in the input text. Then, the method identifies a conversion destination word after converting a word or a conversion destination phrase after converting a phrase based on the relation information of the word combinations or the relation information of the phrase combinations. Then, the method generates a converted input text in which the word or phrase included in the input text is converted with the conversion destination word or conversion destination phrase. Then, the method inputs the converted input text into an answer generation system, and obtains an output text outputted with respect to the converted input text from the answer generation system.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to a text conversion method and a text conversion system. [Background technology]

[0002] A method is disclosed for accepting text data of questions from users and acquiring text data of answers using generative AI (Artificial Intelligence) such as ChatGPT and GPT-3. Also, a method is disclosed for improving the convenience of these generative AIs by providing them with additional learning data or combining them with external tools such as search engines and calculators. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Tatsuro Inaba, Kanichi Kiyomaru, Fei Cheng, Sadao Kurohashi, “An inference framework using multiple external tools based on a large-scale language model”, [online], 2023, Proceedings of the 29th Annual Conference of the Association for Natural Language Processing, [Retrieved July 10, 2023], Internet <URL:https: / / www.anlp.jp / proceedings / annual_meeting / 2023 / pdf_dir / Q8-10.pdf [Non-Patent Document 2] Aitor Lewkowycz, et al., “Solving quantitative reasoning problems with language models.”, [online], 2022, arXiv preprint arXiv:2206.14858, [Retrieved July 10, 2020], Internet <URL:https: / / arxiv.org / pdf / 2206.14858.pdf Summary of the Invention [Problem to be solved by the invention]

[0004] Here, the above-mentioned non-patent document 2 reports that the ability of the generation AI to respond to inquiries that require arithmetic operations is inferior to that of humans.

[0005] For example, consider the question, "You can get to Shuri from Naha Airport in 30 minutes by Okinawa Airport Shuttle, or in 15 minutes by Yui Rail. Okinawa Churaumi Aquarium is a 20-minute walk from Shuri. What is the shortest travel time to Okinawa Churaumi Aquarium?" The correct answer to this question is that it takes 15 minutes to get to Shuri from Naha Airport by Yui Rail, and then 20 minutes to walk from Shuri, so the total is 35 minutes.

[0006] On the other hand, when this query is input to GPT-3, it does not always give a correct answer, and may answer, for example, "45 minutes." If this question is rewritten as "You can get to Shuri from Naha Airport in 30 minutes by bus or 15 minutes by train. The aquarium is a 20-minute walk from Shuri. What is the shortest time it takes to get to the aquarium?" and input to GPT-3, the answer changes to "50 minutes by bus, 35 minutes by train."

[0007] In addition to the example questions above, if GPT-3 is given the question "Of the following days, Monday, Tuesday, Wednesday, Thursday, Friday, Saturday, and Sunday, the opening days are Monday and Tuesday. What days are the closing days?", it will only answer "Sunday." On the other hand, if the query is rewritten and input to GPT-3 as "Of the following days, Monday, Tuesday, Wednesday, Thursday, Friday, Saturday, and Sunday, the opening days are Monday and Tuesday. What days are the closing days?", the answer will change to "The closing days are Wednesday, Thursday, Friday, Saturday, and Sunday."

[0008] Taking the above-mentioned circumstances into consideration, the present invention aims to improve the accuracy of a generation AI's answers to questions by appropriately rewriting the input text when querying the generation AI. [Means for solving the problem]

[0009] As one aspect for solving the above-mentioned problem, there is provided a text conversion method executed by a text conversion system which converts an input text including a question to be input to an answer generation system which outputs an output text including an answer to the question when the input text including a question is input, the text conversion system having a memory and a processor, the processor extracting a plurality of words or a plurality of phrases which satisfy a predetermined condition from the input text, extracting relationship information indicating a relationship between the combinations of the words or the combinations of the phrases, and extracting the combinations of the words or the phrases to be converted in the input text based on the relationship information between the combinations of the words or the relationship information between the combinations of the phrases. a conversion target word or a conversion target phrase after the words are converted based on the relationship information of the word combination or the relationship information of the phrase combination; generating a converted input text in which the word or the phrase included in the input text is converted with the converted target word or the converted target phrase; inputting the converted input text to the answer generation system; obtaining the output text output from the answer generation system for the converted input text; and converting the converted target word or the converted phrase included in the output text to the word or the phrase before conversion. Effect of the Invention

[0010] According to the present invention, it is possible to improve the accuracy of answers given by a generation AI to questions. [Brief description of the drawings]

[0011] [Figure 1] FIG. 1 is a diagram showing the hardware configuration of a computer that realizes the text conversion system according to the first embodiment. [Diagram 2] FIG. 1 is a block diagram showing the configuration of a text conversion system according to a first embodiment. [Diagram 3] FIG. 2 is a diagram showing an example of the configuration of data in an input text DB according to the first embodiment. [Figure 4] FIG. 2 is a diagram showing an example of the data configuration of a word DB according to the first embodiment. [Diagram 5] FIG. 2 is a diagram showing an example of the data configuration of a word relation DB according to the first embodiment. [Figure 6] FIG. 4 is a diagram showing an example of the data configuration of a conversion pattern DB according to the first embodiment. [Figure 7] FIG. 4 is a diagram showing an example of the data configuration of a template DB. [Figure 8] FIG. 13 is a diagram showing an example of the data configuration of an answer text selection record DB. [Figure 9] 4 is a flowchart showing the operation procedure of the text conversion device according to the first embodiment. [Figure 10] 5 is a flowchart showing a conversion pattern generation process according to the first embodiment. [Figure 11] 5 is a flowchart showing a conversion pattern evaluation data generation process according to the first embodiment. [Figure 12] 5 is a flowchart showing a conversion pattern certainty measurement process according to the first embodiment. [Figure 13] 11 is a flowchart showing a complementary text collection process according to the first embodiment. [Figure 14] 4 is a flowchart showing a conversion keyword selection process according to the first embodiment. [Figure 15] 5 is a flowchart showing a conversion pattern selection process according to the first embodiment. [Figure 16] 4 is a flowchart showing a QA system inquiry process according to the first embodiment. [Figure 17] 4 is a flowchart showing an answer text reverse conversion process according to the first embodiment. [Figure 18] FIG. 4 is a diagram showing a display example of a processing result of the query result visualization processing according to the first embodiment. [Figure 19] 10 is a flowchart showing a conversion pattern evaluation data generation process according to the second embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings. Note that the embodiments described below do not limit the invention according to the claims, and all of the elements and combinations thereof described in the embodiments are not necessarily essential to the solution of the invention.

[0013] In the following description, a program realizing each processing function unit is executed by a processor to perform a predetermined process using a storage resource (e.g., a memory) and / or a communication interface device (e.g., a port) as appropriate. The processor is, for example, a CPU (Central Processing Unit). Therefore, the process executed by each processing function unit may be a process performed by a processor or a computer having the processor.

[0014] In the following explanation, various information is expressed in table format, but the information is not limited to table format and may be in CSV (Comma Separated Values) or other formats. Since various information does not depend on the data format, for example, "XXX table" can be called "XXX information".

[0015] In addition, among the items in the various information, the names of items that include information to distinguish the corresponding record from other records, such as "# (number)", "identification information", "ID", and "name", may be interchangeable. For example, "user ID" may be interchangeable with "user identification information".

[0016] [Embodiment 1] (Hardware configuration of computer 100) FIG. 1 is a diagram showing the hardware configuration of a computer 100 that realizes a text conversion system S according to the first embodiment.

[0017] The text conversion system S is realized by a computer 100 that executes a program corresponding to the processing. The computer 100 is configured by connecting a CPU 110, a memory 120, a hard disk drive 130, an input device 140 such as a keyboard or a mouse, an output device 150 such as a display, and a network device 160 via a bus 170. For example, the text conversion processing according to this embodiment is realized by the CPU 110 executing a text conversion program stored in the hard disk drive 130. However, the implementation of the text conversion system S is not limited to this.

[0018] (Configuration of text conversion system S according to embodiment 1) FIG. 2 is a block diagram showing the configuration of the text conversion system S according to the first embodiment.

[0019] The text conversion system S accepts text input from a user terminal 201, and generates a conversion pattern for the input text, conversion pattern evaluation data, and complementary text. The text conversion system S then applies the conversion pattern, conversion pattern evaluation data, and complementary text to the input text to generate a query text for a QA system (answer generation system) 200. The QA system 200 is a type of generation AI that takes the input text as a question and outputs an answer to the question. The text conversion system S then reverse-converts the answer text obtained from the QA system 200 for the query text, and visualizes the query result.

[0020] The text conversion system S includes a text information DB (Data Base) 210, a conversion pattern information DB 220, a visualization result DB 230, a control unit 240, and a query result visualization unit 250.

[0021] The text information DB210 stores data associated with the input text, which is required for converting the input text and visualizing the results of a query made to the QA system 200. The text information DB210 includes an input text DB211 that stores the input text and a complementary text that complements the input text, a word DB212 that stores keywords to be converted, and a word relationship DB213 that stores relationship information between keywords to be converted. The detailed configurations of the data in the input text DB211, the word DB212, and the word relationship DB213 will be described later with reference to Figs. 3, 4, and 5.

[0022] The data of the input text DB 211 is supplied by a supplementary text collection unit 241. The supplementary text collection unit 241 will be described later. The data of the word DB 212 is supplied by a conversion keyword selection unit 242. The conversion keyword selection unit 242 will be described later. The data of the word relation DB 213 is also supplied by the conversion keyword selection unit 242.

[0023] The conversion pattern information DB220 stores data associated with a conversion method of an input text, among data required for converting an input text and visualizing a result of a query made to the QA system 200. The conversion pattern information DB220 includes a conversion pattern DB221 that stores candidates for a text conversion method, and a template DB222 that stores a model of an evaluation query text used for calculating the certainty of a conversion pattern. The detailed configuration of the data in the conversion pattern DB221 and the template DB222 will be described later with reference to Figs. 6 and 7.

[0024] The data of the conversion pattern DB 221 is supplied by a conversion pattern generation unit 245 and a conversion pattern confidence measurement unit 247. The conversion pattern generation unit 245 and the conversion pattern confidence measurement unit 247 will be described later. The data of the template DB 222 is supplied by a conversion pattern evaluation data generation unit 246. The conversion pattern evaluation data generation unit 246 will be described later.

[0025] The visualization result DB 230 stores data related to visualization of the query result among data required to convert the input text and visualize the query result made to the QA system 200. The visualization result DB 230 includes an answer text selection record DB 231 that stores answer text information selected by the user on the query result visualization screen. The detailed structure of the data in the answer text selection record DB 231 will be described later with reference to FIG. 9.

[0026] The data in DB 231 is supplied by a query result visualization unit 250. The query result visualization unit 250 will be described later.

[0027] The complemented text collection unit 241 receives an input text from the user terminal 201, and extracts complemented text that complements the input text. The complemented text collection unit 241 stores the input text and the complemented text in the input text DB 211.

[0028] The conversion keyword selection unit 242 extracts words and word relation information to be converted using the input text DB 211. The conversion keyword selection unit 242 stores the extracted words and word relation information in the word DB 212 and the word relation DB 213, respectively.

[0029] The conversion pattern selection unit 243 uses the conversion pattern DB 221 to select a conversion method for the input text.

[0030] The QA system inquiry unit 244 converts the input text using the input text DB 211 and the conversion pattern DB 221 , and makes an inquiry to the QA system 200 .

[0031] The conversion pattern generation unit 245 extracts word combinations that satisfy the conditions of the conversion pattern, and extracts word relation information that indicates the relationship between the words. The conversion pattern generation unit 245 stores the extracted word combinations and word relation information in the conversion pattern DB 221.

[0032] The conversion pattern evaluation data generation unit 246 acquires pairs of question texts and answer texts including quantitative reasoning, converts the quantitative expressions into variables, and extracts words that satisfy the application conditions of the conversion pattern. The conversion pattern evaluation data generation unit 246 stores the extracted results in the template DB 222.

[0033] The conversion pattern certainty measurement unit 247 evaluates the certainty of the conversion pattern by using the conversion pattern DB 221 and the template DB 222. The conversion pattern certainty measurement unit 247 stores the evaluation result of the certainty of the conversion pattern in the conversion pattern DB 221.

[0034] The answer text reverse conversion unit 248 receives the result of an inquiry to the QA system 200 and reverse converts the answer text using the conversion pattern DB 221 .

[0035] The query result visualization unit 250 visualizes the query results to the QA system 200 using the results of the inverse conversion performed by the answer text inverse conversion unit 248 .

[0036] (Data structure of each DB) Next, the configuration of data stored in the input text DB 211, the word DB 212, the word relation DB 213, the conversion pattern DB 221, the template DB 222, and the answer text selection record DB 231 will be described.

[0037] (Configuration of data 300 in input text DB211) 3 is a diagram showing an example of the configuration of data 300 in the input text DB 211. The data 300 in the input text DB 211 is composed of one or more "input text DB" records. The "input text DB" record is composed of multiple fields, including "input text ID", "input text", "topic", "complementary text", "source external document", and "applied conversion pattern ID".

[0038] "Input text ID" is identification information for a record. "Input text" holds text information for the query text received from the user terminal 201. "Topic" holds information related to the topic to which the "input text" belongs. "Complementary text" holds text information that complements the contents of the "input text". "Source external document" holds information related to the information source from which the "complementary text" was collected. "Applied conversion pattern ID" holds ID information for the conversion pattern to be applied.

[0039] (Data 400 in Word DB212) 4 is a diagram showing an example of the configuration of data 400 in the word DB 212. Data 300 in the word DB 212 is composed of one or more "word" records. A "word" record is composed of multiple fields, such as "word ID", "corresponding input text ID", "word content", "part of speech information", and "embedded expression".

[0040] "Word ID" is identification information of a record. "Corresponding input text ID" holds ID information of a record in the input text DB 211. "Word content" holds text information of a word. "Part of speech information" holds information related to the part of speech of the corresponding word. "Embedded representation" holds information related to the embedded representation obtained from the corresponding word.

[0041] (Data 500 of word relation DB213) 5 is a diagram showing an example of the configuration of data 500 in the word relationship DB 213. The data 500 in the word relationship DB 213 is composed of one or more "word relationship" records. The "word relationship" record is composed of multiple fields, namely, "word relationship ID", "word relationship information", "first word ID", "first word content", "second word ID", and "second word content".

[0042] The "word relationship ID" is record identification information. The "word relationship information" holds information about the relationship between words identified by the "first word ID" and "first word content" and the "second word ID" and "second word content," respectively.

[0043] The "First Word ID" holds ID information of one of a pair of words having a relationship shown in the "Word Relationship Information" in the word DB 212. The "First Word Content" holds text information of one of a pair of words having a relationship shown in the "Word Relationship Information".

[0044] The "second word ID" holds ID information of the other word of a word pair having a relationship shown in the "word relationship information" in the word DB 212. The "second word content" holds text information of the other word of a word pair having a relationship shown in the "word relationship information".

[0045] (Data 600 in conversion pattern DB214) 6 is a diagram showing an example of the configuration of data 600 in the conversion pattern DB 214. The data 600 in the conversion pattern DB 221 is composed of one or more "conversion pattern" records. The "conversion pattern" record is composed of multiple fields, such as "conversion pattern ID", "first word", "second word", "word relation information", and "confidence level".

[0046] "Conversion pattern ID" is identification information for a record. "First word" holds information regarding the converted word of the conversion pattern. "Second word" holds information regarding the converted word of the conversion pattern. "Word relationship information" holds information regarding the relationship between the "first word" and the "second word". "Confidence" holds information regarding the confidence that indicates the effectiveness of the conversion pattern indicated by the corresponding record. The confidence represents the degree of improvement in bringing the answer text (output text) output by the QA system 200 closer to the correct answer.

[0047] "Word relationship information" of "hypernym is vehicle" indicates that, for example, when the first word and the second word are both words that represent vehicles, the hypernym that represents the higher concept is "vehicle." "Word relationship information" of "embedded representation is [5,4,3,1,...]" indicates that the first word and the second word are both the same embedded representation. "Word relationship information" of "antonym" indicates that the first word and the second word are "synonyms" of each other. "Word relationship information" indicating that the first word and the second word are "synonyms" may also be "synonyms." Various "word relationship information" can be defined for two words that have a relationship that allows the degree of certainty to be measured.

[0048] (Configuration of data 700 in template DB222) 7 is a diagram showing an example of the configuration of data 700 in template DB 222. Data 700 in template DB 222 is composed of one or more "template" records. The "template" record is composed of multiple fields, namely, "template ID," "template question," "template answer," "logical relationship," "first word to be converted," "second word to be converted," and "word relationship information."

[0049] "Template ID" is identification information for a record. "Template Question" holds the question text of the template identified by the corresponding record. "Template Answer" holds the answer text of the template identified by the corresponding record. "Logical Relationship" holds information regarding the logical relationship between variables included in the "Template Question" and "Template Answer" of the template identified by the corresponding record.

[0050] "First word to be converted" holds information about words to be converted in the "template question". "Second word to be converted" holds information about words to be converted in the "template question". "Word relationship information" holds information about the relationship between the "first word to be converted" and the "second word to be converted". The "word relationship information" is the same as the "word relationship information" in conversion pattern DB 214 (Figure 6).

[0051] (Data 800 of Answer Text Selection Record DB231) 8 is a diagram showing an example of the configuration of data 800 in answer text selection record DB 231. Data 800 in answer text selection record DB 231 is composed of one or more "answer text selection record" records. The "answer text selection record" record is composed of multiple fields, namely, "record ID", "input text", "used conversion pattern ID", "answer text", and "selection result".

[0052] "Record ID" is identification information of a record. "Input text" holds information about the input text entered into the QA system 200 for a query. "Used conversion pattern ID" holds ID information of the applied conversion pattern. "Answer text" holds information about the answer text obtained from the QA system 200. "Selection result" holds information about whether the "answer text" was selected by the user as the final answer on the visualization interface.

[0053] (Operation of the text conversion system S according to the first embodiment) FIG. 9 is a flowchart showing the operation procedure of the text conversion system S according to the first embodiment.

[0054] First, in step S901, the text conversion system S executes a conversion pattern generation process upon receiving an input from the user terminal 201. The details of the conversion pattern generation process will be described later with reference to FIG.

[0055] Next, in step S902, the text conversion system S executes a conversion pattern evaluation data generation process for generating data used for evaluating the confidence of a conversion pattern. The conversion pattern evaluation data generation process will be described in detail later with reference to FIG.

[0056] Next, in step S903, the text conversion system S executes a conversion pattern certainty measurement process. The conversion pattern certainty measurement process will be described in detail later with reference to FIG.

[0057] Next, in step S904, the text conversion system S executes a complementary text collection process for collecting text that complements the input text. The complementary text collection process will be described in detail later with reference to FIG.

[0058] Next, in step S905, the text conversion system S executes a conversion keyword selection process for selecting words to be converted. The details of the conversion keyword selection process will be described later with reference to FIG.

[0059] Next, in step S906, the text conversion system S executes a conversion pattern selection process for selecting a conversion method. The details of the conversion pattern selection process will be described later with reference to FIG.

[0060] Next, in step S907, the text conversion system S executes a QA system query process to convert the input text and make a query to the QA system 200. Details of the QA system query process will be described later with reference to FIG.

[0061] Next, in step S908, the text conversion system S executes an answer text reverse conversion process for converting the obtained answer text back to the words before conversion. Details of the answer text reverse conversion process will be described later with reference to FIG.

[0062] Next, in step S909, the text conversion system S executes a query result visualization process to visualize the answer text. The details of the query result visualization process will be described later with reference to FIG.

[0063] Finally, in step S910, the text conversion system S determines whether there is additional input from the user terminal 201. If there is additional input from the user terminal 201 (S910 YES), the text conversion system S returns the process to step S904, and if there is no additional input (S910 NO), the text conversion system S ends the text conversion process.

[0064] (Details of conversion pattern generation process) FIG. 10 is a flowchart showing the conversion pattern generation process (step S901 in FIG. 9).

[0065] First, in step S901a, the conversion pattern generation unit 245 acquires a list of words for the target language. The list may be acquired by referring to a dictionary or by collecting documents on the Internet and constructing the list by dividing the words. The Internet is an example of a second external information source.

[0066] Next, in step S901b, the conversion pattern generation unit 245 extracts words having part-of-speech information that satisfies the conditions of the conversion pattern from the list of words acquired in step S901a. For example, there is a method of extracting only nouns, verbs, and adjectives as parts of speech that satisfy the conditions of the conversion pattern.

[0067] Next, in step S901c, the conversion pattern generation unit 245 extracts words having generality that satisfy the conditions of the conversion pattern from the list of words acquired in step S901a. The generality of a word can be determined by determining that a word has a high frequency of appearance on the Internet above a certain level, or a word to which a named entity label such as time or place is assigned in named entity extraction, as being general. However, the method is not limited to these, and any method that determines that a word is general based on a predetermined criterion may be used.

[0068] Next, in step S901d, the conversion pattern generation unit 245 determines whether there is a word combination for which no related information has been extracted from the set of words extracted in steps S901b and S901c. If there is a word combination for which no related information has been extracted (step S901d YES), the conversion pattern generation unit 245 moves the process to step S901e. On the other hand, if there is no word combination for which no related information has been extracted (step S901d NO), the conversion pattern generation unit 245 ends the conversion pattern generation process.

[0069] In step S901e, the conversion pattern generation unit 245 extracts relationship information between words for the word combinations for which it has been determined in step S901d that the relationship information has not been extracted. The relationship information may be expressed as category information such as "synonyms" or "antonyms" by referring to an ontology such as Wordnet, or may be expressed in vector form by acquiring distributed representations of words and measuring the distance between them. Next, in step S901f, the conversion pattern generation unit 245 stores the relationship information and word combinations extracted in step S901e in the conversion pattern DB 221. When step S901f ends, the conversion pattern generation unit 245 returns the process to step S901d.

[0070] (Details of conversion pattern evaluation data generation process) FIG. 11 is a flowchart showing the conversion pattern evaluation data generation process (step S902 in FIG. 9).

[0071] First, in step S902a, the conversion pattern evaluation data generation unit 246 acquires a pair of question text and answer text including quantitative inference. The pair of question text and answer text including quantitative inference can be collected, for example, using a school examination question book or the Internet. The school examination question book and the Internet here are examples of a third external information source. The pair of question text and answer text including quantitative inference is, for example, the following data. [Question text] "From Naha Airport, you can get to Shuri in 30 minutes by Okinawa Airport Shuttle, or in 15 minutes by Yui Rail. Okinawa Churaumi Aquarium is a 20-minute walk from Shuri. What is the shortest travel time to Okinawa Churaumi Aquarium?" [Answer text] 35 minutes

[0072] Next, in step S902b, the conversion pattern evaluation data generation unit 246 identifies the quantitative expressions in the question text and the answer text. In the case of the pair of question text and answer text including the quantitative inference illustrated in step S902a, "30 minutes", "15 minutes", "20 minutes", and "35 minutes" correspond to the quantitative expressions. Methods that use named entity extraction and regular expressions can be used to identify the quantitative expressions.

[0073] Next, in step S902c, the conversion pattern evaluation data generating unit 246 estimates a logical relationship between the quantitative expressions identified in step S902b. In the case of a pair of question text and answer text including quantitative inference as exemplified in step S902a, it is estimated that there is a logical relationship of "35 minutes = 15 minutes + 20 minutes".

[0074] Next, in step S902d, the conversion pattern evaluation data generation unit 246 converts the quantitative expressions whose logical relations have been estimated in step S902c into variables. In the case of a pair of question text and answer text including a quantitative inference as exemplified in step S902a, the following substitutions are made by the conversion into variables: [Question text] "You can get to Shuri from Naha Airport in xx minutes by Okinawa Airport Shuttle, or in yy minutes by Yui Rail. Okinawa Churaumi Aquarium is zz minutes away on foot from Shuri. What is the shortest travel time to Okinawa Churaumi Aquarium?" [Answer text]ww minutes Also, the logical relationship is replaced with "ww=yy+zz".

[0075] Next, in step S902e, the conversion pattern evaluation data generation unit 246 extracts words having part-of-speech information that satisfies the application conditions of the conversion pattern from the question text and answer text in which the quantitative expressions have been made into variables in step S903d. Step S902e is the same as step S901b. Next, in step S902f, the conversion pattern evaluation data generation unit 246 extracts words having generality that satisfies the application conditions of the conversion pattern. Step S902f is the same as step S901b. Next, in step S902g, the conversion pattern evaluation data generation unit 246 extracts relationship information between the words extracted in steps S902e and S902f.

[0076] Next, in step S902g, the conversion pattern evaluation data generation unit 246 stores in the template DB 222 the question text and answer text that were made into variables in step S902d, the logical relationships between the variables estimated in step S902c, and the words and relationships between the words that satisfy the application conditions of the conversion patterns extracted in steps S902e, S902f, and S902g.

[0077] In this way, by extracting the words to be converted and the relationship information between the words from the question text and answer text containing quantitative inference, and using them as templates for the question text and answer text when calculating the confidence level, the accuracy of the quantitative inference results of the QA system 200 can be improved.

[0078] (Details of conversion pattern confidence measurement process) FIG. 12 is a flowchart showing the conversion pattern certainty measurement process (step S903 in FIG. 9).

[0079] First, in step S903a, the conversion pattern certainty measurement unit 247 extracts a conversion pattern record using the conversion pattern DB 221. Next, in step S903b, the conversion pattern certainty measurement unit 247 determines whether or not there is a conversion pattern record for which the certainty has not been measured among the conversion pattern records extracted in step S903a. If there is a conversion pattern record for which the certainty has not been measured (YES in step S903b), the conversion pattern certainty measurement unit 247 transfers the process to step S903c. On the other hand, if there is no conversion pattern record for which the certainty has not been measured (NO in step S903b), the conversion pattern certainty measurement unit 247 ends the conversion pattern certainty measurement process.

[0080] In step S903c, the conversion pattern certainty measurement unit 247 uses the template DB 222 to extract template records having the same word relation information as the conversion pattern for which it has been determined in step S903b that the certainty has not been measured.

[0081] Next, in step S903d, the conversion pattern certainty measurement unit 247 determines whether there is a template record that has not been used to measure the certainty. If there is a template record that has not been used to measure the certainty (step S903d YES), the conversion pattern certainty measurement unit 247 moves the process to step S903e. On the other hand, if there is no template record that has not been used to measure the certainty (step S903d NO), the conversion pattern certainty measurement unit 247 returns the process to step S903b.

[0082] In step S903e, the conversion pattern certainty measurement unit 247 creates question text and answer text before conversion and question text and answer text after conversion, using the template record determined in step S903d as not being used to measure the certainty and the conversion pattern record determined in step S903b as not having been used to measure the certainty.

[0083] For example, using the template with template ID=1 in FIG. 7 and the conversion pattern with conversion pattern ID=1 in FIG. 6, substituting 30 minutes for the variable xx, 10 minutes for the variable yy, 15 minutes for the variable zz, and 25 minutes for the variable ww will result in the following question text and answer text before conversion, and the question text and answer text after conversion. [Question text before conversion] "From Naha Airport, you can get to Shuri in 30 minutes by Okinawa Airport Shuttle, or in 10 minutes by Yui Rail. Okinawa Churaumi Aquarium is a 15-minute walk from Shuri. What is the shortest travel time to Okinawa Churaumi Aquarium?" [Answer text before conversion] 25 minutes [Question text after conversion] "You can get to Shuri from Naha Airport in 30 minutes by bus or 10 minutes by train. Okinawa Churaumi Aquarium is a 15-minute walk from Shuri. What is the shortest travel time to Okinawa Churaumi Aquarium?" [Converted answer text] 25 minutes

[0084] Next, in step S903f, the conversion pattern certainty measurement unit 247 inputs the two question texts before and after conversion created in step S903e to the QA system 200 and compares the answers to the two question texts before and after conversion. In the comparison result of the two question texts before and after conversion and the answer text exemplified in step S903e, if the answer obtained with the question text before conversion is "40 minutes" and the answer obtained with the question text after conversion is "25 minutes", the question text after conversion provides a correct answer or an answer closer to the correct answer, so it can be determined that this conversion pattern improves the accuracy of queries to the QA system 200.

[0085] Next, in step S903g, the conversion pattern certainty measurement unit 247 measures the certainty and updates the conversion pattern DB 214. In the example case compared in step S903f, the converted question text was 15 minutes closer to the correct answer, so the value of the "certainty" field of the conversion pattern is increased by 15 points, and the certainty is assigned so that the closer the answer is to the correct answer, the higher the certainty. However, the method of calculating the certainty is not limited to this. In addition, the certainty may be calculated by trying multiple patterns of combinations of values ​​to be substituted for variables in steps S903e, S903f, and S903g, and may be a statistic such as the total value or average value of the changes in the answers obtained by these trials.

[0086] Upon completion of step S903g, the conversion pattern certainty measurement unit 247 advances the process to step S903d.

[0087] (Details of the completion text collection process) FIG. 13 is a flowchart showing the complementary text collection process (step S904 in FIG. 9).

[0088] First, in step S904a, the supplemental text collection unit 241 receives a query text for the QA system 200 input from the user terminal 201. Next, in step S904b, the supplemental text collection unit 241 identifies a topic to which the input text received in step S904a belongs. The topic can be identified by using a classification model that estimates topic information such as "politics" or "health" for the input text, or by using a dictionary that records words corresponding to each topic.

[0089] Next, in step S904c, the supplemental text collection unit 241 collects external documents that include descriptions related to the topic identified in step S904b. One method for collecting external documents is to crawl the text of sites related to the topic from the Internet. The Internet, which collects external documents, is an example of a first external information source. Next, in step S904d, the supplemental text collection unit 241 associates the "input text," "topic," "supplemental text," and "source external document" based on the external document acquired in step S904c, and stores them in the input text DB 211.

[0090] By using complementary text in this way, it may be possible to further improve the accuracy of the answer of the output text of the QA system 200 to the input text.

[0091] (Details of conversion keyword selection process) FIG. 14 is a flowchart showing the conversion keyword selection process (step S905 in FIG. 9).

[0092] First, in step S905a, the conversion keyword selection unit 242 extracts an input text record from the input text DB 211. Next, in step S905b, the conversion keyword selection unit 242 extracts a word having part-of-speech information that satisfies the application condition of the conversion pattern from the input text record extracted in step S905a. Step S905b is the same as step S901b. Next, in step S905c, the conversion keyword selection unit 242 extracts a word having generality that satisfies the application condition of the conversion pattern from the input text record extracted in step S905a. Step S905c is the same as step S901c.

[0093] Next, in step S905d, the conversion keyword selection unit 242 determines whether there is a word combination from which related information has not been extracted among the words extracted in steps S905b and S905c. If there is a word combination from which related information has not been extracted (step S905d YES), the conversion keyword selection unit 242 moves the process to step S905e. On the other hand, if there is no word combination from which related information has not been extracted (step S905d NO), the conversion keyword selection unit 242 ends the conversion keyword selection process.

[0094] In step S905e, the conversion keyword selection unit 242 extracts relationship information between the words for which it has been determined in step S905d that relationship information has not been extracted. The extraction of relationship information in step S905e is similar to that in step S901e.

[0095] Next, in step S905f, the converted keyword selection unit 242 stores the word combination for which the relationship information was extracted in step S905e and this relationship information in the word DB 212 and the word relationship DB 213. Upon completing step S905f, the converted keyword selection unit 242 returns the process to step S906.

[0096] (Details of conversion pattern selection process) FIG. 15 is a flowchart showing the conversion pattern selection process (step S906 in FIG. 9).

[0097] First, in step S906a, the conversion pattern selection unit 243 extracts an input text record and a word relation record by using the input text DB 211 and the word relation DB 213. Next, in step S906b, the conversion pattern selection unit 243 extracts, from the conversion pattern DB 221, a conversion pattern record having the same word relation information as the word relation record extracted in step S906a.

[0098] Next, in step S906c, the conversion pattern selection unit 243 selects a conversion pattern to be applied to the input text extracted in step S906a from among the conversion pattern records extracted in step S906b. The conversion pattern to be applied to the input text can be selected by selecting the record with the highest confidence value. Since the conversion pattern is selected based on the confidence in this way, the words of the input text can be efficiently converted so that the accuracy of the answers in the output text of the QA system 200 is more likely to be improved.

[0099] Next, in step S906d, the "applied conversion pattern ID" in the input text DB 211 is updated with the ID of the conversion pattern selected in step S906c.

[0100] (Details of QA system inquiry processing) FIG. 16 is a flowchart showing the QA system inquiry process (step S907 in FIG. 9).

[0101] First, in step S907a, the QA system inquiry unit 244 extracts input text records, word records, and conversion pattern records from the input text DB 211, the word DB 212, and the conversion pattern DB 221, respectively. Next, in step S907b, the QA system inquiry unit 244 selects a source word and a destination word. In step S907b, it is determined which word to convert the first word of the conversion pattern record into, and which word to convert the second word into. One method of determination is to select words with the closest embedded representations as a combination of source and destination words.

[0102] Next, in step S907c, the QA system inquiry unit 244 converts the input text using the selection result in step S907b (generating converted input text). Next, in step S907d, the QA system inquiry unit 244 inputs the input text converted in step S907c to the QA system 200 to make an inquiry. In step S907c, the QA system inquiry unit 244 inputs the input text converted in step S907c and the complementary text collected in step S904 to the QA system 200.

[0103] Next, in step S907e, the QA system inquiry unit 244 acquires from the QA system 200 an answer text (output text) to the input text input in step S907d.

[0104] (Details of the answer text reverse conversion process) FIG. 17 is a flowchart showing the answer text reverse conversion process (step S908 in FIG. 9).

[0105] First, in step S908a, the answer text reverse conversion unit 248 extracts input text records, word records, and conversion pattern records from the input text DB 211, the word DB 212, and the conversion pattern DB 221, respectively. Next, in step S908b, the answer text reverse conversion unit 248 selects source words and destination words in the same way as in step S907b. Next, in step S908c, the answer text reverse conversion unit 248 converts the answer text acquired in step S907e (FIG. 16) using the source words and destination words selected in step S908b.

[0106] (Example of query result visualization processing results) 18 is a diagram showing an example of a display of the query result visualization process (step S909 in FIG. 9). Query result visualization unit 250 displays query result visualization result 250D as shown in FIG.

[0107] Query result visualization result 250D includes a query text display 250D1 and a complement text display 250D2. Query result visualization result 250D also includes a conversion pattern candidate display 250D31 and a conversion pattern candidate display 250D32. The number of conversion pattern candidate displays is not limited to two.

[0108] Conversion pattern candidate display 250D31 shows that the certainty of conversion pattern ID:1, which converts the word "opening day" in query text display 250D1 to "A" and "closing day" to "B (other than A)", is "0.5". Conversion pattern candidate display 250D31 also shows that when question text in which words have been converted using this conversion pattern is input to QA system 200, answer text "Closing days are Saturday and Sunday" is obtained.

[0109] On the other hand, conversion pattern candidate display 250D32 shows that the confidence level of conversion pattern ID:2, which converts the word "opening days" in query text display 250D1 to "business days" and "closing days" to "holidays", is "0.8". Conversion pattern candidate display 250D32 also shows that if question text with words converted using this conversion pattern is input to QA system 200, the answer text "Closing days are Wednesday, Thursday, Friday, Saturday, and Sunday" will be obtained.

[0110] Comparing the conversion pattern candidate display 250D31 and the conversion pattern candidate display 250D32, the word conversion based on the conversion pattern candidate display 250D32 has a higher degree of certainty, and the answer text is closer to the correct one. Therefore, the user who operates the user terminal 201 presses the selection button 250D41 corresponding to the conversion pattern candidate display 250D31 and the selection button 250D42 corresponding to the conversion pattern candidate display 250D32, and the conversion pattern with the conversion pattern ID: 2 is selected.

[0111] In this way, a plurality of candidates for output text output from the QA system 200 are visualized for the user on a predetermined display screen. Then, the user can perform a selection operation on the predetermined display screen to determine one or more output texts to be output from the QA system 200 from the plurality of candidates.

[0112] According to the first embodiment, by rewriting the input text before making a query to the QA system, it is possible to improve the accuracy of answers to questions that include calculations of the four arithmetic operations or operations related to sets.

[0113] (Modification of the first embodiment) In the first embodiment, words included in sentences related to the quantities included in the question text and answer text are converted into other words. However, the conversion is not limited to words, and phrases included in sentences related to the quantities included in the question text and answer text may be converted into other phrases. In other words, the conversion can be carried out by replacing "words" in the first embodiment with "phrases".

[0114] In the first embodiment, step S901b of the conversion pattern generation process (FIG. 10) "extracts words having part-of-speech information that satisfies the conditions of the conversion pattern," and step S901c "extracts words having generality that satisfies the conditions of the conversion pattern." However, this is not limiting, and either one of steps S901b and S901c may be executed.

[0115] In the first embodiment, in step S902e of the conversion pattern evaluation data generation process (FIG. 11), "words having part-of-speech information that satisfies the application conditions of the conversion pattern are extracted," and in step S902f, "words having uniqueness that satisfies the application conditions of the conversion pattern are extracted." However, this is not limiting, and either one of steps S902e and S902f may be executed.

[0116] In the first embodiment, step S905b of the conversion keyword selection process (FIG. 14) "extracts words having part-of-speech information that satisfies the application condition of the conversion pattern," and step S905c "extracts words having uniqueness that satisfies the application condition of the conversion pattern." However, this is not limiting, and either step S905b or S905c may be executed.

[0117] By diversifying the criteria for extracting words from a population in this way, the accuracy of converting words in text can be improved.

[0118] [Embodiment 2] In the first embodiment, as shown in steps S902a to S902d of the conversion pattern evaluation data generation process (FIG. 11), words contained in sentences related to quantities contained in the question text and answer text are converted to more general words based on relationships such as antonyms and synonyms. Then, the input text entered when making a query to the QA system 200 is converted to a more general expression using the conversion to words that have a higher degree of improvement (certainty) in the answer from the QA system 200 to the question text.

[0119] However, conversion to a more general expression based on relationships such as antonyms and synonyms is not limited to words contained in sentences related to quantity, and may be words contained in sentences that are not related to quantity. This example will be described below as embodiment 2. In the description of embodiment 2, differences from embodiment 1 will be mainly described, and descriptions of the same parts will be omitted.

[0120] That is, in the conversion pattern evaluation data generation process (FIG. 19) of the second embodiment, step S902aB is executed instead of step S902a, as compared with the conversion pattern evaluation data generation process (FIG. 11) of the first embodiment. In step S902aB, the conversion pattern evaluation data generation unit 246 acquires a pair of question text and answer text that does not include a quantitative inference.

[0121] In the conversion pattern evaluation data generation process of the second embodiment, the execution of steps S902b to S902d is canceled, as compared with the conversion pattern evaluation data generation process of the first embodiment. That is, in the second embodiment, since there is no logical relationship of the extracted quantity, "NULL" is stored in the "logical relationship" of the template DB 222.

[0122] In step S903g of the conversion pattern certainty measurement process (FIG. 12) in the second embodiment, each sentence vector of the answer text for the question text before and after word conversion is calculated, and the degree of change in the sentence vector is measured as the certainty. How close the answer text is to the correct answer is represented, for example, by the inner product with the sentence vector of the correct answer (cosine similarity). Note that the index representing how close the answer text is to the correct answer is not limited to cosine similarity.

[0123] In cosine similarity, if the dot product between the sentence vector of the answer text and the sentence vector of the correct answer is "1", the two sentences are a perfect match. If the dot product between the sentence vector of the answer text and the sentence vector of the correct answer is "-1", the two sentences are a perfect mismatch. If the dot product between the sentence vector of the answer text and the sentence vector of the correct answer is greater than "-1" and less than "1", the closer it is to "1", the higher the similarity between the two sentences. The value of the "confidence" field of the corresponding conversion pattern is increased by the value of the dot product between the sentence vector of the answer text to the question text before the word conversion and the sentence vector of the answer text to the question text after the word conversion and the correct answer, which is closer to "1". Or, the value of the "confidence" field of the corresponding conversion pattern is decreased by the value of the dot product between the sentence vector of the answer text to the question text before the word conversion and the sentence vector of the answer text to the question text after the word conversion and the correct answer, which is closer to "-1".

[0124] As described above, in the second embodiment, expressions of sentences included in question texts and answer texts that do not include sentences related to quantity are converted to more general expressions based on relationships such as antonyms and synonyms.Then, by converting to expressions that have a higher degree of improvement (certainty) of the answer from the QA system 200 to the question text, the input text to be input when making an inquiry to the QA system 200 is converted to a more general expression.

[0125] (Modification of the second embodiment) Also, the first embodiment may be applied to question texts and answer texts including sentences related to quantity, and the second embodiment may be applied to question texts and answer texts not including sentences related to quantity. In this way, regardless of whether the question texts and answer texts include sentences related to quantity or not, the question texts for the QA system 200 can be appropriately converted to obtain a correct answer or an answer text closer to the correct answer from the QA system 200.

[0126] Although the embodiments of the present disclosure have been described above in detail, the present disclosure is not limited to the above-described embodiments, and various modifications can be made without departing from the spirit of the present disclosure. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and the present disclosure is not necessarily limited to those having all of the configurations described. In addition, it is possible to add, delete, or replace part of the configuration of the above-described embodiments with other configurations.

[0127] Furthermore, the above-mentioned configurations, functional units, processing units, etc. may be realized in part or in whole by hardware, for example, by designing them as integrated circuits. The above-mentioned configurations, functions, etc. may be realized in software by a processor interpreting and executing a program that realizes each function. Information such as the program, table, file, etc. that realizes each function can be stored in a memory, a storage device such as an HDD or SSD, or a recording medium such as an IC card, an SD card, or a DVD.

[0128] In addition, in each of the above figures, the control lines and information lines are shown as those considered necessary for the explanation, and do not necessarily show all the control lines and information lines in the actual implementation. For example, it may be considered that almost all the components are actually connected to each other.

[0129] The above-mentioned arrangement of each function and data of the user terminal 201, text information DB 210, conversion pattern information DB 220, visualization result DB 20, control unit 240, and query result visualization unit 250 is merely an example. The arrangement, integration, and distribution of each function and data can be changed to an optimal form in terms of hardware and software performance, processing efficiency, communication efficiency, and the like. [Explanation of symbols]

[0130] S: text conversion system, 100: computer, 110: CPU, 120: memory, 200: QA system.

Claims

1. A text conversion method executed by a text conversion system that converts an input text including a question into an answer generation system that outputs an output text including an answer to the question when the input text includes a question, comprising the steps of: The text conversion system includes a memory and a processor. The processor, extracting from the input text a plurality of words or a plurality of phrases that satisfy a predetermined condition, extracting relationship information indicating a relationship between the combinations of the words or the combinations of the phrases, and specifying the combinations of the words or the combinations of the phrases to be converted in the input text based on the relationship information between the combinations of the words or the relationship information between the combinations of the phrases; Identifying a conversion target word after converting the word or a conversion target phrase after converting the phrase based on the relationship information of the combination of words or the relationship information of the combination of phrases; generating a converted input text by converting the word or phrase included in the input text with the conversion target word or the conversion target phrase; inputting the converted input text into the answer generation system; obtaining the output text outputted from the answer generation system in response to the converted input text; The conversion target word or the conversion target phrase included in the output text is converted to the word or phrase before conversion. A text conversion method comprising the steps of:

2. 2. A text conversion method according to claim 1, comprising: The processor, collecting complementary text from a first external information source that complements the input text with respect to a topic of the input text; The completion text is input to the answer generation system together with the converted input text, and the output text for the converted input text and the completion text is obtained from the answer generation system. A text conversion method comprising the steps of:

3. 2. A text conversion method according to claim 1, comprising: The processor, extracting a plurality of the words or a plurality of the phrases that satisfy the predetermined condition from the words or the phrases collected from a second external information source, extracting the relationship information of the word combination or the relationship information of the phrase combination, and generating a word conversion pattern record that associates the word combination with the relationship information of the word combination, or a phrase conversion pattern record that associates the phrase combination with the relationship information of the phrase combination, Collecting question text and answer text to the question text from a third external information source; extracting the words or phrases that satisfy the predetermined condition from the question text and the answer text, extracting the relationship information of the word combination or the relationship information of the phrase combination, and generating a word template record that associates the question text, the answer text, the word combination, and the relationship information of the word combination, or a phrase template record that associates the question text, the answer text, the phrase combination, and the relationship information of the phrase combination, extracting from the template a record having the same relationship information as the conversion pattern, converting the word combination or the phrase combination associated with the pre-conversion question text and the pre-conversion answer text in the record into the word combination or the phrase combination in the conversion pattern record having the same relationship information, thereby generating the converted question text, and inputting the converted question text to the answer generation system; based on a comparison result between the converted answer text output from the answer generation system for the converted question text and the answer text before conversion, a confidence level indicating a degree of improvement that the conversion of the combination of words or the combination of phrases in the question text before conversion using the conversion pattern will improve the answer text output by the answer generation system; storing the confidence level in association with the conversion pattern; Identifying the converted word or the converted phrase based on the degree of certainty from the combination of the words or the combination of the phrases associated in the conversion pattern. A text conversion method comprising the steps of:

4. 4. A text conversion method according to claim 3, comprising: The predetermined condition is satisfied when the word or phrase corresponds to a specific part of speech and / or when the word or phrase is general.

13. A text conversion method comprising:

5. 4. A text conversion method according to claim 3, comprising: The processor, collecting, from the third external information source, the question text including an inference about a quantity and the answer text to the question text; Inferring logical relationships between the quantitative expressions involved in the inferences about said quantities; Converting the quantitative expressions into variables, Extracting the words or phrases that satisfy the predetermined conditions from the question text and the answer text in which the quantitative expressions are converted into variables. A text conversion method comprising the steps of:

6. 2. A text conversion method according to claim 1, comprising: The processor, visualizing the plurality of output text candidates output from the answer generation system to a user on a predetermined display screen; The output text to be output from the answer generation system is determined from the candidates in response to a selection operation on the predetermined display screen by the user. A text conversion method comprising the steps of:

7. A text conversion system that converts an input text including a question into an output text including an answer to the question, the output text being input to an answer generation system that outputs the output text including an answer to the question, The text conversion system includes a memory and a processor. The processor, extracting from the input text a plurality of words or a plurality of phrases that satisfy a predetermined condition, extracting relationship information indicating a relationship between the combinations of the words or the combinations of the phrases, and specifying the combinations of the words or the combinations of the phrases to be converted in the input text based on the relationship information between the combinations of the words or the relationship information between the combinations of the phrases; Identifying a conversion target word after converting the word or a conversion target phrase after converting the phrase based on the relationship information of the combination of words or the relationship information of the combination of phrases; generating a converted input text by converting the word or phrase included in the input text with the conversion target word or the conversion target phrase; inputting the converted input text into the answer generation system; obtaining the output text outputted from the answer generation system in response to the converted input text; The conversion target word or the conversion target phrase included in the output text is converted to the word or phrase before conversion. A text conversion system comprising:

8. 8. A text conversion system according to claim 7, comprising: The processor, collecting complementary text from a first external information source that complements the input text with respect to a topic of the input text; The completion text is input to the answer generation system together with the converted input text, and the output text for the converted input text and the completion text is obtained from the answer generation system. A text conversion system comprising:

9. 8. A text conversion system according to claim 7, comprising: The processor, extracting a plurality of the words or a plurality of the phrases that satisfy the predetermined condition from the words or the phrases collected from a second external information source, extracting the relationship information of the word combination or the relationship information of the phrase combination, and generating a word conversion pattern record that associates the word combination with the relationship information of the word combination, or a phrase conversion pattern record that associates the phrase combination with the relationship information of the phrase combination, Collecting question text and answer text to the question text from a third external information source; extracting the words or phrases that satisfy the predetermined condition from the question text and the answer text, extracting the relationship information of the word combination or the relationship information of the phrase combination, and generating a word template record that associates the question text, the answer text, the word combination, and the relationship information of the word combination, or a phrase template record that associates the question text, the answer text, the phrase combination, and the relationship information of the phrase combination, extracting from the template a record having the same relationship information as the conversion pattern, converting the word combination or the phrase combination associated with the pre-conversion question text and the pre-conversion answer text in the record into the word combination or the phrase combination in the conversion pattern record having the same relationship information, thereby generating the converted question text, and inputting the converted question text to the answer generation system; based on a comparison result between the converted answer text output from the answer generation system for the converted question text and the answer text before conversion, a confidence level indicating a degree of improvement that the conversion of the combination of words or the combination of phrases in the question text before conversion using the conversion pattern will improve the answer text output by the answer generation system; storing the confidence level in association with the conversion pattern; Identifying the converted word or the converted phrase based on the degree of certainty from the combination of the words or the combination of the phrases associated in the conversion pattern. A text conversion system comprising:

10. 10. The text conversion system of claim 9, The predetermined condition is satisfied when the word or phrase corresponds to a specific part of speech and / or when the word or phrase is general. A text conversion system comprising:

11. 10. The text conversion system of claim 9, The processor, collecting, from the third external information source, the question text including an inference about a quantity and the answer text to the question text; Inferring logical relationships between the quantitative expressions involved in the inferences about said quantities; Converting the quantitative expressions into variables, Extracting the words or phrases that satisfy the predetermined conditions from the question text and the answer text in which the quantitative expressions are converted into variables. A text conversion system comprising:

12. 8. A text conversion system according to claim 7, comprising: The processor, visualizing the plurality of output text candidates output from the answer generation system to a user on a predetermined display screen; The output text to be output from the answer generation system is determined from the candidates in response to a selection operation on the predetermined display screen by the user. A text conversion system comprising:

Images