JP7313757B1 - Text generation device and text generation method - Google Patents

Abstract

Kind Code: A1 A prompt can be generated by adding effective sentences as reference information to an input question sentence within a predetermined character limit. SOLUTION: Based on a question text input by a user, the total number of characters combined with the number of characters of the question text does not exceed the limit of the number of characters that can be input to an answer generation unit 20 of a large scale language model (LLM) 1. An additional sentence generation unit 11 for generating additional sentences related to the question sentence with the number of characters, and a prompt is generated by adding the additional sentences generated by the additional sentence generation unit 11 to the input question sentence as reference information. and a prompt generation unit 12 for generating an additional sentence so that the number of characters does not exceed the character limit that can be input to the LLM 1 as a prompt, so that the input question sentence can be answered within the determined character limit. It is possible to generate a prompt to which effective sentences are added as reference information. [Selection drawing] Fig. 2

Description

TECHNICAL FIELD The present invention relates to a sentence generation device and a sentence generation method, and more particularly to a technology for generating sentences to be input as prompts to a large-scale language model.

Conventionally, systems have been provided that perform natural language processing such as question-and-answer sessions using large language models (LLMs). LLM is a model of natural language processing trained using a large amount of text data, takes sentences as input, and outputs sentences. When LLM is applied to a question-and-answer system, when a question sentence is input to LLM, an answer sentence is output from LLM. In LLM, an input text is called a "Prompt".

In recent years, the utilization of LLM has progressed, but there is a big problem with the accuracy of answers. That is, the behavior of the LLM confidently outputting an incorrect answer is regarded as a problem. As a means to solve this problem, adding information to the question sentence to generate a prompt is being performed. If there is additional information, the LLM will refer to the additional information to generate an answer to the question. For example, as shown in FIG. 13, by adding reference information to a question to create a prompt, it becomes easier to obtain an appropriate answer from the LLM considering the reference information.

Conventionally, as a technology for generating prompts using additional information, a method combining a Completion model and an Embedding model has been proposed. The Completion model is a model that inputs prompts including question sentences and reference sentences and outputs answer sentences generated based on them. The embedding model is a model that converts an input sentence into a numerical vector based on its meaning. If two sentences are close in meaning, the parameters of the model are adjusted so that the vectors generated from them are close. A method that combines these models is a style in which vectors of the Embedding model are used to search for information that is deeply related to the question, and the completion model is executed after the search results are reflected in the prompt.

A technique of adding reference information to an input question sentence to generate a prompt, and inputting the generated prompt into a language model to generate an answer sentence is also disclosed in Japanese Unexamined Patent Application Publication No. 2002-200310. Regarding adding information to a prompt, Patent Literature 1 discloses adding an instruction including an explanation about a chatbot character to the prompt, adding a character dialogue sample to the prompt, adding a dialogue scene obtained from a dialogue storage database in relation to the dialogue theme to the prompt, and replacing words included in question sentences with synonyms and the like and adding them to the prompt.

Further, Patent Documents 2 and 3 disclose that another question is generated from an input question and added to the prompt. Here, Patent Literature 2 discloses that another question sentence is generated based on the past answer history. Patent Literature 3 discloses that words included in a question sentence are replaced with shortened terms or the like to generate another question sentence.

Japanese Patent Application Laid-Open No. 2022-180282 Japanese Patent Publication No. 2001-519559 Japanese Patent Publication No. 2002-508863

As described above, inputting a prompt generated by adding reference information to a question text into the LLM is effective in that an appropriate answer text can be obtained more easily than when the reference information is not added. However, since the LLM of the Completion model has a limit on the number of input characters, there is a problem that sentences generated based on input question sentences cannot be added unlimitedly as reference information.

The present invention has been made to solve such problems, and aims to generate a prompt to which effective sentences are added as reference information to an input question sentence within the limit of the determined number of characters.

In order to solve the above-described problem, in the present invention, based on an input question sentence, the total number of characters including the number of characters of the question sentence does not exceed the limit of the number of characters that can be input to the large-scale language model. Here, candidate sentences related to the input question sentence are acquired, and additional sentences are generated by adjusting the acquired candidate sentences.

According to the present invention configured as described above, when generating a prompt by adding sentences related to an input question sentence, the additional sentence is generated so that the number of characters does not exceed the character limit that can be input to the large-scale language model as a prompt.

It is a figure which shows the structural example of the language model system provided with the sentence production|generation apparatus of this embodiment. It is a block diagram which shows the functional structural example of the sentence production|generation part (sentence production|generation apparatus) by this embodiment. FIG. 10 is a block diagram showing a functional configuration example of a candidate sentence obtaining unit that implements the first candidate sentence obtaining method; FIG. 10 is a diagram showing an example of a prompt configured using candidate sentences respectively acquired from three sentence DBs; FIG. 11 is a block diagram showing a functional configuration example of a candidate sentence obtaining unit that implements a second candidate sentence obtaining method; FIG. 11 is a block diagram showing a functional configuration example of a candidate sentence obtaining unit that implements a third candidate sentence obtaining method; FIG. 10 is a diagram showing an example of a prompt and another question sentence generated by applying the third candidate sentence acquisition method; FIG. 10 is a diagram showing another example of a prompt and another question sentence generated by applying the third candidate sentence acquisition technique; FIG. 12 is a block diagram showing a functional configuration example of a candidate sentence obtaining unit that implements a fourth candidate sentence obtaining method; FIG. 10 shows an example of prompts for generating question-answer pairs for supervised learning. FIG. 10 is a diagram showing two examples of candidate sentences adjusted by the first candidate sentence adjustment method; It is a flowchart which shows the operation example of the sentence production|generation part (sentence production|generation apparatus) by this embodiment. It is a figure which shows an example of the content which inputs into LLM the prompt produced|generated by adding reference information to the question sentence, and obtains an answer sentence.

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram showing a configuration example of a language model system 100 equipped with a sentence generation device of this embodiment. A language model system 100 of this embodiment comprises a large scale language model (LLM) 1 and a text database (DB) 2 .

The LLM 1 is connected to a user terminal 200 via a communication network 300 such as the Internet or a mobile phone network, inputs a question from the user terminal 200 , generates an answer to the question, and outputs the answer to the user terminal 200 . When the LLM 1 generates an answer sentence, it generates an additional sentence related to the question sentence and generates a prompt including the question sentence and the additional sentence.

The text DB2 is a database used when the LLM1 generates additional information based on the question text input from the user terminal 200. FIG. Although FIG. 1 shows a configuration in which the text DB2 is connected to the LLM1, the configuration is not limited to this. For example, part or all of the text DB 2 may be configured to be connected to the communication network 300 . In this case, the LLM1 accesses the text DB2 via the communication network 300 and obtains the necessary information.

The LLM 1 is software that performs natural language processing according to a model trained using a large amount of text data. The text generation unit 10 corresponds to the text generation device of this embodiment, and its function is realized by the operation of a program stored in a storage medium such as a computer's RAM, ROM, hard disk, or semiconductor memory under the control of the CPU. The function of the answer generation unit 20 is similarly realized.

The text generation unit 10 of the present embodiment generates a prompt for input to the LLM 1 by adding reference information to the question text input from the user terminal 200 . A prompt for input to the LLM 1 is specifically a prompt for input to the answer generator 20 . That is, the text generation part 10 generates an additional text based on the question text input from the user terminal 200, generates a text with the additional text added as reference information for the question text, and inputs the text to the answer generation part 20 as a prompt.

The answer generation unit 20 receives the prompt generated by the text generation unit 10, and generates an answer text to the question text included in the prompt by using the additional text included in the prompt as reference information. That is, the answer generation unit 20 generates an answer sentence with a higher degree of conformity to the reference information.

FIG. 2 is a block diagram showing a functional configuration example of the text generation unit 10. As shown in FIG. As shown in FIG. 2, the text generation unit 10 of the present embodiment includes an additional text generation unit 11 and a prompt generation unit 12 as functional configurations. The additional text generation unit 11 includes a candidate text acquisition unit 111 and a text adjustment unit 112 as more specific functional configurations.

Based on the question text input from the user terminal 200, the additional text generating part 11 generates the additional text related to the input question text with the total number of characters combined with the number of characters of the question text not exceeding the limit of the number of characters that can be input to the answer generating part 20 of the LLM 1. - 特許庁The prompt generation unit 12 generates a prompt by adding the additional text generated by the additional text generation unit 11 as reference information to the question text input from the user terminal 200 .

For example, as shown in FIG. 13, when a question such as "Is Hiroshima tomorrow suitable for hanging laundry?" In the example of Fig. 13, a sentence such as "Please answer the following #questions. When answering, please refer to the contents written in #reference facts and answer based on them." is added at the beginning of the prompt. At this time, the additional text generation unit 11 generates the additional text so that the total number of characters of the prompt composed of the input question text and the additional text does not exceed the input character limit of the LLM1.

Based on the question text input from the user terminal 200, the candidate text acquisition unit 111 of the additional text generation unit 11 acquires candidate texts related to the question text. In order for the answer generator 20 to generate a more appropriate answer to the question, it is important to generate appropriate additional sentences to the question. In order to generate appropriate additional sentences, it is important to acquire appropriate candidate sentences from the sentence DB2. For this purpose, the candidate sentence acquisition unit 111 acquires candidate sentences using one or more of the methods described later.

The sentence adjustment unit 112 generates additional sentences by adjusting the candidate sentences acquired by the candidate sentence acquisition unit 111 . When the candidate text acquired by the candidate text acquisition part 111 has a large number of characters, if the candidate text is added to the question text as it is to generate a prompt, the total number of characters may exceed the limit of the number of input characters of the LLM1. In such a case, the sentence adjustment unit 112 adjusts the candidate sentences using one or more of the methods described later so that the total number of characters is within the character number limit.

A method of acquiring candidate sentences by the candidate sentence acquisition unit 111 will be described in detail below. The candidate sentence acquisition unit 111 can acquire candidate sentences by applying one or more of the techniques described below.

<First Candidate Sentence Acquisition Method>
FIG. 3 is a block diagram showing a functional configuration example of the candidate sentence obtaining unit 111 that implements the first candidate sentence obtaining method. In the example shown in FIG. 3, the candidate sentence acquisition unit 111 includes a keyword extraction unit 111a and a keyword search unit 111b.

The keyword extraction unit 111a extracts a search keyword from the question text input from the user terminal 200. FIG. For example, the keyword extraction unit 111a extracts words as search keywords from the input question sentence by morphological analysis. Here, words may be extracted by limiting to specific parts of speech (e.g., nouns, verbs, adjectives, adjectives, etc.), or words may be extracted by excluding specific parts of speech (e.g., particles, auxiliary verbs, adverbs, adnominals, conjunctions, interjections, etc.). One or a plurality of words may be extracted, but if a plurality of words are extracted and used as a search keyword, it is possible to retrieve and acquire sentences that are more closely related to the question sentence.

The keyword search unit 111b uses the search keyword extracted by the keyword extraction unit 111a to search the sentence DB2 in which a plurality of sentences are recorded, thereby obtaining sentences including the search keyword from the sentence DB2 as candidate sentences. When the keyword extraction unit 111a extracts a plurality of search keywords, the keyword search unit 111b acquires sentences including all of the plurality of search keywords from the sentence DB2 as candidate sentences. If no relevant sentence is found, the search may be re-executed by omitting any of the search keywords.

If there are multiple sentences that match the search keyword, the keyword search unit 111b may acquire all of the multiple sentences. In addition, when all the plural sentences that match the search keyword are acquired, the total number of characters is likely to exceed the input character limit of the LLM1. Therefore, a predetermined number of sentences may be acquired as candidate sentences in descending order of degree of matching with the search keyword.

Here, as shown in FIG. 3, there may be a plurality of types of text DB2 to be searched. FIG. 3 shows three of fact information DB 2a, persona information DB 2b, and correct example information DB 2c. The keyword search unit 111b executes a sentence search for one or more of these three sentence DBs 2a to 2c using the words extracted by the keyword extraction unit 111a as search keywords. In order to improve the accuracy of the answer sentence generated by the answer generator 20, it is preferable to search multiple databases.

The factual information DB 2a is a database that holds, for example, web pages of specific companies, Wikipedia (registered trademark), or files such as txt, csv, word, excel, and pdf uploaded by the user using the user terminal 200 in a referable state. The persona information DB 2b is a database that holds a history of past interactions with the LLM 1 by a specific user, or information extracted therefrom from which the user's literacy can be inferred (e.g., degree of pre-knowledge regarding a specific technical field, past travel history, demographic attributes such as age, gender, family composition, etc.). The correct example information DB 2c is a database that holds information obtained by manually correcting a past interaction history with the LLM 1 by an arbitrary user, such as "LLM 1 should have replied in this way".

FIG. 4 is a diagram showing an example of a prompt constructed using candidate sentences respectively obtained from the three sentence DBs 2a-2c. In FIG. 4, "# referenceable information is a candidate sentence acquired from the fact information DB 2a, "# questioner's persona" is a candidate sentence acquired from the persona information DB 2b, and "# reference answer example" is a candidate sentence acquired from the correct example information DB 2c. In the example shown in FIG. 4, since the total number of characters in all the sentences included in the prompt does not exceed the input character limit of LLM1, there is no need to execute the processing of the sentence adjustment unit 112, and the prompt shown in FIG.

<Second Candidate Sentence Acquisition Method>
FIG. 5 is a block diagram showing a functional configuration example of the candidate sentence obtaining unit 111 that implements the second candidate sentence obtaining method. In the example shown in FIG. 5, the candidate sentence acquisition unit 111 includes a vector calculation unit 111c and a vector search unit 111d.

The vector calculation unit 111c calculates the feature vector of the text from the question text input from the user terminal 200. FIG. For example, the vector calculation unit 111c has a small-scale language model of an embedding model, and uses the embedding model to convert an input question text into a numerical vector. A small language model is a natural language processing model trained on less data than LLM1. Note that the feature vector calculation method is not limited to this, and other known techniques may be applied. For example, words may be extracted from an input question sentence by morphological analysis, and the meanings of the words and the relationships between the words may be vectorized.

In order to apply the second candidate sentence acquisition method, feature vectors are also calculated in advance for the plurality of sentences recorded in the sentence DBs 2a to 2c, and the plurality of sentences are recorded in association with the respective feature vectors in advance. The vector search unit 111d searches the text DBs 2a to 2c using the feature vectors calculated by the vector calculation unit 111c for the input question sentence, thereby obtaining sentences associated with the feature vectors that satisfy a predetermined condition as candidate sentences from the text DBs 2a to 2c.

Predetermined conditions for obtaining text can be arbitrarily determined. For example, it is possible to obtain a predetermined number of sentences in descending order of feature vector similarity. Alternatively, sentences having feature vector similarities greater than a predetermined threshold may be obtained. In this case, it is possible to acquire all the sentences whose feature vector similarity is greater than a predetermined threshold, or if the number of sentences whose similarity is greater than the predetermined threshold exceeds a predetermined number, a predetermined number of sentences having the highest similarity may be acquired.

<Third Candidate Sentence Acquisition Method>
FIG. 6 is a block diagram showing a functional configuration example of the candidate sentence obtaining unit 111 that implements the third candidate sentence obtaining method. In the example shown in FIG. 6, the candidate sentence acquisition unit 111 further includes a question sentence generation unit 111e in addition to the vector calculation unit 111c and the vector search unit 111d shown in FIG. In addition, the sentence DB2 further includes an exemplary question sentence DB2d in which a plurality of exemplary question sentences are recorded.

The question text generation unit 111 e generates another question text from the question text input from the user terminal 200 . In addition to the feature vector of the question text input from the user terminal 200, the vector calculation unit 111c calculates the feature vector of another question text generated by the question text generation unit 111e. The vector search unit 111d uses the feature vector calculated from the question text input from the user terminal 200 and the feature vector calculated from another question text generated by the question text generation unit 111e to search the text DBs 2a to 2c to obtain candidate texts.

Here, the question text generation unit 111e has a small-scale language model generated by applying Few-shot learning, for example, constructs a prompt by adding an example question text to the input question text, and inputs it to the small-scale language model, thereby outputting another question text from the small-scale language model. Exemplary question sentences are acquired from the exemplary question sentence DB 2d. FIG. 7 is a diagram showing an example of a prompt and another question sentence generated by the candidate sentence acquisition unit 111 in this case. In the example of FIG. 7, a prompt is formed by adding an exemplary question sentence extracted from the exemplary question sentence DB 2d using "Okinawa trip" included in the question sentence as a keyword to the input question sentence, and another question sentence is generated as indicated by the example question sentence.

In addition, the question text generation unit 111e may dynamically generate a reference question generation example according to a question text input from the user terminal 200 to compose a prompt. For example, the question text generation unit 111e can extract example question texts that match the question text input from the user terminal 200 from the example question text DB 2d, generate a question generation example that serves as a reference from the extracted example question texts, and compose a prompt. In this example, the exemplary question sentence that matches the input question sentence can be, for example, an exemplary question sentence associated with a feature vector whose similarity to the feature vector calculated from the input question sentence is equal to or greater than a threshold.

FIG. 8 is a diagram showing an example of a prompt and another question sentence generated by the candidate sentence acquisition unit 111 in this case. In the example of FIG. 8, a plurality of exemplary question sentences that match the input question sentence are extracted from the exemplary question sentence DB 2d, and a question generation example that serves as a reference is generated from these to compose a prompt. The small-scale language model uses multiple question generation examples included in this prompt to dynamically generate another question sentence. In this example, the small-scale language model is generated by training using the past conversation history, and instead of generating another question sentence according to the question generation example, another question sentence is generated by inputting the question generation example and the conversation history.

<Fourth Candidate Sentence Acquisition Method>
FIG. 9 is a block diagram showing a functional configuration example of the candidate sentence acquisition unit 111 that implements the fourth candidate sentence acquisition method. In the example shown in FIG. 9A, the candidate sentence acquisition unit 111 includes a question generation example generation unit 111f and a question/answer generation unit 111g. Further, the text DB2 does not include the text DB2a to 2c, but includes an exemplary question text DB2d.

The question generation example generation unit 111f generates a question generation example to be used as a reference from the question text input from the user terminal 200. FIG. The method of generating a reference question generation example by the question generation example generation unit 111f may be the same as the generation method by the question generation unit 111e illustrated in FIGS. That is, the question generation example generation unit 111f extracts from the example question sentence DB 2d an example question sentence that matches the question sentence input from the user terminal 200, and generates a question generation example that serves as a reference from the extracted example question sentence. Alternatively, since the question/answer generation unit 111g is a small-scale language model with flexible reading comprehension ability, the question input from the user terminal 200 may be directly input to the question/answer generation unit 111g without generating question generation examples. In this case, the question generation example generation unit 111f and the example question sentence DB 2d may be deleted from the configuration.

The question/answer generation unit 111g inputs the reference question generation examples generated by the question generation example generation unit 111f into the small-scale language model, thereby obtaining combinations of question sentences and answer sentences as candidate sentences from the small-scale language model.

The small-scale language model provided in the question/answer generation unit 111g can be, for example, a model obtained by fine-tuning the LLM1 according to a predetermined task. Although the small-scale language model is a small-scale model, it is practically difficult to perform fine-tuning for each user and always store multiple small-scale language models generated by this on the memory of the server. As means for solving this problem, a known method of additional learning called LoRA may be used.

LoRA learns by adding a very small language model called Adapter to an original small language model, and fine tuning can be performed efficiently with low memory. While the original small-scale language model is usually several dozen GBytes, Adapter usually stays at several dozen MBytes. Therefore, even if the small language model resides in memory and only the Adapter is loaded on demand according to the user's request, the load time will be negligible. Alternatively, it becomes possible to place a large number of Adapters on a limited memory. This method is not limited to the question/answer generation unit 111g, and can be generally used in scenes where language models are used.

In fine-tuning the small-scale language model of the question/answer generator 111g, either self-instructed learning or supervised learning may be used. When self-supervised learning is used, for example, it is possible to learn by extracting input-output pairs from past dialogue history catchballs. On the other hand, if supervised learning is used, the question-answer pairs can be generated by another LLM that references past interaction history through prompts, and this question-answer pair can be learned. A prompt for obtaining a question-answer pair is illustrated in FIG.

In the example shown in FIG. 9B, the candidate sentence acquisition unit 111 includes a question sentence generation unit 111h and an answer sentence acquisition unit 111j. The question text generation unit 111h generates another question text from the question text input from the user terminal 200 . The method of generating another question text by the question text generating unit 111h may be the same as the method of generating another question text by the question text generating unit 111e illustrated in FIGS. Alternatively, since the answer sentence acquisition unit 111j is a small-scale language model with flexible reading comprehension ability, the question input from the user terminal 200 may be directly input to the answer sentence acquisition unit 111j without generating the question sentence. In this case, the question sentence generator 111h and the example question sentence DB 2d may be deleted from the configuration.

The answer sentence acquisition unit 111j inputs another question sentence generated by the question sentence generation unit 111h to the small-scale language model, thereby acquiring an answer sentence to the other question sentence as a candidate sentence from the small-scale language model. In the example shown in FIG. 9B, the candidate sentence acquisition unit 111 outputs a combination of the question sentence generated by the question sentence generation unit 111h and the answer sentence acquired by the answer sentence acquisition unit 111j as a candidate sentence. In this case, the question/answer generation unit in the scope of claims is configured by the question sentence generation unit 111h and the answer sentence acquisition unit 111j.

When candidate sentence acquisition unit 111 acquires candidate sentences using the above-described method, the number of candidate sentences acquired is often plural, and the acquired candidate sentences may be redundant. Therefore, if the candidate text acquired by the candidate text acquisition unit 111 is added to the question text as it is to generate a prompt, the total number of characters of the prompt may exceed the input character limit of the LLM1. In that case, processing by the text adjustment unit 112 is required.

A method of adjusting candidate sentences by the sentence adjustment unit 112 will be described in detail below. The sentence adjuster 112 can apply any one or more of the techniques described below to adjust the candidate sentences.

<First Candidate Sentence Adjustment Method>
A text adjustment unit 112 inputs a question text input from a user terminal 200 and candidate texts acquired by a candidate text acquisition unit 111 into a small-scale language model, thereby generating an additional text with a reduced number of characters by using a text part related to the question text out of the candidate texts. The small-scale language model provided in the sentence adjustment unit 112 may be, for example, the LLM1 of the present embodiment that has been independently fine-tuned for the above task, or may be a small-scale language model generated for the above task.

FIG. 11 is a diagram showing two adjustment examples of candidate sentences by the first candidate sentence adjustment method. In the example shown in FIG. 11, the candidate sentence acquired by the candidate sentence acquisition unit 111 is "Speaking of travel to Okinawa, it's the beach in summer, and the aquarium in winter!".

FIG. 11A shows an adjustment example of candidate sentences when the input question sentence is "I want to go on a trip to Okinawa in winter. What is the recommended plan for 4 days and 3 nights?" Since the question is about a trip to Okinawa in winter, the small-scale language model underestimates the relevance of the part of the candidate sentences "in summer to the sea" to the question sentence, and uses the part of the candidate sentences related to the question sentence "in winter, aquarium" to generate an additional sentence "I recommend an aquarium for a trip to Okinawa".

On the other hand, FIG. 11B shows an adjustment example of candidate sentences when the input question sentence is "I want to go on a trip to Okinawa in the summer. What is the recommended plan for 4 days and 3 nights?" Since the question is about a trip to Okinawa in summer, the small-scale language model underestimates the relevance of the part of the candidate sentences "Aquarium in winter" to the question sentence, and uses the part of the candidate sentences related to the question sentence "If summer is the sea" to generate an additional sentence "Travel to Okinawa by the sea".

Although an example of adjustment of one candidate sentence is shown here for explanation, simply adding one candidate sentence shown here to the question sentence does not exceed the limit of the number of characters that can be input to the LLM1. Actually, there are a plurality of candidate sentences, and if the total number of characters exceeds the character number limit when all of them are added to the question sentence, the adjustment process described above is executed for each candidate sentence. This also applies to the second candidate sentence adjustment method described below.

<Second Candidate Sentence Adjustment Method>
The sentence adjustment unit 112 may generate additional sentences with a reduced number of characters by summarizing the candidate sentences acquired by the candidate sentence acquisition unit 111 . Candidate sentences may be summarized using a small-scale language model generated for the task, or may be performed using a known algorithm.

In addition, the sentence adjustment unit 112 may generate an additional sentence with a reduced number of characters by summarizing the main points related to the question sentence input from the user terminal 200 for the candidate sentences acquired by the candidate sentence acquisition unit 111. In this case, the sentence adjustment unit 112, for example, receives the question sentence input from the user terminal 200 and the candidate sentence acquired by the candidate sentence acquisition unit 111, and outputs a summarized candidate sentence. It can be configured with a trained small-scale language model.

FIG. 12 is a flow chart showing an operation example of the text generation unit 10 according to this embodiment configured as described above. The flowchart shown in FIG. 12 outlines the procedure of processing executed for one question sentence input from the user terminal 200 to the LLM1.

First, the additional sentence generation unit 11 of the LLM 1 inputs a question sentence from the user terminal 200 (step S1). When the additional sentence generation unit 11 inputs a question sentence, the candidate sentence acquisition unit 111 applies one or more of the first to fourth candidate sentence acquisition methods to acquire one or more candidate sentences related to the input question sentence (step S2).

Next, the sentence adjustment unit 112 determines whether the total number of characters, which is the number of characters of the question sentence input from the user terminal 200 and the number of characters of one or more candidate sentences acquired by the candidate sentence acquisition unit 111, exceeds the limit of the number of characters that can be input to the LLM 1 (step S3). Here, when it is determined that the total number of characters exceeds the limit of the number of input characters, the text adjustment unit 112 applies one or both of the first and second candidate text adjustment methods to reduce the number of characters of the candidate text so that the total number of characters is within the range of the limit of the number of input characters (step S4).

Here, for example, the text adjusting unit 112 adjusts all candidate texts acquired by the candidate text acquiring unit 111 to reduce the number of characters, and determines again whether the total number of characters after reduction exceeds the limit of the number of input characters. If the total number of characters still exceeds the input character limit, adjust the candidate sentences again to reduce the number of characters. Alternatively, any of the candidate sentences acquired by the candidate sentence acquisition unit 111 may be deleted. By repeating such processing, the total number of characters is kept within the limit of the number of input characters. One or more candidate sentences adjusted in this way are determined as additional sentences for the question sentence.

When the additional text for the question text is generated by adjusting the candidate text by the text adjusting unit 112 as described above, the prompt generating unit 12 generates a prompt by adding the additional text generated by the text adjusting unit 112 in step S4 as reference information to the question text input from the user terminal 200 in step S1 (step S5).

If it is determined in step S3 that the total number of characters does not exceed the limit of the number of input characters, the processing of the sentence adjusting unit 112 in step S4 is not performed, and the candidate sentence acquired by the candidate sentence acquiring unit 111 is determined as it is as an additional sentence to the question sentence. In this case, the prompt generating unit 12 generates a prompt by adding the candidate text (=additional text) obtained by the candidate text obtaining unit 111 in step S2 as reference information to the question text input from the user terminal 200 in step S1 (step S5).

Then, the prompt generator 12 inputs the generated prompt to the answer generator 20 (step S6). The answer generation unit 20 uses the additional text included in the prompt input from the text generation unit 10 as reference information to generate a response text to the question text included in the prompt, and outputs the generated response text to the user terminal 200 (step S7). As a result, the processing of the flowchart shown in FIG. 12 ends.

As described above in detail, in this embodiment, based on a question text input from the user terminal 200 to the LLM 1, the total number of characters combined with the number of characters of the question text does not exceed the character limit that can be input to the answer generation unit 20 of the LLM 1. Additional text related to the question text is generated, and the generated additional text is added to the question text as reference information, thereby generating a prompt for input to the answer generation unit 20.

According to this embodiment configured in this way, when generating a prompt by adding sentences related to an input question sentence, the additional sentence is generated so that the number of characters does not exceed the limit of the number of characters that can be input to the answer generating part 20 as the prompt.

In addition, although the said embodiment demonstrated the example provided with the text production|generation part 10 as a structure of LLM1, this invention is not limited to this. For example, the text generator 10 may be configured separately from the LLM1.

Further, in the above embodiment, the user may be allowed to select the output format of the answer sentence output by the answer generator 20 . For example, the tone of the answer sentence may be selected. This can be realized by specifying additional sentences such as "Reply with a tone of XX" when the sentence generating unit 10 generates a prompt. Due to the nature of LLMs, when reference information is provided in a prompt, there is a tendency for the tone specification to be ignored and the reference tone to take precedence. To solve this problem, the LLM may be instructed to output in two stages. In other words, the prompt may include additional text that says, "First, write 'Normal answer:' followed by your normal answer, then 'Character of XX:' and then answer in the tone of the character of XX." In this case, by outputting the part after "character of XX:" in the output of the LLM, it is possible to present to the user only the part imitating the tone of the character. Also, the format of the reply sentence may be selected. For example, according to the format of administrative declaration documents, etc., it may be possible to output according to chapters such as "summary", "business contents", "application items", and "additional items". This can also be accomplished by specifying additional text in the prompt.

Also, the text DBs 2a-2d or the small-scale language models used in the above embodiments can be sold to companies. It can be a lump sum payment or a subscription type. When selling text DBs 2a to 2d or small-scale language models, it is necessary to remove personal information and company confidential information. The work may be done manually or using an LLM created for the task. It should be noted that the output format of the above-described reply text may also be made available for sale to companies.

In addition, the above-described embodiments are merely examples of specific implementations of the present invention, and the technical scope of the present invention should not be construed in a limited manner. Thus, the invention may be embodied in various forms without departing from its spirit or essential characteristics.

1 Large scale language model (LLM)
2 Text database
2a Fact information DB
2b Persona information DB
2c Correct example information DB
10 text generation unit (text generation device)
11 additional text generation unit 12 prompt generation unit 20 answer generation unit 111 candidate text acquisition unit 111a keyword extraction unit 111b keyword search unit 111c vector calculation unit 111d vector search unit 111e question text generation unit 111f question generation example generation unit 111g question/answer generation unit 111h question text generation unit (question/answer generation unit)
111j Answer sentence acquisition unit (question/answer generation unit)
112 Sentence Coordinator

Claims (9)

A text generation device that generates a prompt for inputting into a large-scale language model by adding reference information to an input question text,
an additional sentence generation unit for generating an additional sentence related to the question sentence based on the input question sentence, with the total number of characters combined with the number of characters of the question sentence not exceeding the limit of the number of characters that can be input to the large-scale language model;
a prompt generation unit that generates the prompt by adding the additional text generated by the additional text generation unit as the reference information to the input question text ;
The above additional sentence generation unit
a candidate sentence acquisition unit for acquiring candidate sentences related to the question sentence based on the input question sentence;
a sentence adjustment unit for generating the additional sentence by adjusting the candidate sentence acquired by the candidate sentence acquisition unit.
A text generation device characterized by: 2. The sentence generating apparatus according to claim 1 , wherein the sentence adjustment unit inputs the input question sentence and the candidate sentence acquired by the candidate sentence acquiring unit into a small-scale language model, thereby generating the additional sentence using a sentence part related to the question sentence among the candidate sentences. 2. The text generation apparatus according to claim 1 , wherein the text adjustment unit generates the additional text by summarizing the candidate texts acquired by the candidate text acquisition unit. 4. The text generation device according to claim 3 , wherein the text adjustment unit generates the additional text by summarizing the key points related to the input question text for the candidate texts acquired by the candidate text acquisition unit. The above candidate sentence acquisition unit
a keyword extraction unit for extracting a search keyword from the input question text;
5. The sentence generation device according to claim 1, further comprising a keyword search unit that searches a sentence database in which a plurality of sentences are recorded using the search keyword extracted by the keyword extraction unit, and obtains sentences containing the search keyword from the sentence database as the candidate sentences. The above candidate sentence acquisition unit
a vector calculation unit that calculates a feature vector of the text from the input question text;
5. The sentence generating apparatus according to claim 1, further comprising: a vector searching unit that searches a sentence database in which a plurality of sentences are recorded in advance in association with the feature vectors calculated by the vector calculating unit, using the feature vectors calculated by the vector calculating unit, and acquires sentences associated with the feature vectors that satisfy a predetermined condition regarding similarity to the feature vectors calculated by the vector calculating unit from the sentence database as the candidate sentences. The above candidate sentence acquisition unit
further comprising a question text generation unit that generates another question text from the input question text;
The vector calculation unit calculates, in addition to the feature vector of the input question sentence, the feature vector of the another question sentence generated by the question sentence generation unit,
7. The sentence generation device according to claim 6 , wherein the vector search unit searches a sentence database using a feature vector calculated from the input question sentence and a feature vector calculated from the other question sentence. The sentence generation device according to any one of claims 1 to 4, wherein the candidate sentence acquisition unit includes a question/answer generation unit that generates, as the candidate sentence, a combination of another question sentence and an answer sentence using a small-scale language model based on the input question sentence. A text generation method for generating a prompt for inputting into a large-scale language model by adding reference information to an input question text,
A first step of generating an additional sentence related to the question sentence by an additional sentence generation unit of a computer, based on the input question sentence, with a total number of characters including the number of characters of the question sentence not exceeding the limit of the number of characters that can be input to the large-scale language model;
a second step of generating the prompt by adding the additional text generated by the additional text generation unit as the reference information to the input question text, wherein the prompt generation unit of the computer generates the prompt ;
In the first step, the additional sentence generation unit acquires candidate sentences related to the question sentence based on the input question sentence, and generates the additional sentence by adjusting the acquired candidate sentences.
A sentence generation method characterized by: