A question-answer pair generation method, device, storage medium and equipment

Abstract

The present application discloses a method, device, storage medium and equipment for generating question-answer pairs, the method comprising: firstly generating a first target question text and a first target answer text in a preset format according to text corpus data; then constructing a first prompt using the first target question text, inputting it into a preset large language model, rewriting the first target question text, and obtaining a second target question text and a second target answer text output by the model; and constructing a second prompt using the text corpus data, inputting it into a large language model, obtaining a third target question text and a third target answer text output by the model, then segmenting the text corpus data, and constructing a third prompt using the segmentation result, inputting it into a large language model, and obtaining a fourth target question text and a fourth target answer text output by the model; and then generating question-answer pairs using these target question texts and target answer texts. Thus, the generation efficiency, diversity, and professionalism of question-answer pairs can be improved.

Description

Technical Field

[0001] The present application relates to the technical field of natural language processing, and in particular to a method, apparatus, storage medium and device for generating question and answer pairs. Background Art

[0002] With the rapid development of artificial intelligence and natural language processing technology, question-answering systems have been widely used in various fields such as medicine, education, and agriculture. In question-answering systems, users' questions are usually answered through pre-set question-answer pairs, so how to generate question-answer pairs is very important.

[0003] At present, the mainstream question-answer generation system generally learns to generate question-answer pairs based on experience. This question-answer generation method usually includes three implementation methods: one is to generate question-answer pairs based on templates. Although this method is simple, it lacks flexibility and diversity, and has high requirements for the diversity of templates. The second is to generate question-answer pairs based on retrieval. This method relies on the relevance of retrieval quality and text paragraphs, which is more labor-intensive, and the diversity, richness, and professionalism of the constructed question-answer pairs are also insufficient. The third method is to use a pre-trained model to construct question-answer pairs, but the question-answer pairs generated by this method are more dependent on training data, lack generalization ability, and require operators to have a high level of code development. It can be seen that the existing question-answer generation methods mentioned above cannot achieve the desired effect. Summary of the invention

[0004] The main purpose of the embodiments of the present application is to provide a question-answer pair generation method, device, storage medium and equipment, which can improve the efficiency, diversity, professionalism and operability of question-answer pair generation, achieve ideal question-answer pair generation effect, and thus improve the user's question-answer experience.

[0005] The present application embodiment provides a method for generating a question-answer pair, including:

[0006] Acquire text corpus data; and generate a first target question text and a first target answer text in a preset format according to the text corpus data;

[0007] Using the first target question text to construct a first prompt instruction prompt, inputting it into a preset large language model, rewriting the first target question text, and obtaining a second target question text and a second target answer text output by the large language model;

[0008] Using the text corpus data to construct a second prompt instruction prompt, inputting it into a preset large language model, and obtaining a third target question text and a third target answer text output by the large language model;

[0009] Segment the text corpus data, and construct a third prompt instruction prompt using the segmentation result, and input it into a preset large language model to obtain a fourth target question text and a fourth target answer text output by the large language model;

[0010] Question-answer pairs are generated using the first target question text and the first target answer text, the second target question text and the second target answer text, the third target question text and the third target answer text, and the fourth target question text and the fourth target answer text.

[0011] In a possible implementation, before generating a question-answer pair using the first target question text and the first target answer text, the second target question text and the second target answer text, the third target question text and the third target answer text, and the fourth target question text and the fourth target answer text, the method further includes:

[0012] Using the third target question text to construct a fourth prompt instruction prompt, inputting it into a preset large language model, rewriting the third target question text, and obtaining a fifth target question text and a fifth target answer text output by the large language model;

[0013] And / or, the fourth target question text is used to construct a fifth prompt instruction prompt, which is input into a preset large language model, and the fourth target question text is rewritten to obtain a sixth target question text and a sixth target answer text output by the large language model.

[0014] In a possible implementation, generating a question-answer pair using the first target question text and the first target answer text, the second target question text and the second target answer text, the third target question text and the third target answer text, and the fourth target question text and the fourth target answer text includes:

[0015] Question-answer pairs are generated using the first target question text and the first target answer text, the second target question text and the second target answer text, the third target question text and the third target answer text, the fourth target question text and the fourth target answer text, and the fifth target question text and the fifth target answer text, and / or the sixth target question text and the sixth target answer text.

[0016] In a possible implementation, the text corpus data is segmented, and the segmentation result is used to construct a third prompt instruction prompt, which is input into a preset large language model to obtain a fourth target question text and a fourth target answer text output by the large language model, including:

[0017] According to a preset segmentation method, the text corpus data is segmented to obtain N slice texts, where N is a positive integer greater than 0;

[0018] Based on the correlation between the title and the content, the N slice texts are filtered to obtain M slice texts; M is a positive integer less than N;

[0019] Based on the relevance of the text content, the M slice texts are fused to obtain K slice texts; K is a positive integer less than M;

[0020] A third prompt instruction prompt is constructed using the K slice texts, and is input into a preset large language model to obtain a fourth target question text and a fourth target answer text output by the large language model.

[0021] In a possible implementation, the preset segmentation method includes at least one of segmentation based on paragraphs, segmentation based on sentences, and segmentation based on preset byte lengths.

[0022] In a possible implementation, after segmenting the text corpus data according to a preset segmentation method to obtain N sliced ​​texts, the method further includes:

[0023] The N sliced ​​texts are used to construct a sixth prompt instruction prompt, which is input into a preset large language model, and the N sliced ​​texts are highlighted and extracted to obtain P paragraph summary texts within a preset number of words output by the large language model; P is a positive integer less than N;

[0024] Based on the correlation between the title and the content, the N slice texts are filtered to obtain M slice texts, including:

[0025] Based on the relevance between the title and the content, the P paragraph summary texts are filtered to obtain M slice texts.

[0026] In a possible implementation, generating a question-answer pair using the first target question text and the first target answer text, the second target question text and the second target answer text, the third target question text and the third target answer text, and the fourth target question text and the fourth target answer text includes:

[0027] Performing clustering processing on the first target question text, the second target question text, the third target question text and the fourth target question text to obtain T grouping clustering results; wherein T is a positive integer greater than 0;

[0028] T target question texts are extracted from the T grouping clustering results by classification, and T question-answer pairs are generated using the T target question texts and their corresponding target answer texts.

[0029] The present application also provides a question-answer pair generating device, including:

[0030] An acquisition unit, configured to acquire text corpus data; and generate a first target question text and a first target answer text in a preset format according to the text corpus data;

[0031] A first input unit is used to construct a first prompt instruction prompt using the first target question text, input it into a preset large language model, rewrite the first target question text, and obtain a second target question text and a second target answer text output by the large language model;

[0032] A second input unit is used to construct a second prompt instruction prompt using the text corpus data, and input it into a preset large language model to obtain a third target question text and a third target answer text output by the large language model;

[0033] A third input unit is used to segment the text corpus data, and construct a third prompt instruction prompt using the segmentation result, and input it into a preset large language model to obtain a fourth target question text and a fourth target answer text output by the large language model;

[0034] A generating unit is used to generate question-answer pairs using the first target question text and the first target answer text, the second target question text and the second target answer text, the third target question text and the third target answer text, and the fourth target question text and the fourth target answer text.

[0035] In a possible implementation manner, the device further includes:

[0036] a fourth input unit, configured to construct a fourth prompt instruction prompt using the third target question text, input the fourth prompt instruction prompt into a preset large language model, rewrite the third target question text, and obtain a fifth target question text and a fifth target answer text output by the large language model;

[0037] And / or, the fourth target question text is used to construct a fifth prompt instruction prompt, which is input into a preset large language model, and the fourth target question text is rewritten to obtain a sixth target question text and a sixth target answer text output by the large language model.

[0038] In a possible implementation, the generating unit is specifically used for:

[0039] Question-answer pairs are generated using the first target question text and the first target answer text, the second target question text and the second target answer text, the third target question text and the third target answer text, the fourth target question text and the fourth target answer text, and the fifth target question text and the fifth target answer text, and / or the sixth target question text and the sixth target answer text.

[0040] In a possible implementation, the third input unit includes:

[0041] A segmentation subunit, used to segment the text corpus data according to a preset segmentation method to obtain N slice texts; N is a positive integer greater than 0;

[0042] A filtering subunit, configured to filter the N slice texts based on the relevance between the title and the content to obtain M slice texts; wherein M is a positive integer less than N;

[0043] A fusion subunit is used to fuse the M slice texts based on the relevance of the text contents to obtain K slice texts; K is a positive integer less than M;

[0044] The first input subunit is used to construct a third prompt instruction prompt using the K slice texts, and input it into a preset large language model to obtain a fourth target question text and a fourth target answer text output by the large language model.

[0045] In a possible implementation, the preset segmentation method includes at least one of segmentation based on paragraphs, segmentation based on sentences, and segmentation based on preset byte lengths.

[0046] In a possible implementation, the third input unit further includes:

[0047] The second input subunit is used to construct a sixth prompt instruction prompt using the N slice texts, input the sixth prompt instruction prompt into a preset large language model, perform highlight extraction on the N slice texts, and obtain P paragraph summary texts within a preset number of words output by the large language model; P is a positive integer less than N;

[0048] The filtering subunit is specifically used for:

[0049] Based on the relevance between the title and the content, the P paragraph summary texts are filtered to obtain M slice texts.

[0050] In a possible implementation, the generating unit includes:

[0051] A clustering subunit, configured to perform clustering processing on the first target question text, the second target question text, the third target question text and the fourth target question text to obtain T grouping clustering results; wherein T is a positive integer greater than 0;

[0052] A generating subunit is used to classify and extract T target question texts from the T grouping clustering results, and generate T question-answer pairs using the T target question texts and their corresponding target answer texts.

[0053] The embodiment of the present application also provides a question-answer pair generating device, including: a processor, a memory, and a system bus;

[0054] The processor and the memory are connected via the system bus;

[0055] The memory is used to store one or more programs, and the one or more programs include instructions, which, when executed by the processor, enable the processor to execute any one of the implementation methods of the above-mentioned question-answer pair generation method.

[0056] An embodiment of the present application also provides a computer-readable storage medium, in which instructions are stored. When the instructions are executed on a terminal device, the terminal device executes any one of the implementations of the above-mentioned question-answer pair generation method.

[0057] An embodiment of the present application also provides a computer program product. When the computer program product is run on a terminal device, the terminal device executes any one of the implementations of the above-mentioned question-answer pair generation method.

[0058] The embodiments of the present application provide a method, apparatus, storage medium and device for generating question-answer pairs. The method first acquires text corpus data, generates a first target question text and a first target answer text in a preset format according to the text corpus data, constructs a first prompt instruction prompt using the first target question text, inputs the prompt into a preset large language model, rewrites the first target question text, and obtains a second target question text and a second target answer text output by the large language model, and constructs a second prompt instruction prompt using the text corpus data, inputs the prompt into a preset large language model, and obtains a third target question text and a third target answer text output by the large language model. Then, the text corpus data is segmented, and a third prompt instruction prompt is constructed using the segmentation result, inputs the prompt into the preset large language model, and obtains a fourth target question text and a fourth target answer text output by the large language model. Then, the first target question text and the first target answer text, the second target question text and the second target answer text, the third target question text and the third target answer text, and the fourth target question text and the fourth target answer text can be used to generate question-answer pairs.

[0059] It can be seen that since this application constructs different prompts about text corpus data from multiple perspectives and inputs them into a preset large language model, the model can output target question texts and target answer texts from different perspectives to form question-answer pairs. This can improve the generation efficiency, diversity, professionalism, and specificity of question-answer pairs, and achieve an ideal question-answer pair generation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0061] Figure 1 A flowchart of a method for generating question-answer pairs provided in an embodiment of the present application;

[0062] Figure 2 A schematic diagram of the composition of a question-answer pair generating device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0063] With the development of network technology and computer technology, intelligent question-answering systems have been widely used. However, existing question-answering systems still face great challenges in generating high-quality, context-relevant question-answer pairs. Traditional question-answer pair generation methods rely on manually written rules or small-scale data sets, which often do not work well when dealing with complex problems and are difficult to meet the needs of large-scale applications. In recent years, deep learning technology has achieved remarkable results in the field of natural language processing, especially the emergence of pre-trained models, which has provided new ideas for the research and development of question-answering systems.

[0064] Specifically, the current mainstream question-answer generation systems generally learn to generate question-answer pairs based on experience. This question-answer generation method usually includes three implementation methods:

[0065] (1) Template-based question-answer pair generation, that is, using predefined templates to generate question-answer pairs. Although this method is simple, it lacks flexibility and diversity, and has high requirements on the diversity of templates.

[0066] (2) Generating question-answer pairs based on retrieval, that is, retrieving relevant paragraphs from a large-scale document library and then generating question-answer pairs. This method relies on the retrieval quality and the relevance of the text paragraphs, which is relatively labor-intensive. At the same time, the diversity, richness, and professionalism of the constructed question-answer pairs are also insufficient, and the professional level of the constructors is also required to be high.

[0067] (3) Directly use the pre-trained language model to generate question-answer pairs. Although this method is flexible, the generated question-answer pairs are more dependent on the training data, and the generalization ability and diversity of the generated question-answer pairs are insufficient. At the same time, it requires a high level of code development for operators and is not friendly to non-developers.

[0068] It is understandable that the pre-trained language model can be, but is not limited to, a generative pre-trained transformer (Generative Pre-trained Transformer, GPT), a bidirectional transformer model (Bidirectional Encoder Representations from Transformers, BERT), or a text-to-text conversion transformer (Text-To-Text Transfer Transformer, T5), etc. Among them, the GPT model can generate text in an autoregressive manner, which can be applied to generating question-answer pairs, and the model learns the statistical characteristics of the language in the pre-training stage, and is further optimized by the question-answering dataset in the fine-tuning stage. The BERT model is pre-trained by a masked language model and a next sentence prediction task, which can be applied to the generation and understanding of question-answer pairs, and can generally be used to extract contextual information and then generate answers. The T5 model converts all tasks into text generation tasks, including question-answer pair generation tasks. And T5 uses a unified text-to-text framework in both the pre-training and fine-tuning stages. However, even after these pre-trained language models generate question-answer pairs, in order to improve the application effect of the model in the current scenario, it is often necessary to construct training data (i.e., question-answer pair data) for fine-tuning. Therefore, it is still necessary to consider how to generate high-quality question-answer pairs.

[0069] Therefore, the existing question-answer pair generation methods cannot achieve the ideal effect, thereby reducing the user's question-answering experience.

[0070] To solve the above-mentioned defects, the present application provides a method for generating question-answer pairs, which firstly obtains text corpus data; and generates a first target question text and a first target answer text in a preset format according to the text corpus data; then uses the first target question text to construct a first prompt instruction prompt, which is input into a preset large language model, and the first target question text is rewritten to obtain a second target question text and a second target answer text output by the large language model; and, uses the text corpus data to construct a second prompt instruction prompt, which is input into a preset large language model, and a third target question text and a third target answer text output by the large language model are obtained; then, the text corpus data is segmented, and a third prompt instruction prompt is constructed using the segmentation result, which is input into the preset large language model, and a fourth target question text and a fourth target answer text output by the large language model are obtained; and then the first target question text and the first target answer text, the second target question text and the second target answer text, the third target question text and the third target answer text, and the fourth target question text and the fourth target answer text can be used to generate question-answer pairs.

[0071] It can be seen that since this application constructs different prompts about text corpus data from multiple perspectives and inputs them into a preset large language model, the model can output target question texts and target answer texts from different perspectives to form question-answer pairs. This can improve the generation efficiency, diversity, professionalism, and specificity of question-answer pairs, and achieve an ideal question-answer pair generation effect.

[0072] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0073] First embodiment

[0074] See also Figure 1 , is a flow chart of a question-answer pair generation method provided in this embodiment, the method comprising the following steps:

[0075] S101: Acquire text corpus data; and generate a first target question text and a first target answer text in a preset format according to the text corpus data.

[0076] In this embodiment, in order to improve the efficiency, diversity, professionalism, etc. of question-answer pair generation and achieve an ideal question-answer pair generation effect, a large amount of text corpus data (including but not limited to document title, author, document content, document background information, etc.) can be first obtained as the data basis for subsequent generation of question-answer pairs.

[0077] Furthermore, the present embodiment does not limit the language type of the acquired text corpus data, for example, the text corpus data may be Chinese text corpus data or English text corpus data, etc. The present embodiment does not limit the field to which the text corpus data belongs, for example, the text corpus data may belong to various fields such as medical, educational or agricultural. The present embodiment does not limit the length of each text in the text corpus data, that is, it may be a sentence text (i.e., a collection of individual words) or a chapter text (i.e., a collection of individual sentences), etc. And the present embodiment does not limit the method of acquiring the text corpus data, for example, the text corpus data may be acquired from an e-book, an academic paper, a news report, a magazine article, etc., or it may also be acquired from the public web page texts on the Internet, such as some publicly recorded and reported web page contents, etc., as text corpus data.

[0078] Furthermore, after obtaining the text corpus data, we can first generate a preset format of question text (which can be understood as a query instruction / request / demand, and represented by query) and answer text (which can be represented by answer) based on the document title, author, document content, document background information, etc. contained in the text corpus data, and define them as the first target question text and the first target answer text, respectively, to execute the subsequent step S102.

[0079] Among them, the present application does not limit the specific format content of the preset format, and it can be set according to actual conditions and experience values. For example, based on the document content contained in the text corpus data, the first target question text generated can be: "What is the original text of "XXX"? ", "What is the main content of "XXX"? ", "What does "XXX" talk about?", etc. Or, based on the document title contained in the text corpus data, the first target question text generated can be: "What articles has XXX written?", "Which article does this sentence XXX come from?", etc. Or, based on the document author contained in the text corpus data, the first target question text generated can be: "Who wrote "XXX"? ", "Who wrote this sentence XXX?", etc. And it can be understood that the answers corresponding to these first target question texts (i.e., the content of the first target answer text) can all be obtained from the text corpus data.

[0080] In this way, based on the acquired text corpus data, some more common question texts and answer texts (i.e., first target question texts and first target answer texts) can be generated, but these first target question texts and first target answer texts can only cover a part of the dialogue scenarios and often lack diversity, i.e., the generated first target question texts are relatively general and may not be specific or professional. Therefore, it is necessary to continue to execute subsequent steps S102-S105 to further improve the generation efficiency, diversity, professionalism and specificity of question and answer pairs.

[0081] S102: Use the first target question text to construct a first prompt instruction prompt, input it into a preset large language model, rewrite the first target question text, and obtain a second target question text and a second target answer text output by the large language model.

[0082] In this embodiment, after generating a first target question text in a preset format according to the acquired text corpus data through step S102, the first target question text can be further integrated into a first prompt instruction (prompt) and input into a preset large language model (LLM) so as to rewrite the first target question text using the large language model LLM, and output a second target question text and a second target answer text to execute the subsequent step S105.

[0083] Among them, the large language model LLM can be a language model based on deep learning, which can generate new language expressions such as text, sentences, paragraphs and even articles based on the input text content. The large language model LLM uses a large-scale language data set to train language rules and patterns through autoregressive generation, and can simulate human instructions to generate language expressions (such as text data). Specifically, when the large language model LLM generates new text data, it predicts the possibility of the next language unit based on the previously generated content until the complete text data is generated.

[0084] For example: The example of the first prompt command prompt is as follows:

[0085] Please select the questions and answers from the above paragraphs and strictly abide by the following requirements:

[0086] 1. The extracted questions must contain a subject, predicate, and object, and must not contain unclear words such as this study or this experiment.

[0087] 2. Please make sure that the answers to the extracted questions can be found in the original text.

[0088] 3. Make sure each question and answer contains sufficient complexity and depth to avoid generating simple restatements of facts.

[0089] 4. Answers should provide additional information or explanation, not cite the question.

[0090] The specific requirement is that the correct answers to these constructed questions {query} are the full text of the ancient poem {title} and its appreciation.

[0091] Output format

[0092] After marking the query command (query), please output the results in the following JSON format:

[0093]

[0094] Please follow the above requirements to process the query you received: {query, title}. ”

[0095] When the example of the first prompt instruction prompt is input into the preset large language model LLM, an example of the second target question text and the second target answer text output by the large language model can be obtained as follows:

[0096]

[0097]

[0098] Among them, the key points of constructing the first prompt instruction prompt include: 1) the generated query must be clear and easy to understand, and have sufficient complexity and depth; 2) no words with unclear references can appear in the query; 3) ensure that all current document information (i.e., the text corpus data obtained through step S101) can accurately answer the question of the query, thereby improving the diversity of the query (i.e., the question text), that is, the second target question text output by the large language model LLM has a stronger diversity than the first target question text.

[0099] S103: Use the text corpus data to construct a second prompt instruction prompt, input it into a preset large language model, and obtain a third target question text and a third target answer text output by the large language model.

[0100] In this embodiment, after obtaining the text corpus data through step S101, in order to further improve the specificity and professionalism of the question-answer pair generated according to the text corpus data, the document title, author, document background information, etc. contained in the text corpus data can be further incorporated into the second prompt instruction prompt as answer alternative content, and input into the preset large language model LLM, giving the large language model LLM sufficient freedom to utilize the knowledge depth and breadth of the large language model LLM itself to mine out specific question texts (query) and answer texts (answer), such as the question text (query) can be "Which poem of Li Bai expresses his grief and indignation at not being recognized for his talent?", and the corresponding answer text (answer) can be "Jiang Jinjiu", which is used as the third target question text and the third target answer text output by the model to execute the subsequent step S105. It can be seen that the third target question text (query) combines the information such as the document author and document background contained in the text corpus data, and can also give a clear answer. This reflects the specificity and professionalism of the third target question text (query) and the third target answer text (answer), while enhancing the diversity of the third target question text (query) and the third target answer text (answer).

[0101] For example: The example of the second prompt command prompt is as follows:

[0102] "##Role

[0103] You are an experienced lawyer and have a clear understanding of various legal terms.

[0104] ##Task Description

[0105] Now I will give you some legal regulations {title}. Based on these regulations, I will build a series of questions and make sure to strictly abide by the following requirements:

[0106] 1. The extracted questions must contain a subject, predicate, and object, and must not contain unclear words such as this study or this experiment.

[0107] 2. Please make sure that the answers to the extracted questions can be found in the original text.

[0108] 3. Make sure each question and answer contains sufficient complexity and depth to avoid generating simple restatements of facts.

[0109] 4. Answers should provide additional information or explanation, not cite the question.

[0110] The specific requirement is that the correct answers to the constructed questions {query} are the brief introduction and full text of the laws and regulations of {title}.

[0111] Output format

[0112] After marking the query command (query), please output the results in the following JSON format:

[0113]

[0114] Please follow the above requirements to process the query you received: {title}. ”

[0115] When the example of the second prompt instruction prompt is input into the preset large language model LLM, an example of the third target question text and the third target answer text output by the large language model is obtained as follows:

[0116]

[0117] Among them, the key points of constructing the second prompt instruction prompt include: 1) the generated query must be clear and easy to understand, and have sufficient complexity and depth; 2) no unclear words should appear in the query; 3) ensure that all current document information (i.e., the document title, author, document background information, etc. contained in the text corpus data) can accurately answer the query question, thereby improving the diversity, specificity and professionalism of the query (i.e., the question text), that is, the third target question text output by the large language model LLM is more diverse, specific and professional than the first target question text.

[0118] S104: Segment the text corpus data, and use the segmentation results to construct a third prompt instruction prompt, which is input into a preset large language model to obtain a fourth target question text and a fourth target answer text output by the large language model.

[0119] In this embodiment, after obtaining the text corpus data through step S101, in order to further improve the specificity and professionalism of the question-answer pairs generated according to the text corpus data and to enrich the number and content of the generated question-answer pairs, the text corpus data can be segmented using existing or future text segmentation methods, and the obtained segmentation results (i.e., the obtained sliced ​​text) are integrated into the third prompt instruction prompt and input into the preset large language model LLM, so as to use the large language model LLM to mine more professional, detailed and fine-grained question texts (query) and answer texts (answer) as the fourth target question text and fourth target answer text output by the model, so as to execute the subsequent step S105.

[0120] Specifically, an optional implementation manner is that the specific implementation process of step S104 may include the following steps AD:

[0121] Step A: Segment the text corpus data according to a preset segmentation method to obtain N slice texts; where N is a positive integer greater than 0.

[0122] In this implementation, there may be multiple preset segmentation methods, including but not limited to at least one of segmentation based on paragraphs, segmentation based on sentences, and segmentation based on preset byte length (such as 128 bytes, etc., the specific value is not limited). The specific content of the preset segmentation method is not limited and can be set according to actual conditions and experience values. In practical applications, based on the type and length of the document corpus data content, a suitable method can be selected for slicing. For example, when the document corpus data content is standard content, since each sentence in the standard content usually has an accurate practical meaning, the standard content can be segmented by using a sentence-based text segmentation method; or, when the document corpus data content is prose content, since some sentences without practical meaning such as exclamatory sentences may appear in the prose content, in order to ensure the integrity of the sliced ​​content, the prose can be segmented by using a paragraph-based text segmentation method.

[0123] It should be noted that when the text corpus data is segmented based on the preset byte length, there may be a situation where the whole sentence is not completely segmented. For example, when a whole sentence is 164 bytes, if it is segmented based on 128 bytes, the whole sentence will be cut off. For this kind of whole sentence cutting, the cut sentence can be discarded or supplemented based on the specific situation of the cut sentence to ensure the integrity of the sentence after slicing.

[0124] Step B: Based on the correlation between the title and the content, filter the N slice texts to obtain M slice texts; where M is a positive integer less than N.

[0125] In this implementation, after the text corpus data is segmented and processed through step A to obtain N slice texts, considering the length and content integrity of the text itself, these N slices may have the problem of being taken out of context. To avoid this situation, these N slices can be optimized. Specifically, an optional implementation method is to first use these N slice texts to integrate into the sixth prompt instruction prompt, and input them into the preset large language model LLM, so as to use the large language model LLM to highlight the N slice texts and obtain P paragraph summary texts within a preset number of words (the specific value is not limited and can be set according to actual conditions, such as 200 words, etc.) output by the large language model, where P is a positive integer less than N to reduce the number of slice texts.

[0126] For example: The example of the sixth prompt command prompt is as follows:

[0127] "If you are an experienced scholar, please rewrite the following.

[0128] The output format is as follows. Please make sure your output can be parsed by JSON and does not contain any redundant information:

[0129] ```

[0130] {

[0131] "tit1e":"Title",

[0132] "content":"Main text",

[0133] "summary": "Summary",

[0134] "origin_content":"Original text"

[0135] }

[0136] ```

[0137] Output requirements:

[0138] 1. The title should not exceed 20 words, should be attractive and contain enough key information;

[0139] 2. The length of the text is about 200 words, and the text should cover the content of the article as much as possible;

[0140] 3. The abstract should be within 100 words and should be summarized in one or two sentences;

[0141] 4. The effective information of the title, abstract and text must come from the webpage I provide. You only need to organize the key sentences and do not generate irrelevant content. Please rewrite according to the above requirements. "

[0142] When the example of the sixth prompt instruction prompt is input into the preset large language model LLM, an example of the text title, summary, text and other contents including P paragraph summaries output by the large language model is obtained as follows:

[0143] "{

[0144] "title":"Measures to support the dissolution and bankruptcy liquidation of XXX cooperative",

[0145] "content":"In order to support and guide the healthy development of XXX Cooperative, standardize the dissolution and bankruptcy liquidation... the situations where dissolution is required... the persons who violate the regulations will be held accountable accordingly.",

[0146] "summary":"This regulation stipulates the disposal behavior of supporting the dissolution and bankruptcy liquidation of XXX Cooperative. According to the specific circumstances, those who violate the provisions of this regulation will be held accountable according to law.",

[0147] "origin_content":"Article 1 To support and guide the healthy development of cooperatives...\nArticle 2 These measures apply to...\nArticle 3 Responsible for organizing...\nArticle 4...\nArticle 5...\nArticle 9 These measures shall come into force on July 1, 2019."

[0148] }”

[0149] It should be noted that the ellipsis in the above examples is only used to simplify the relevant document content and does not make any other limitation. And the above processing step of using the sixth prompt instruction prompt and the large language model LLM to extract highlights from N slices is only one of the optimization methods and can be selectively executed.

[0150] Then, based on the correlation between the title and the content (the specific calculation method is not limited, for example, similarity can be reflected by calculating text similarity), these N sliced ​​texts (or the P paragraph summary texts obtained after processing by the large language model LLM) can be filtered to obtain M sliced ​​texts, so that the sliced ​​text cleaning operation can be implemented on the basis of retaining the core content of the text body to further reduce the number of slices.

[0151] Step C: Based on the relevance of the text content, the M slice texts are fused to obtain K slice texts; wherein K is a positive integer less than M.

[0152] In this implementation, after filtering N slice texts (or P paragraph summary texts obtained after processing by the large language model LLM) in step B to obtain M slice texts, for slice contents with strong correlation or contextual relationship, the M slice texts can be further fused according to the correlation of the text content (the specific method is not limited and can be set according to actual conditions, such as directly splicing the slice texts, etc.) to obtain K slice texts, so as to ensure the relevance and integrity of the contents of these K slice texts.

[0153] It should be noted that, in the case where a query text needs to be constructed based on the original text of the article, for example, when the query text is to explain a word in a poem (i.e., the original text), the processing of the above steps BC can be omitted to ensure the accuracy of the text content.

[0154] Step D: Use the K slice texts to construct a third prompt instruction prompt, input it into the preset large language model, and obtain the fourth target question text and the fourth target answer text output by the large language model.

[0155] In this implementation, after obtaining K slice texts through step C (or N slice texts obtained through step A without processing through step BC), these K slice texts (or N slice texts obtained through step A without processing through step BC) can be further integrated into the third prompt instruction prompt, and the core of constructing the question text in the third prompt instruction prompt is to use all or part of the information in the document as the answer to the question, and then input the third prompt instruction prompt into the preset large language model LLM, giving the large language model LLM sufficient freedom to use the large language model LLM to mine detailed and professional question texts (query) and answer texts (answer), as the fourth target question text and fourth target answer text output by the model, to execute the subsequent step S105. It can be seen that since the fourth target question text (query) is based on a specific fragment of the text corpus data (such as K slice texts), it grasps the content details to raise questions, so that the generated question text (query) is more professional and detailed, and better fits the academic professional question and answer scenario in the real scene. Among them, the advantage of slicing text is that it can construct a more detailed and professional question text (query) and obtain the corresponding answer (answer) based on the specific content of the document, that is, the fourth target question text and the fourth target answer text

[0156] For example: The example of the third prompt command prompt is as follows:

[0157] "##Role

[0158] You are a knowledgeable scholar of ancient poetry and know it like the back of your hand.

[0159] ##Task Description

[0160] Now I will give you some ancient poems {content} and their corresponding ancient poems {title}.

[0161] Please construct a question-answer pair based on the given poem and its title, and be sure to strictly abide by the following requirements:

[0162] 1. The extracted questions must contain a subject, predicate, and object, and must not contain unclear words such as this study or this experiment.

[0163] 2. Please make sure that the answers to the extracted questions can be found in the original text.

[0164] 3. Make sure each question and answer contains sufficient complexity and depth to avoid generating simple restatements of facts.

[0165] 4. Answers should provide additional information or explanation, not cite the question.

[0166] The specific requirement is that the correct answers to these constructed questions {query} are the full text of the ancient poem {title} and its appreciation.

[0167] Output format

[0168] After marking the query command (query), please output the results in the following JSON format:

[0169]

[0170] Please follow the above requirements to process the query you received: {content, title}. ”

[0171] When the example of the third prompt instruction prompt is input into the preset large language model LLM, an example of the fourth target question text and the fourth target answer text output by the large language model is obtained as follows:

[0172]

[0173]

[0174]

[0175] The key points of constructing the third prompt include: 1) the generated query must be clear and easy to understand, and have sufficient complexity and depth; 2) no unclear words should appear in the query; 3) ensure that all current document information (i.e. the sliced ​​text content information corresponding to the text corpus data) can accurately answer the query, thereby improving the meticulousness and professionalism of the query (i.e. the fourth target question text).

[0176] It can be seen that for any length of text corpus data, the length of the slice can be adjusted through the above segmentation method, and the problem of the prompt being too long due to the length of the document content can be avoided. Thus, it can adapt to the generation of question-answer pairs of any length of text.

[0177] S105: Generate question-answer pairs using the first target question text and the first target answer text, the second target question text and the second target answer text, the third target question text and the third target answer text, and the fourth target question text and the fourth target answer text.

[0178] In this embodiment, after obtaining the first target question text and the first target answer text through step S101, obtaining the second target question text and the second target answer text through step S102, obtaining the third target question text and the third target answer text through step S103, and obtaining the fourth target question text and the fourth target answer text through step S104, further, the first target question text, the second target question text, the third target question text and the fourth target question text can be clustered by using an existing or future text clustering method. For example, an existing or future embedding vector extraction method (such as extracting text embedding vectors using an embedding model, etc.) can be used to extract the embedding vectors corresponding to each target question text. By calculating the vector similarity, the target question texts having a similarity higher than a preset threshold (the specific value is not limited, such as a value of 0.8, etc.) are clustered into one category, and a total of T grouping clustering results are obtained, where T is a positive integer greater than 0. Then, T target question texts are extracted from the T grouping and clustering results, and T question-answer pairs are generated using these T target question texts and their corresponding target answer texts. In this way, the coverage of the generated question-answer pairs is guaranteed, and duplicate question-answer pair data is eliminated, achieving an ideal question-answer pair generation effect.

[0179] In addition, an optional implementation method is that in order to further enrich the number of question-answer pairs and achieve data enhancement, a similar method to step S102 can be used to construct a fourth prompt instruction prompt using the third target question text, and input it into a preset large language model LLM to rewrite the third target question text using the large language model LLM to obtain the fifth target question text and the fifth target answer text output by the model. The specific implementation process will not be repeated here, and can be processed by referring to the implementation process of step S102.

[0180] And / or, a method similar to step S102 may be adopted to construct a fifth prompt instruction prompt using the fourth target question text, and input it into a preset large language model, LLM, so as to rewrite the fourth target question text using the large language model LLM to obtain a sixth target question text and a sixth target answer text output by the model. The specific implementation process will not be repeated here, and may refer to the implementation process of step S102 for processing.

[0181] On this basis, we can use the first target question text and the first target answer text, the second target question text and the second target answer text, the third target question text and the third target answer text, the fourth target question text and the fourth target answer text, as well as the fifth target question text and the fifth target answer text, and / or the sixth target question text and the sixth target answer text, these richer and larger number of question texts and answer texts, through clustering methods such as embedding models, to dig out more specific and professional questions, and then generate more question and answer pairs with wider coverage. The specific implementation process will not be repeated here.

[0182] In this way, by executing the above steps S101-S105, the generation efficiency, diversity, professionalism and operability of question-answer pairs can be improved, so that users can simply operate and use them flexibly in a short time.

[0183] In summary, the present embodiment provides a method for generating a question-answer pair, which first acquires text corpus data; and generates a first target question text and a first target answer text in a preset format according to the text corpus data; then uses the first target question text to construct a first prompt instruction prompt, which is input into a preset large language model, and the first target question text is rewritten to obtain a second target question text and a second target answer text output by the large language model; and, uses the text corpus data to construct a second prompt instruction prompt, which is input into a preset large language model, and a third target question text and a third target answer text output by the large language model are obtained; then, the text corpus data is segmented, and a third prompt instruction prompt is constructed using the segmentation result, which is input into the preset large language model, and a fourth target question text and a fourth target answer text output by the large language model are obtained; and then the first target question text and the first target answer text, the second target question text and the second target answer text, the third target question text and the third target answer text, and the fourth target question text and the fourth target answer text can be used to generate a question-answer pair.

[0184] It can be seen that since this application constructs different prompts about text corpus data from multiple perspectives and inputs them into a preset large language model, the model can output target question texts and target answer texts from different perspectives to form question-answer pairs. This can improve the generation efficiency, diversity, professionalism, and specificity of question-answer pairs, and achieve an ideal question-answer pair generation effect.

[0185] Second embodiment

[0186] This embodiment will introduce a question-answer pair generation device. For related content, please refer to the above method embodiment.

[0187] See also Figure 2 , is a schematic diagram of the composition of a question-answer pair generating device provided in this embodiment, the device 200 includes:

[0188] The acquisition unit 201 is used to acquire text corpus data; and generate a first target question text and a first target answer text in a preset format according to the text corpus data;

[0189] A first input unit 202 is used to construct a first prompt instruction prompt using the first target question text, input it into a preset large language model, rewrite the first target question text, and obtain a second target question text and a second target answer text output by the large language model;

[0190] The second input unit 203 is used to construct a second prompt instruction prompt using the text corpus data, and input it into a preset large language model to obtain a third target question text and a third target answer text output by the large language model;

[0191] The third input unit 204 is used to segment the text corpus data, and construct a third prompt instruction prompt using the segmentation result, and input it into a preset large language model to obtain a fourth target question text and a fourth target answer text output by the large language model;

[0192] The generation unit 205 is used to generate question-answer pairs using the first target question text and the first target answer text, the second target question text and the second target answer text, the third target question text and the third target answer text, and the fourth target question text and the fourth target answer text.

[0193] In an implementation of this embodiment, the device further includes:

[0194] a fourth input unit, configured to construct a fourth prompt instruction prompt using the third target question text, input the fourth prompt instruction prompt into a preset large language model, rewrite the third target question text, and obtain a fifth target question text and a fifth target answer text output by the large language model;

[0195] And / or, the fourth target question text is used to construct a fifth prompt instruction prompt, which is input into a preset large language model, and the fourth target question text is rewritten to obtain a sixth target question text and a sixth target answer text output by the large language model.

[0196] In an implementation of this embodiment, the generating unit 205 is specifically configured to:

[0197] Question-answer pairs are generated using the first target question text and the first target answer text, the second target question text and the second target answer text, the third target question text and the third target answer text, the fourth target question text and the fourth target answer text, and the fifth target question text and the fifth target answer text, and / or the sixth target question text and the sixth target answer text.

[0198] In an implementation of this embodiment, the third input unit 204 includes:

[0199] A segmentation subunit, used to segment the text corpus data according to a preset segmentation method to obtain N slice texts; N is a positive integer greater than 0;

[0200] A filtering subunit, configured to filter the N slice texts based on the relevance between the title and the content to obtain M slice texts; wherein M is a positive integer less than N;

[0201] A fusion subunit is used to fuse the M slice texts based on the relevance of the text contents to obtain K slice texts; K is a positive integer less than M;

[0202] The first input subunit is used to construct a third prompt instruction prompt using the K slice texts, and input it into a preset large language model to obtain a fourth target question text and a fourth target answer text output by the large language model.

[0203] In an implementation of this embodiment, the preset segmentation method includes at least one of segmentation based on paragraphs, segmentation based on sentences, and segmentation based on preset byte lengths.

[0204] In an implementation of this embodiment, the third input unit 204 further includes:

[0205] The second input subunit is used to construct a sixth prompt instruction prompt using the N slice texts, input the sixth prompt instruction prompt into a preset large language model, perform highlight extraction on the N slice texts, and obtain P paragraph summary texts within a preset number of words output by the large language model; P is a positive integer less than N;

[0206] The filtering subunit is specifically used for:

[0207] Based on the relevance between the title and the content, the P paragraph summary texts are filtered to obtain M slice texts.

[0208] In an implementation of this embodiment, the generating unit 205 includes:

[0209] A clustering subunit, configured to perform clustering processing on the first target question text, the second target question text, the third target question text and the fourth target question text to obtain T grouping clustering results; wherein T is a positive integer greater than 0;

[0210] A generating subunit is used to classify and extract T target question texts from the T grouping clustering results, and generate T question-answer pairs using the T target question texts and their corresponding target answer texts.

[0211] Furthermore, an embodiment of the present application also provides a question-answer pair generating device, comprising: a processor, a memory, and a system bus;

[0212] The processor and the memory are connected via the system bus;

[0213] The memory is used to store one or more programs, and the one or more programs include instructions, which, when executed by the processor, enable the processor to execute any one of the implementation methods of the above-mentioned question-answer pair generation method.

[0214] Furthermore, an embodiment of the present application also provides a computer-readable storage medium, in which instructions are stored. When the instructions are executed on a terminal device, the terminal device executes any one of the implementation methods of the above-mentioned question-answer pair generation method.

[0215] Furthermore, an embodiment of the present application also provides a computer program product, which, when executed on a terminal device, enables the terminal device to execute any one of the implementation methods of the above-mentioned question-answer pair generation method.

[0216] It can be known from the description of the above implementation mode that those skilled in the art can clearly understand that all or part of the steps in the above-mentioned embodiment method can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a storage medium such as ROM / RAM, a disk, an optical disk, etc., including several instructions for enabling a computer device (which can be a personal computer, a server, or a network communication device such as a media gateway, etc.) to execute the methods described in the various embodiments of the present application or certain parts of the embodiments.

[0217] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0218] It should also be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0219] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for generating question-answer pairs, characterized in that: include: Get text corpus data; and generating a first target question text and a first target answer text in a preset format according to the text corpus data; Using the first target question text to construct a first prompt instruction prompt, inputting it into a preset large language model, rewriting the first target question text, and obtaining a second target question text and a second target answer text output by the large language model; Using the text corpus data to construct a second prompt instruction prompt, inputting it into a preset large language model, and obtaining a third target question text and a third target answer text output by the large language model; Segment the text corpus data, and construct a third prompt instruction prompt using the segmentation result, and input it into a preset large language model to obtain a fourth target question text and a fourth target answer text output by the large language model; Question-answer pairs are generated using the first target question text and the first target answer text, the second target question text and the second target answer text, the third target question text and the third target answer text, and the fourth target question text and the fourth target answer text.

2. The method according to claim 1, characterized in that Before generating a question-answer pair using the first target question text and the first target answer text, the second target question text and the second target answer text, the third target question text and the third target answer text, and the fourth target question text and the fourth target answer text, the method further includes: Using the third target question text to construct a fourth prompt instruction prompt, inputting it into a preset large language model, rewriting the third target question text, and obtaining a fifth target question text and a fifth target answer text output by the large language model; And / or, the fourth target question text is used to construct a fifth prompt instruction prompt, which is input into a preset large language model, and the fourth target question text is rewritten to obtain a sixth target question text and a sixth target answer text output by the large language model.

3. The method according to claim 2, characterized in that The step of generating a question-answer pair by using the first target question text and the first target answer text, the second target question text and the second target answer text, the third target question text and the third target answer text, and the fourth target question text and the fourth target answer text comprises: Question-answer pairs are generated using the first target question text and the first target answer text, the second target question text and the second target answer text, the third target question text and the third target answer text, the fourth target question text and the fourth target answer text, and the fifth target question text and the fifth target answer text, and / or the sixth target question text and the sixth target answer text.

4. The method according to claim 1, characterized in that The segmentation process is performed on the text corpus data, and the segmentation result is used to construct a third prompt instruction prompt, which is input into a preset large language model to obtain a fourth target question text and a fourth target answer text output by the large language model, including: According to a preset segmentation method, the text corpus data is segmented to obtain N slice texts, where N is a positive integer greater than 0; Based on the correlation between the title and the content, the N slice texts are filtered to obtain M slice texts; M is a positive integer less than N; Based on the relevance of the text content, the M slice texts are fused to obtain K slice texts; K is a positive integer less than M; A third prompt instruction prompt is constructed using the K slice texts, and is input into a preset large language model to obtain a fourth target question text and a fourth target answer text output by the large language model.

5. The method according to claim 4, characterized in that The preset segmentation method includes at least one of segmentation based on paragraphs, segmentation based on sentences, and segmentation based on preset byte lengths.

6. The method according to claim 4, characterized in that After the text corpus data is segmented according to the preset segmentation method to obtain N sliced ​​texts, the method further includes: The N sliced ​​texts are used to construct a sixth prompt instruction prompt, which is input into a preset large language model, and the N sliced ​​texts are highlighted and extracted to obtain P paragraph summary texts within a preset number of words output by the large language model; P is a positive integer less than N; Based on the correlation between the title and the content, the N slice texts are filtered to obtain M slice texts, including: Based on the relevance between the title and the content, the P paragraph summary texts are filtered to obtain M slice texts.

7. The method according to claim 1, characterized in that The step of generating a question-answer pair by using the first target question text and the first target answer text, the second target question text and the second target answer text, the third target question text and the third target answer text, and the fourth target question text and the fourth target answer text comprises: Performing clustering processing on the first target question text, the second target question text, the third target question text and the fourth target question text to obtain T grouping clustering results; wherein T is a positive integer greater than 0; T target question texts are extracted from the T grouping clustering results by classification, and T question-answer pairs are generated using the T target question texts and their corresponding target answer texts.

8. A question-answer pair generating device, characterized in that: include: An acquisition unit, used for acquiring text corpus data; and generating a first target question text and a first target answer text in a preset format according to the text corpus data; A first input unit is used to construct a first prompt instruction prompt using the first target question text, input it into a preset large language model, rewrite the first target question text, and obtain a second target question text and a second target answer text output by the large language model; A second input unit is used to construct a second prompt instruction prompt using the text corpus data, and input it into a preset large language model to obtain a third target question text and a third target answer text output by the large language model; A third input unit is used to segment the text corpus data, and construct a third prompt instruction prompt using the segmentation result, and input it into a preset large language model to obtain a fourth target question text and a fourth target answer text output by the large language model; A generating unit is used to generate question-answer pairs using the first target question text and the first target answer text, the second target question text and the second target answer text, the third target question text and the third target answer text, and the fourth target question text and the fourth target answer text.

9. A question-answer pair generating device, characterized in that: include: Processor, memory, system bus; The processor and the memory are connected via the system bus; The memory is used to store one or more programs, wherein the one or more programs include instructions, and when the instructions are executed by the processor, the processor executes the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when the instructions are executed on a terminal device, the terminal device executes the method according to any one of claims 1 to 7.

Images