CN118377881A - Intelligent question answering method, system, device, computer equipment and readable storage medium - Google Patents

Abstract

The present application relates to an intelligent question-answering method, system, device, computer equipment and readable storage medium. The method comprises: obtaining a question text input by a user; searching a preset knowledge base based on the similarity between a first vector representation corresponding to the question text and each second vector representation existing in a preset knowledge base, and obtaining multiple candidate results corresponding to the question text; constructing a task prompt template based on the first vector representation and each candidate result; analyzing and processing the correlation between the first vector representation and the candidate results in the task prompt template based on a large language model, and obtaining a target result corresponding to the question text. The use of this method can improve the accuracy of the intelligent question-answering method in replying.

Description

Intelligent question answering method, system, device, computer equipment and readable storage medium

Technical Field

The present application relates to the field of artificial intelligence technology, and in particular to an intelligent question-answering method, system, device, computer equipment, and readable storage medium.

Background technique

As the business of an enterprise develops, factory employees need to receive training before taking up their posts to clarify the contents of the operating procedure (SOP) manual. Since the operating procedure manual is usually large, new employees need to spend a lot of time searching when faced with specific problems.

In traditional technology, enterprises can build a question-and-answer system for the corresponding field of the factory, and use a rule-based or keyword matching method to search and match the questions raised by users. First, the traditional question-and-answer system needs to determine the keywords contained in the questions raised by users, and search based on the keywords and element data such as file names or tags in the document management system, and use the files corresponding to the keywords as the answers to the questions raised by users.

However, in traditional question-and-answer systems, due to the limitations of rule-based or keyword-based matching methods, it is difficult for the question-and-answer system to determine the semantics and intentions of questions asked by different users. In addition, the answers output by the question-and-answer system are in the form of files with more content and lower relevance to the questions asked by users, resulting in poor accuracy of the answers output by the traditional question-and-answer system.

Summary of the invention

Based on this, it is necessary to provide an intelligent question-answering method, system, device, computer equipment and readable storage medium to address the above technical issues.

In a first aspect, the present application provides an intelligent question-answering method, comprising:

Get the question text entered by the user;

Based on the similarity between the first vector representation corresponding to the question text and each second vector representation existing in the preset knowledge base, a search is performed in the preset knowledge base to obtain multiple candidate results corresponding to the question text;

constructing a task prompt template according to the first vector representation and each of the candidate results;

Based on the large language model, the correlation between the first vector representation and the candidate result in the task prompt template is analyzed and processed to obtain a target result corresponding to the question text.

In one embodiment, the similarity between the first vector representation corresponding to the question text and each second vector representation existing in the preset knowledge base is searched in the preset knowledge base to obtain multiple candidate results corresponding to the question text, including:

Performing vectorization processing on the question text according to a text vectorization model to obtain a first vector representation corresponding to the question text;

Calculating the similarity between the first vector representation and each second vector representation in a preset knowledge base based on a preset vector similarity algorithm;

Each second vector representation whose similarity is greater than a preset threshold is taken as a candidate result corresponding to the question text.

In one embodiment, the analyzing and processing the association relationship between the first vector representation and the candidate result in the task prompt template based on the large language model to obtain the target result corresponding to the question text includes:

Based on the large language model, semantic understanding is performed on the first vector representation and the candidate result in the task prompt template to obtain first semantic information corresponding to the question text and second semantic information corresponding to the candidate result;

Performing a correlation analysis on the first semantic information and the second semantic information based on the large language model to obtain semantic correlation information;

The candidate results are screened based on the semantic association information to obtain target candidate results, and a target result is generated based on the target candidate results.

In one of the embodiments, after screening the candidate results based on the semantic association information to obtain a target candidate result, and generating a target result based on the target candidate result, the method further includes:

Determine the text block tag corresponding to the target candidate result in the preset knowledge base;

Determine the target file path or link of the target candidate result according to the correspondence between the text block mark and the file path;

The target file path or the link is fed back.

In one embodiment, before obtaining the question text input by the user, the method further includes:

Acquire multimodal knowledge data;

Slicing the multimodal knowledge data to obtain a plurality of paragraph slices;

Performing data processing on each of the paragraph slices based on a text vectorization model to obtain a second vector representation corresponding to each of the paragraph slices, and determining a text block tag corresponding to each of the second vector representations;

A preset knowledge base is constructed according to the second vector representation and the text block mark corresponding to the second vector representation.

In one embodiment, the slicing process of the multimodal knowledge data to obtain a plurality of paragraph slices includes:

Converting the multimodal knowledge data into text data based on a multimodal model;

The text data is segmented into paragraphs based on paragraph marks to obtain paragraph slices and text block marks corresponding to the paragraph slices.

In one embodiment, the data processing of each paragraph slice based on the text vectorization model to obtain a second vector representation corresponding to each paragraph slice, and determining a text block mark corresponding to each second vector representation, includes:

For each of the paragraph slices, segment the paragraph slices according to a segmentation tool to obtain a word sequence of each subword in each of the paragraph slices;

Mapping the word sequence contained in each paragraph slice based on a text vectorization model to obtain an initial second vector representation;

The initial second vector representations contained in each of the paragraph slices are merged or weighted averaged to obtain the second vector representation corresponding to each of the paragraph slices.

In a second aspect, the present application also provides an intelligent question-answering system, comprising:

The client terminal is used to obtain the question text and feed the question text back to the server terminal;

The server is used to obtain a question text input by a user; search the preset knowledge base based on the similarity between the first vector representation corresponding to the question text and each second vector representation existing in the preset knowledge base to obtain multiple candidate results corresponding to the question text; construct a task prompt template based on the first vector representation and each candidate result; analyze and process the association relationship between the first vector representation and the candidate results in the task prompt template based on a large language model to obtain a target result corresponding to the question text.

In a third aspect, the present application also provides an intelligent question-answering device, comprising:

The first acquisition module is used to obtain the question text input by the user;

A retrieval module, configured to search the preset knowledge base based on the similarity between the first vector representation corresponding to the question text and each second vector representation existing in the preset knowledge base, to obtain a plurality of candidate results corresponding to the question text;

A first construction module, configured to construct a task prompt template according to the first vector representation and each of the candidate results;

The analysis module is used to analyze and process the association between the first vector representation and the candidate result in the task prompt template based on the large language model to obtain the target result corresponding to the question text.

In one embodiment, the retrieval module is specifically used to perform vectorization processing on the question text according to a text vectorization model to obtain a first vector representation corresponding to the question text;

Calculating the similarity between the first vector representation and each second vector representation in a preset knowledge base based on a preset vector similarity algorithm;

Each second vector representation whose similarity is greater than a preset threshold is taken as a candidate result corresponding to the question text.

In one embodiment, the analysis module is specifically used to perform semantic understanding on the first vector representation and the candidate result in the task prompt template based on a large language model, to obtain first semantic information corresponding to the question text and second semantic information corresponding to the candidate result;

Performing a correlation analysis on the first semantic information and the second semantic information based on the large language model to obtain semantic correlation information;

The candidate results are screened based on the semantic association information to obtain target candidate results, and a target result is generated based on the target candidate results.

In one embodiment, the device further comprises:

A first determination module is used to determine the text block mark corresponding to the target candidate result in the preset knowledge base;

A second determination module, configured to determine a target file path or link of the target candidate result according to a correspondence between the text block mark and the file path;

The feedback module is used to provide feedback on the target file path or the link.

In one embodiment, the device further comprises:

A second acquisition module is used to acquire multimodal knowledge data;

A slicing module, used for slicing the multimodal knowledge data to obtain a plurality of paragraph slices;

A vectorization processing module, used for performing data processing on each of the paragraph slices based on a text vectorization model to obtain a second vector representation corresponding to each of the paragraph slices, and determining a text block mark corresponding to each of the second vector representations;

The second construction module is used to construct a preset knowledge base according to the second vector representation and the text block mark corresponding to the second vector representation.

In one of the embodiments, the slicing module is specifically used to convert the multimodal knowledge data into text data based on a multimodal model;

The text data is segmented into paragraphs based on paragraph marks to obtain paragraph slices and text block marks corresponding to the paragraph slices.

In one embodiment, the vectorization processing module is specifically used to perform word segmentation processing on each paragraph slice according to a word segmentation tool to obtain a word sequence of each subword in each paragraph slice;

Mapping the word sequence contained in each paragraph slice based on a text vectorization model to obtain an initial second vector representation;

The initial second vector representations contained in each of the paragraph slices are merged or weighted averaged to obtain the second vector representation corresponding to each of the paragraph slices.

In a fourth aspect, the present application further provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

Get the question text entered by the user;

Based on the similarity between the first vector representation corresponding to the question text and each second vector representation existing in the preset knowledge base, a search is performed in the preset knowledge base to obtain multiple candidate results corresponding to the question text;

constructing a task prompt template according to the first vector representation and each of the candidate results;

Based on the large language model, the correlation between the first vector representation and the candidate result in the task prompt template is analyzed and processed to obtain a target result corresponding to the question text.

In a fifth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the following steps are implemented:

Get the question text entered by the user;

Based on the similarity between the first vector representation corresponding to the question text and each second vector representation existing in the preset knowledge base, a search is performed in the preset knowledge base to obtain multiple candidate results corresponding to the question text;

constructing a task prompt template according to the first vector representation and each of the candidate results;

Based on the large language model, the correlation between the first vector representation and the candidate result in the task prompt template is analyzed and processed to obtain a target result corresponding to the question text.

In a sixth aspect, the present application further provides a computer program product, including a computer program, which implements the following steps when executed by a processor:

Get the question text entered by the user;

Based on the similarity between the first vector representation corresponding to the question text and each second vector representation existing in the preset knowledge base, a search is performed in the preset knowledge base to obtain multiple candidate results corresponding to the question text;

constructing a task prompt template according to the first vector representation and each of the candidate results;

Based on the large language model, the correlation between the first vector representation and the candidate result in the task prompt template is analyzed and processed to obtain a target result corresponding to the question text.

The above-mentioned intelligent question-answering method, system, device, computer equipment and readable storage medium can retrieve candidate results corresponding to the question text according to the similarity between the first vector representation corresponding to the question text and each second vector representation existing in the preset knowledge base, and can quickly determine the candidate results with semantic similarity to the question text in the second vector representation representing a large amount of multimodal knowledge data. Furthermore, the candidate results are combined with the question text to construct a task prompt template, so that the large language model can clarify the semantics of the question text in the task prompt template based on the association between the question text and the candidate results, and generate a target result that fits the context of the question text. The candidate results in the task prompt template can also limit the output range of the large language model, so that the large language model is more focused on generating content that is more relevant to the question text, and provide more accurate and specific target results, thereby avoiding the large language model from generating generalized descriptions and improving the accuracy of the intelligent question-answering method in replying.

BRIEF DESCRIPTION OF THE DRAWINGS

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

FIG1 is a schematic diagram of a flow chart of an intelligent question-answering method in one embodiment;

FIG2 is a schematic diagram of a process of retrieving candidate results in a preset knowledge base in one embodiment;

FIG3 is a schematic diagram of a process of generating a target result by a large language model in one embodiment;

FIG4 is a schematic diagram of a process of providing feedback on extended reference content of a target result in one embodiment;

FIG5 is a schematic diagram of a process for constructing a preset knowledge base in one embodiment;

FIG6 is a schematic diagram of building a preset knowledge base in a specific embodiment;

FIG7 is a schematic diagram of a process of slicing multimodal knowledge data in one embodiment;

FIG8 is a schematic diagram of a process of vectorizing multimodal knowledge data in one embodiment;

FIG9 is a flow chart of an example of an intelligent question-answering method according to an embodiment;

FIG10 is a block diagram of a smart question-answering device according to an embodiment;

FIG. 11 is a diagram showing the internal structure of a computer device in one embodiment.

Detailed ways

In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

In one embodiment, as shown in FIG1 , an intelligent question-answering method is provided. This embodiment is illustrated by applying the method to a server. It is understandable that the method can also be applied to a terminal, or to a system including a terminal and a server, and implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

Step 102: Obtain the question text input by the user.

In the embodiment of the present application, before new factory employees of the enterprise take up their posts, they need to receive training and understand the contents of the operating procedures (SOP) manual in industrial scenarios, and in their daily work, when faced with specific problems, they need to solve the problems according to the provisions in the operating procedures manual. At this time, the factory employees ask questions as users, that is, input the current problems or query solutions in certain situations into the terminal, and then the server receives and obtains the question text entered by the user.

In an optional embodiment, the user can also ask questions by voice, and the terminal responds to the user's question operation, receives audio data containing the user's question content, uploads the audio data to a pre-set voice recognition service, converts the audio data into text data through the voice recognition service, and then obtains the question text, and sends the question text to the server. Among them, a noise reduction function can also be added to the microphone. This means that users can interact with the system through voice without being disturbed by on-site noise, and the interactive method of voice input improves the operating efficiency at the work site, especially suitable for scenes with inconvenient hand operation or limited vision. Through voice input, workers can ask questions more conveniently when operating machines or carrying items, making information acquisition faster and more natural.

In an optional embodiment, due to the special requirements of time-sensitive processes and SOPs in the factory, the preset knowledge base has a preset expiration date, and the administrator of the preset knowledge base can flexibly set the expiration time of the knowledge data while maintaining the preset knowledge base. Specifically, each piece of knowledge data in the preset knowledge base contains a corresponding expiration time. For knowledge data that exceeds the expiration time, the server can delete it, or, when retrieving candidate results later, the server only retrieves the second vector representation corresponding to the knowledge data that has not exceeded the expiration time, to prevent the generation of target results containing expired knowledge data, thereby ensuring the accuracy of the intelligent question-answering method's response.

In an optional embodiment, the server clearly divides the contents of the preset knowledge base that users can access based on their roles and permission levels. For example, factory operators may only be able to access basic knowledge data related to operating procedures, equipment usage, etc., engineers can access more in-depth technical information, and managers can retrieve core data such as finance and operations.

Step 104 , searching the preset knowledge base based on the similarity between the first vector representation corresponding to the question text and each second vector representation existing in the preset knowledge base, to obtain a plurality of candidate results corresponding to the question text.

In the embodiment of the present application, the preset knowledge base includes a second vector representation corresponding to pre-stored knowledge data, and the knowledge data may be related instruction documents, technical documents, FAQs, etc. in an industrial operation environment obtained from a variety of data sources. These documents may contain various information related to industrial operations, such as operating procedures, equipment instructions, safety specifications, etc.

After the server obtains the question text input by the user, in order to determine the knowledge data that has a certain relevance to the content of the current question text, the knowledge data that has relevance to the question text is used as a reference, and the terminal can vectorize the question text according to the embedding model to obtain the first vector representation. Among them, the model structure of the embedding model can be Word2Vec (Word to Vector, a group of related models used to generate word vectors), GloVe (Global Vectors for Word Representation, a word representation tool based on global word frequency statistics), FastText (a fast text classifier), etc.

The server traverses the second vector representations stored in the preset knowledge base, and calculates the similarity between the first vector representation and each second vector representation through a similarity algorithm (e.g., a cosine similarity algorithm). Based on the similarity between the first vector representation and the second vector representation, a candidate result related to the question text can be determined. The similarity between the first vector representation and the second vector representation can characterize the correlation between the question text and the knowledge data stored in the preset knowledge base. The basis for determining the second vector representation as a candidate result can be to analyze the similarity and compare the similarity with a preset threshold. The server can use the second vector representation greater than the threshold as a candidate result, or the terminal can also rank the similarity to obtain a ranking result of the similarity, and determine the second vector representation corresponding to the TopK similarities in the ranking result as a candidate result.

Step 106: construct a task prompt template according to the first vector representation and each candidate result.

In an embodiment of the present application, the task prompt template may be structured data, which is used to guide the model to generate a reply text related to a specific topic, content, or focus. Specifically, the server may concatenate the first vector representation with multiple candidate results (the second vector representation is obtained by screening the second vector representation based on the similarity between the first vector representation and the second vector representation) to construct a task prompt template.

In an optional embodiment, a task prompt template can be dynamically generated through the user's question text. Based on pre-stored template keywords and template fragments, the template keywords and template fragments are matched according to the keywords or phrases contained in the question text, and the keywords and template fragments are combined with the keywords or phrases originally contained in the question text according to a pre-set template structure to construct a task prompt template.

Step 108: Analyze and process the correlation between the first vector representation in the task prompt template and the candidate results based on the large language model to obtain the target result corresponding to the question text.

In the embodiment of the present application, the server uses a large language model (Large Language Model) to perform semantic understanding on the text in the task prompt template, understands the relationship between the question text and the candidate results through the large language model, and uses the candidate results provided in the task prompt template as a basis to extract the information or answer (i.e., the target result) most relevant to the question text from the candidate results. The information or answer most relevant to the question text is fed back to the terminal for output display.

In the above-mentioned intelligent question-answering method, the candidate results corresponding to the question text are retrieved according to the similarity between the first vector representation corresponding to the question text and each second vector representation existing in the preset knowledge base, and the candidate results with semantic similarity to the question text can be quickly determined in the second vector representation of a large amount of multimodal knowledge data. Furthermore, the candidate results are combined with the question text to construct a task prompt template, so that the large language model can clarify the semantics of the question text in the task prompt template based on the association between the question text and the candidate results, and generate a target result that fits the context of the question text. The candidate results in the task prompt template can also limit the output range of the large language model, so that the large language model can focus more on generating content that is more relevant to the question text, and provide more accurate and specific target results, thereby avoiding the large language model from generating generalized descriptions and improving the accuracy of the intelligent question-answering method's responses.

In an exemplary embodiment, as shown in FIG2 , step 104 includes steps 202 to 206. Among them:

Step 202: vectorize the question text according to the text vectorization model to obtain a first vector representation corresponding to the question text.

In an embodiment of the present application, the server vectorizes the question text according to the text vectorization model, and maps the question text into a high-dimensional first vector representation, wherein the number of dimensions of the first vector representation is determined by the model structure and parameters of the text vectorization model. The high-dimensional first vector representation can contain richer semantic information, and each dimension can be regarded as a feature of the question text. When a vector is used to represent the text, each dimension may correspond to a word or some linguistic feature. For example, the feature may be the frequency of occurrence of a word, part of speech, contextual information, etc.

Step 204: Calculate the similarity between the first vector representation and each second vector representation in the preset knowledge base based on a preset vector similarity algorithm.

In the embodiment of the present application, the preset vector similarity algorithm may be a similarity calculation method based on cosine similarity or Euclidean distance. According to the preset vector similarity algorithm, the server may calculate the similarity between the first vector representation and the second vector representation through the dot product between the first vector representation and the second vector representation. The similarity calculation evaluates the degree of association between the question text and the knowledge data by comparing the degree of similarity between the first vector representation and the second vector representation, specifically by measuring the difference in semantic similarity between the first vector representation and the second vector representation through cosine similarity or Euclidean distance, thereby obtaining the degree of association between the question text and the knowledge data.

Step 206: taking each second vector representation having a similarity greater than a preset threshold as a candidate result corresponding to the question text.

In the embodiment of the present application, different question texts may have different semantic similarities with documents in the knowledge base. By setting a similarity threshold, knowledge data with a low semantic relevance to the question text can be filtered out to a certain extent. After calculating the similarity between the first vector representation of the question text and the second vector representation of each candidate text, the server will filter out candidate results with a similarity greater than a preset threshold, which represents knowledge data with semantics similar to the question text of the question raised by the user.

In an optional embodiment, the server may also perform semantic matching based on historical conversation records, calculate similarity between the vector representation corresponding to the historical conversation record and the first vector corresponding to the question text, and use the vector representation corresponding to the historical conversation record whose similarity is greater than a preset threshold as a candidate result to provide a reference for data analysis of the large language model.

In this embodiment, by using a preset knowledge base including the second vector representation and a similarity algorithm, knowledge data that is semantically similar to the question text input by the user can be quickly retrieved from a large amount of knowledge data, thereby improving the speed and accuracy of information retrieval.

In an exemplary embodiment, as shown in FIG3 , step 108 includes steps 302 to 306. Among them:

Step 302 , based on the large language model, semantic understanding is performed on the first vector representation and the candidate results in the task prompt template respectively, to obtain first semantic information corresponding to the question text and second semantic information corresponding to the candidate results.

In an embodiment of the present application, the server inputs the task prompt template into the large language model, and the large language model performs semantic understanding on the first vector representation to obtain the first semantic information, which can understand the semantics and intention of the user's question at a deeper level. And the candidate results are semantically understood, and the candidate results are semantically understood at the vocabulary, phrase and sentence levels to obtain the second semantic information. Among them, the first semantic information includes the understanding of the intention, theme, and entities or concepts contained in the question text. The first semantic information can help the large language model understand what the user's question is, and the fields and scopes that the user expects to get. The second semantic information includes the entities, events, concepts involved in the knowledge data, and background information related to the question, which helps the large language model understand the possible answers provided in the candidate results or the context related to the question, and helps to determine which information is most important for answering the question.

Based on the high-dimensional first vector representation and the second vector representation, the large language model can perform semantic understanding from the content of the first vector representation and the second vector representation, and also include contextual analysis of the first vector representation and the second vector representation to understand the meaning of the question text and the candidate results in the context. Through the contextual analysis of the question text and the candidate results, the background of the user's question and the context corresponding to the candidate results can be determined as the basis for generating the target result, thereby ensuring the accuracy of the output content of the large language model.

Step 304: Perform a correlation analysis on the first semantic information and the second semantic information based on the large language model to obtain semantic correlation information.

In an embodiment of the present application, after the first semantic information and the second semantic information are obtained through semantic understanding, the server analyzes the semantic association between the question text and the candidate results through a language model, including associations at the vocabulary, phrase or sentence level, that is, performs a correlation analysis on the first semantic information and the second semantic information, analyzes the degree of association between the question text and the candidate results, including the semantic similarity between the first semantic information and the second semantic information, the logical relationship, co-occurrence relationship or other correlation indicators between the question text and the candidate results, and obtains semantic association information that characterizes the first vector representation and the candidate results with correlation characteristics.

Step 306 , screening the candidate results based on the semantic association information to obtain target candidate results, and generating target results based on the target candidate results.

In an embodiment of the present application, based on the semantic association information of the correlation analysis results between the first vector representation and the candidate results, the large language model can screen the candidate results and select the candidate answers that are most relevant to the question text to exclude some irrelevant or inappropriate candidate results, thereby narrowing the analysis scope of the large language model. Specifically, the large language model first calculates the similarity between the first vector representation of the question text and each second vector representation as a candidate result, and the similarity calculation includes cosine similarity, Euclidean distance, Pearson correlation coefficient or Manhattan distance, etc. Then, the large language model can sort according to the calculated similarity results, put the candidate results with higher similarity in front, and use the candidate results with the highest similarity as a basis for further processing to generate the target result corresponding to the question text. Among them, the process of generating the target result based on the candidate results also includes the process of text repetition, text matching, context analysis and logical reasoning to generate a target result with coherent word order, reasonable logic and fitting the context of the question.

In an optional embodiment, users can provide feedback based on the returned target results to help improve the preset knowledge base, and users can perform maintenance on issues that are not within the scope of the preset knowledge base. After the maintenance is completed, the new content can be added to the preset knowledge base for subsequent retrieval after being reviewed and approved by the reviewer.

In this embodiment, the large language model extracts semantic information from the question text and the candidate results respectively, and analyzes the semantic association between the two, so as to better understand the relevance between the question and the candidate results, so that the large language model can more comprehensively and accurately understand the background of the question and the context of the candidate results, thereby generating more accurate target results and improving the accuracy of the intelligent question-answering method's responses.

In an exemplary embodiment, after the server generates the target result, it can also output and display the basis for generating the target result, as shown in FIG4 , step 306 includes steps 402 to 406. Among them:

Step 402: determine the text block tag corresponding to the target candidate result in the preset knowledge base.

In an embodiment of the present application, in a preset knowledge base, the second vector representation of each knowledge data has its corresponding text block tag. The server determines the text block tag corresponding to the target candidate result based on the correspondence between the target candidate result and the second vector representation and the text block tag obtained by screening the candidate results with a large language model, thereby providing a basis for subsequent positioning and retrieval based on the specific text block tag corresponding to the target candidate result in the preset knowledge base.

Step 404: Determine the target file path or link of the target candidate result according to the correspondence between the text block mark and the file path.

In an embodiment of the present application, the server determines the specific file path or link where the target candidate result is located based on the correspondence between the text block mark and the file path. The correspondence can be a mapping relationship in an index table or a preset knowledge base, which is used to correspond the text block mark with the corresponding file path or link.

Step 406, feeding back the target file path or link.

In the embodiment of the present application, the server feeds back the determined target file path or link to the user terminal. In this way, the user can directly access the relevant knowledge base documents or information through the provided link or file path, thereby obtaining more detailed and comprehensive information support. Among them, the feedback content can be presented in text form, or it can be a clickable link or file path, and the user can jump to the corresponding document, web page or folder after clicking it.

In this embodiment, by clearly outputting the basis of the target result (i.e., the knowledge data or information in the corresponding preset knowledge base), the user can easily view and obtain relevant information, which helps the user to refer to and confirm the authenticity and validity of the answer, avoiding the illusion problem that may be caused by the large language model. By clarifying the basis and source of the target result, the user's trust and reliability in the target result are enhanced, the transparency and credibility of the target result are improved, and the accuracy of the target result is guaranteed.

In an exemplary embodiment, as shown in FIG5 , before step 102, the method further includes steps 502 to 508. Among them:

Step 502: Acquire multimodal knowledge data.

In an embodiment of the present application, the multimodal knowledge data may be relevant instruction documents, technical documents, FAQs, etc. in the industrial operation environment of the factory. The server reads the pre-stored multimodal knowledge data to facilitate further processing of the multimodal knowledge data.

Step 504: Slice the multimodal knowledge data to obtain multiple paragraph slices.

In an embodiment of the present application, the server segments the multimodal knowledge data into paragraphs, and segments the document into paragraph slices according to paragraph marks (for example, line breaks, paragraph tags, page breaks, or pre-set lengths, etc.), with each paragraph slice serving as a basic unit of slicing processing.

Step 506 , performing data processing on each paragraph slice based on the text vectorization model to obtain a second vector representation corresponding to each paragraph slice, and determining a text block mark corresponding to each second vector representation.

In the embodiment of the present application, the preset knowledge base is used as the business knowledge base for explanation. As shown in FIG6 , the server imports the document content, question-answer pairs, corpus or knowledge base containing business text data into the text vectorization model, and vectorizes each paragraph slice through the text vectorization model to obtain the second vector representation corresponding to each paragraph slice, and obtains the business data in vector form. During the paragraph slicing process, the server can pre-store the text block mark corresponding to the paragraph slice after slicing according to the slicing process, characterize the document to which the paragraph slice belongs, and record the position information of the text block in the original document, so as to restore the semantic structure of the original document later. Then, according to the text block mark corresponding to the paragraph slice, the text block mark corresponding to the second vector representation can be determined, providing a basis for constructing a vector index.

Step 508: construct a preset knowledge base according to the second vector representation and the text block mark corresponding to the second vector representation.

In an embodiment of the present application, the server stores the business data in vector form, i.e., the second vector representation and the text block mark corresponding to the second vector representation, and can use a database or file system to store the vector to ensure that it can be efficiently retrieved and used later to construct a preset knowledge base containing the vector index and the second vector representation corresponding to the knowledge data.

In this embodiment, by constructing a preset knowledge base by slicing and vectorizing the multimodal knowledge data, the efficiency of retrieving candidate results in the intelligent question-answering method can be improved, the relevance with the original multimodal knowledge data can be maintained, and the integrity and accuracy of the storage of the multimodal knowledge data can be ensured.

In an exemplary embodiment, as shown in FIG7 , step 504 includes steps 702 to 704. Among them:

Step 702: Convert the multimodal knowledge data into text data based on the multimodal model.

In the embodiment of the present application, multimodal knowledge data can be obtained from knowledge sources in various formats, such as Word (word processor application) documents, TXT (a text format) documents, PDF (Portable Document Format) documents or pictures, etc. The server can convert the multimodal knowledge data in different formats according to the targeted document loader or multimodal model, and convert it into text data in a plain text format that can be understood by the large language model. For example, for multimodal knowledge data in PDF file format, the server can use a PDF extractor to extract the text content in the multimodal knowledge data; for multimodal knowledge data in picture format, the server can use OCR (Optical Character Recognition) technology to identify and convert the text information therein to obtain text data.

In an optional embodiment, after the server completes the format conversion for the multimodal knowledge data, the server may preprocess the text data after the format conversion, including removing special characters, blank characters, etc. in the text data to reduce noise data.

Step 704, segment the text data into paragraphs based on the paragraph marks to obtain paragraph slices and text block marks corresponding to the paragraph slices.

In the embodiment of the present application, the server performs paragraph segmentation on the text data according to the same principle as the paragraph segmentation on the multimodal knowledge data in step 504, and obtains paragraph slices and text block marks corresponding to the paragraph slices. The specific process of paragraph segmentation will not be described in detail in this embodiment.

In this embodiment, by segmenting multimodal knowledge data in a text data format into paragraphs to obtain paragraph slices and text block tags corresponding to the paragraph slices, the efficiency of information retrieval and the accuracy of subsequent text positioning can be improved.

In an exemplary embodiment, as shown in FIG8 , step 506 includes steps 802 to 806. Among them:

Step 802: for each paragraph slice, perform word segmentation processing on the paragraph slice according to the word segmentation tool to obtain the word sequence of each subword in each paragraph slice.

In the embodiment of the present application, the word segmentation tool can be NLTK (Natural Language Toolkit) or Spacy (an industrial-grade natural language processing tool), etc. The server uses the word segmentation tool to perform word segmentation on the paragraph slices, and splits the text data corresponding to the paragraph slices into word sequences of words or subwords. In natural language processing, a word sequence is a sequence formed by arranging consecutive words in the text in the order in which they appear in the text after a piece of text data has been processed by word segmentation. Each word occupies a position in the word sequence, and the order of the positions represents the order in which the words appear in the original text.

Step 804 , mapping the word sequence contained in each paragraph slice based on the text vectorization model to obtain an initial second vector representation.

In the embodiment of the present application, the server maps the segmented words or word sequences to a text vectorization model (e.g., a word embedding model) to obtain the initial second vector representation corresponding to each word, i.e., the word vector. If the word sequence is not in the vocabulary of the word embedding model, a special processing method (e.g., random initialization vector) can be used for word vectorization.

Step 806: merge or weighted average the initial second vector representations contained in each paragraph slice to obtain the second vector representation corresponding to each paragraph slice.

In an embodiment of the present application, the server merges or weighted averages the first vector representations of all words in the text data of each paragraph slice to obtain a second vector representation of the complete text data of each paragraph slice. A common method includes obtaining the average vector or weighted average vector of all words, and storing the second vector representation in a vector library to obtain a preset knowledge base. Among them, the server can use a database or a file system to store the second vector representation to ensure that it can be efficiently retrieved and used later.

In this embodiment, a preset knowledge base is constructed by converting text data into a second vector representation and storing the generated second vector representation in a vector library to facilitate subsequent text retrieval, similarity calculation and other tasks, thereby improving the efficiency of document processing and retrieval and ensuring the integrity and accuracy of information.

In a specific embodiment, an example of an intelligent question-answering method is provided, as shown in FIG9 , including steps 901 to 905. Among them:

Step 901, inputting the terminal user's question into the text vectorization model to obtain a first vector representation corresponding to the user's question;

Step 902, inputting the first vector representation into a business vector database;

Step 903, searching based on the similarity between the first vector representation and the business vector in the business vector database to obtain TopK search results as candidate results;

Step 904, integrating the TopK candidate results into a task prompt template in the form of prompt, and inputting it into the LLM question answering model;

Step 905: Return the question and answer result generated by the question and answer model and the search result of the first vector representation in the business vector database to the user terminal.

In one embodiment, an intelligent question-answering system is provided, the system comprising a user end and a server end, wherein:

The client side is used to obtain the question text and feed the question text back to the server side;

The server is used to obtain the question text input by the user; search in the preset knowledge base based on the similarity between the first vector representation corresponding to the question text and each second vector representation existing in the preset knowledge base to obtain multiple candidate results corresponding to the question text; construct a task prompt template based on the first vector representation and each candidate result; analyze and process the correlation between the first vector representation and the candidate results in the task prompt template based on the large language model to obtain the target result corresponding to the question text.

Optionally, after obtaining the target result corresponding to the question text, the server feeds back the target result to the terminal.

It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

Based on the same inventive concept, the embodiment of the present application also provides an intelligent question-answering device for implementing the intelligent question-answering method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more intelligent question-answering device embodiments provided below can refer to the limitations of the intelligent question-answering method above, and will not be repeated here.

In an exemplary embodiment, as shown in FIG10 , an intelligent question-answering device is provided, comprising: a first acquisition module 1001, a retrieval module 1002, a first construction module 1003 and an analysis module 1004, wherein:

The first acquisition module 1001 is used to acquire the question text input by the user;

A retrieval module 1002 is used to search in a preset knowledge base based on the similarity between the first vector representation corresponding to the question text and each second vector representation existing in the preset knowledge base, and obtain multiple candidate results corresponding to the question text;

A first construction module 1003, configured to construct a task prompt template according to the first vector representation and each candidate result;

The analysis module 1004 is used to analyze and process the association between the first vector representation and the candidate results in the task prompt template based on the large language model to obtain the target result corresponding to the question text.

In one embodiment, the retrieval module 1002 is specifically used to perform vectorization processing on the question text according to the text vectorization model to obtain a first vector representation corresponding to the question text;

Calculating the similarity between the first vector representation and each second vector representation in the preset knowledge base based on a preset vector similarity algorithm;

Each second vector representation having a similarity greater than a preset threshold is regarded as a candidate result corresponding to the question text.

In one embodiment, the analysis module 1004 is specifically used to perform semantic understanding on the first vector representation and the candidate results in the task prompt template based on the large language model, and obtain the first semantic information corresponding to the question text and the second semantic information corresponding to the candidate results;

Performing a correlation analysis on the first semantic information and the second semantic information based on the large language model to obtain semantic correlation information;

The candidate results are screened based on the semantic association information to obtain the target candidate results, and the target results are generated based on the target candidate results.

In one embodiment, the apparatus 1000 further includes:

A first determination module is used to determine the text block mark corresponding to the target candidate result in a preset knowledge base;

A second determination module is used to determine a target file path or link of a target candidate result according to a correspondence between a text block mark and a file path;

The feedback module is used to feedback the target file path or link.

In one embodiment, the apparatus 1000 further includes:

A second acquisition module is used to acquire multimodal knowledge data;

The slicing module is used to slice the multimodal knowledge data to obtain multiple paragraph slices;

A vectorization processing module is used to perform data processing on each paragraph slice based on a text vectorization model to obtain a second vector representation corresponding to each paragraph slice, and determine a text block mark corresponding to each second vector representation;

The second building module is used to build a preset knowledge base according to the second vector representation and the text block mark corresponding to the second vector representation.

In one of the embodiments, the slicing module is specifically used to convert the multimodal knowledge data into text data based on the multimodal model;

The text data is segmented into paragraphs based on the paragraph marks to obtain paragraph slices and text block tags corresponding to the paragraph slices.

In one embodiment, the vectorization processing module is specifically used to perform word segmentation processing on each paragraph slice according to a word segmentation tool to obtain a word sequence of each subword in each paragraph slice;

Based on the text vectorization model, the word sequence contained in each paragraph slice is mapped to obtain an initial second vector representation;

The initial second vector representations contained in each paragraph slice are merged or weighted averaged to obtain the second vector representation corresponding to each paragraph slice.

Each module in the above intelligent question-answering device can be implemented in whole or in part by software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to each module.

In an exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be shown in FIG11. The computer device includes a processor, a memory, an input/output interface (Input/Output, referred to as I/O) and a communication interface. The processor, the memory and the input/output interface are connected via a system bus, and the communication interface is connected to the system bus via the input/output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the second vector representation present in the preset knowledge base. The input/output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, an intelligent question-answering method is implemented.

Those skilled in the art will understand that the structure shown in FIG. 11 is merely a block diagram of a partial structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different arrangement of components.

In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

Get the question text entered by the user;

Based on the similarity between the first vector representation corresponding to the question text and each second vector representation existing in the preset knowledge base, a search is performed in the preset knowledge base to obtain multiple candidate results corresponding to the question text;

constructing a task prompt template according to the first vector representation and each of the candidate results;

Based on the large language model, the correlation between the first vector representation and the candidate result in the task prompt template is analyzed and processed to obtain a target result corresponding to the question text.

In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

Performing vectorization processing on the question text according to a text vectorization model to obtain a first vector representation corresponding to the question text;

Calculating the similarity between the first vector representation and each second vector representation in a preset knowledge base based on a preset vector similarity algorithm;

Each second vector representation whose similarity is greater than a preset threshold is taken as a candidate result corresponding to the question text.

In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

Based on the large language model, semantic understanding is performed on the first vector representation and the candidate result in the task prompt template to obtain first semantic information corresponding to the question text and second semantic information corresponding to the candidate result;

Performing a correlation analysis on the first semantic information and the second semantic information based on the large language model to obtain semantic correlation information;

The candidate results are screened based on the semantic association information to obtain target candidate results, and a target result is generated based on the target candidate results.

In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

Determine the text block tag corresponding to the target candidate result in the preset knowledge base;

Determine the target file path or link of the target candidate result according to the correspondence between the text block mark and the file path;

The target file path or the link is fed back.

In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

Acquire multimodal knowledge data;

Slicing the multimodal knowledge data to obtain a plurality of paragraph slices;

Performing data processing on each of the paragraph slices based on a text vectorization model to obtain a second vector representation corresponding to each of the paragraph slices, and determining a text block tag corresponding to each of the second vector representations;

A preset knowledge base is constructed according to the second vector representation and the text block mark corresponding to the second vector representation.

In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

Converting the multimodal knowledge data into text data based on a multimodal model;

The text data is segmented into paragraphs based on paragraph marks to obtain paragraph slices and text block marks corresponding to the paragraph slices.

In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

For each of the paragraph slices, segment the paragraph slices according to a segmentation tool to obtain a word sequence of each subword in each of the paragraph slices;

Mapping the word sequence contained in each paragraph slice based on a text vectorization model to obtain an initial second vector representation;

The initial second vector representations contained in each of the paragraph slices are merged or weighted averaged to obtain the second vector representation corresponding to each of the paragraph slices.

In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above method embodiments are implemented.

In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic unit, a data processing logic unit based on quantum computing, an artificial intelligence (AI) processor, etc., but are not limited to this.

The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims (12)

1. An intelligent question-answering method, characterized in that the method comprises: Get the question text entered by the user; Based on the similarity between the first vector representation corresponding to the question text and each second vector representation existing in the preset knowledge base, a search is performed in the preset knowledge base to obtain multiple candidate results corresponding to the question text; constructing a task prompt template according to the first vector representation and each of the candidate results; Based on the large language model, the correlation between the first vector representation and the candidate result in the task prompt template is analyzed and processed to obtain a target result corresponding to the question text. 2. The method according to claim 1, characterized in that the step of searching the preset knowledge base based on the similarity between the first vector representation corresponding to the question text and each second vector representation existing in the preset knowledge base to obtain multiple candidate results corresponding to the question text comprises: Performing vectorization processing on the question text according to a text vectorization model to obtain a first vector representation corresponding to the question text; Calculating the similarity between the first vector representation and each second vector representation in a preset knowledge base based on a preset vector similarity algorithm; Each second vector representation whose similarity is greater than a preset threshold is taken as a candidate result corresponding to the question text. 3. The method according to claim 1, characterized in that the step of analyzing and processing the correlation between the first vector representation in the task prompt template and the candidate result based on the large language model to obtain the target result corresponding to the question text comprises: Based on the large language model, semantic understanding is performed on the first vector representation and the candidate result in the task prompt template to obtain first semantic information corresponding to the question text and second semantic information corresponding to the candidate result; Performing a correlation analysis on the first semantic information and the second semantic information based on the large language model to obtain semantic correlation information; The candidate results are screened based on the semantic association information to obtain target candidate results, and a target result is generated based on the target candidate results. 4. The method according to claim 3, characterized in that after screening the candidate results based on the semantic association information to obtain a target candidate result and generating a target result based on the target candidate result, the method further comprises: Determine the text block tag corresponding to the target candidate result in the preset knowledge base; Determine the target file path or link of the target candidate result according to the correspondence between the text block mark and the file path; The target file path or the link is fed back. 5. The method according to claim 1, characterized in that before obtaining the question text input by the user, the method further comprises: Acquire multimodal knowledge data; Slicing the multimodal knowledge data to obtain a plurality of paragraph slices; Performing data processing on each of the paragraph slices based on a text vectorization model to obtain a second vector representation corresponding to each of the paragraph slices, and determining a text block mark corresponding to each of the second vector representations; A preset knowledge base is constructed according to the second vector representation and the text block mark corresponding to the second vector representation. 6. The method according to claim 5, characterized in that the slicing process of the multimodal knowledge data to obtain a plurality of paragraph slices comprises: Converting the multimodal knowledge data into text data based on a multimodal model; The text data is segmented into paragraphs based on paragraph marks to obtain paragraph slices and text block marks corresponding to the paragraph slices. 7. The method according to claim 5, characterized in that the data processing of each paragraph slice based on the text vectorization model to obtain a second vector representation corresponding to each paragraph slice, and determining a text block mark corresponding to each second vector representation, comprises: For each of the paragraph slices, segment the paragraph slices according to a segmentation tool to obtain a word sequence of each subword in each of the paragraph slices; Mapping the word sequence contained in each paragraph slice based on a text vectorization model to obtain an initial second vector representation; The initial second vector representations contained in each of the paragraph slices are merged or weighted averaged to obtain the second vector representation corresponding to each of the paragraph slices. 8. An intelligent question-answering system, characterized in that the system comprises: The client terminal is used to obtain the question text and feed the question text back to the server terminal; The server is used to obtain a question text input by a user; search the preset knowledge base based on the similarity between the first vector representation corresponding to the question text and each second vector representation existing in the preset knowledge base to obtain multiple candidate results corresponding to the question text; construct a task prompt template based on the first vector representation and each candidate result; analyze and process the association relationship between the first vector representation and the candidate results in the task prompt template based on a large language model to obtain a target result corresponding to the question text. 9. An intelligent question-answering device, characterized in that the device comprises: The first acquisition module is used to obtain the question text input by the user; A retrieval module, configured to search the preset knowledge base based on the similarity between the first vector representation corresponding to the question text and each second vector representation existing in the preset knowledge base, to obtain a plurality of candidate results corresponding to the question text; A first construction module, configured to construct a task prompt template according to the first vector representation and each of the candidate results; The analysis module is used to analyze and process the association between the first vector representation and the candidate result in the task prompt template based on the large language model to obtain the target result corresponding to the question text. 10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein the processor implements the steps of the method according to any one of claims 1 to 7 when executing the computer program. 11. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented. 12. A computer program product, comprising a computer program, characterized in that when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.