Answer generation method, knowledge question-answering system and electronic equipment

Abstract

The invention provides an answer generation method, a knowledge question-answering system and electronic equipment, and belongs to the technical field of artificial intelligence. According to the method, through multi-dimensional feature extraction, features in different dimensions can be extracted for subsequent complexity classification judgment. The complexity of subsequent processing can be reduced by screening out the core feature which can best reflect the complexity of the problem from the multi-dimensional feature set. And aggregating each core feature in the core feature set by using an adaptive aggregation strategy to obtain an aggregated feature vector, inputting the aggregated feature vector into a hierarchical complexity shunt, and for a query problem with a lower complexity level, directly outputting the complexity level by a lightweight classification layer. And for the query problem with a relatively high complexity level, the complexity level is output by the deep classification layer, so that the accuracy and efficiency of complexity classification can be ensured. And finally, according to the complexity level corresponding to the query question, a corresponding retrieval and answer generation strategy is selected, so that the flexibility and efficiency of answer generation can be improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to an answer generation method, a knowledge question answering system and an electronic device. Background Art

[0002] In the prior art, retrieval-augmented generation (RAG) models usually use fixed strategies for retrieval and generation, which cannot be dynamically adjusted according to the complexity of the query. For simple queries, using complex multi-step methods will lead to unnecessary computational overhead, while for complex queries, using simple methods may not provide accurate answers. Summary of the invention

[0003] The main purpose of the embodiments of the present application is to propose an answer generation method, a knowledge question and answer system and an electronic device, which aims to automatically select appropriate retrieval and answer generation strategies to generate answers according to the complexity of the query question, thereby improving the flexibility and efficiency of answer generation.

[0004] To achieve the above objective, a first aspect of an embodiment of the present application proposes an answer generation method, the method comprising:

[0005] Obtaining a user's query question and a corresponding historical interaction record, and performing multidimensional feature extraction on the query question and the corresponding historical interaction record to obtain a multidimensional feature set;

[0006] Analyze each feature in the multidimensional feature set, identify core features, and construct a core feature set, wherein the core features are the features that best reflect the complexity of the problem;

[0007] Using an adaptive aggregation strategy, aggregating each core feature in the core feature set to obtain an aggregated feature vector;

[0008] Inputting the aggregated feature vector into a hierarchical complexity splitter for processing to obtain a complexity level corresponding to the query question, wherein the hierarchical complexity splitter includes a lightweight classification layer and a deep classification layer, and when the lightweight classification layer cannot determine the complexity level of the query question, the deep classification layer uses a multi-head complexity decoder to determine the complexity level corresponding to the query question;

[0009] According to the complexity level corresponding to the query question, a corresponding retrieval and answer generation strategy is selected to generate an answer to the query question.

[0010] In one embodiment of the present application, multidimensional feature extraction is performed on the query question and the corresponding historical interaction record to obtain a multidimensional feature set, including:

[0011] The query question is integrated with the historical interaction record through a bidirectional attention mechanism to generate context information;

[0012] Extracting different types of features according to the context information to obtain a candidate multidimensional feature set, wherein the candidate multidimensional feature set includes grammatical features, semantic features, and retrieval information features;

[0013] By using a plurality of complexity-aware attention heads, attention weights at different complexity levels are calculated respectively, and the weight of each feature in the candidate multidimensional feature set is adjusted according to the attention weights;

[0014] The candidate multidimensional feature set is subjected to feature transformation, filtering, reconstruction and completion processing to obtain a multidimensional feature set.

[0015] In one embodiment of the present application, analyzing each feature in the multidimensional feature set, identifying core features, and constructing a core feature set includes:

[0016] Constructing a complexity anchor point, wherein the complexity anchor point is a benchmark for measuring complexity;

[0017] Comparing each feature in the multidimensional feature set with the complexity anchor point to screen out core features;

[0018] The filtered core features are collected to obtain a core feature set.

[0019] In one embodiment of the present application, the use of the adaptive aggregation strategy to aggregate the core features in the core feature set to obtain an aggregated feature vector includes:

[0020] Determine the weight corresponding to each core feature according to the relative importance of each core feature in the core feature set and the matching degree between the core feature and the complexity anchor point;

[0021] According to the weights corresponding to each core feature, an adaptive aggregation strategy is used to aggregate each core feature to obtain an aggregated feature vector.

[0022] In one embodiment of the present application, the step of inputting the aggregated feature vector into a hierarchical complexity splitter for processing to obtain a complexity level corresponding to the query question includes:

[0023] Inputting the aggregated feature vector into a hierarchical complexity splitter, obtaining the aggregated feature vector by the lightweight classification layer, and determining the complexity of the query question;

[0024] If the lightweight classification layer cannot determine the complexity of the query question, the deep classification layer uses a multi-head complexity decoder to determine the complexity level corresponding to the query question and output it;

[0025] The deep classification layer is used to predict the number of sub-questions, domain knowledge call tags, and knowledge verification tags.

[0026] In one embodiment of the present application, the complexity level includes a first level, a second level, a third level and a fourth level, wherein the complexity represented by the first level, the second level, the third level and the fourth level increases in sequence, and the selecting a corresponding retrieval and answer generation strategy according to the complexity level corresponding to the query question to generate an answer to the query question includes:

[0027] If the complexity level corresponding to the query question is the first level, the internal knowledge of the large language model is used to generate a corresponding answer;

[0028] If the complexity corresponding to the query question is the second level, documents related to the query question are retrieved from an external knowledge base, and the query question and the retrieved documents are input into a large language model to generate an answer to the query question;

[0029] If the complexity corresponding to the query question is the third level, the query question is decomposed into a plurality of sub-questions, and each of the sub-questions is searched and a sub-answer of each sub-question is generated respectively, so as to generate an answer to the query question according to all the sub-answers;

[0030] If the complexity corresponding to the query question is the fourth level, multi-source information related to the query question is retrieved, and an answer to the query question is generated according to the multi-source information.

[0031] In one embodiment of the present application, after selecting a corresponding retrieval and answer generation strategy according to the complexity level corresponding to the query question to generate an answer to the query question, the method further includes:

[0032] Obtaining user feedback, wherein the user feedback includes user feedback on the generated answer to the query question;

[0033] Retrieval and answer generation strategies are optimized based on the user feedback.

[0034] To achieve the above-mentioned purpose, the second aspect of the embodiment of the present application proposes a knowledge question answering system, including a context-aware multidimensional feature extraction model, a multi-level complexity-aware fusion network and an improved retrieval-based enhanced generation model, wherein:

[0035] The context-aware multidimensional feature extraction model is used to perform the following steps:

[0036] Obtaining a user's query question and a corresponding historical interaction record, and performing multidimensional feature extraction on the query question and the corresponding historical interaction record to obtain a multidimensional feature set;

[0037] The multi-level complexity-aware fusion network is used to perform the following steps:

[0038] Analyze each feature in the multidimensional feature set, identify core features, and construct a core feature set, wherein the core features are the features that best reflect the complexity of the problem;

[0039] Using an adaptive aggregation strategy, aggregating each core feature in the core feature set to obtain an aggregated feature vector;

[0040] Inputting the aggregated feature vector into a hierarchical complexity splitter for processing to obtain a complexity level corresponding to the query question, wherein the hierarchical complexity splitter includes a lightweight classification layer and a deep classification layer, and when the lightweight classification layer cannot determine the complexity level of the query question, the deep classification layer uses a multi-head complexity decoder to determine the complexity level corresponding to the query question;

[0041] The improved search-based enhanced generation model is used to perform the following steps:

[0042] According to the complexity level corresponding to the query question, a corresponding retrieval and answer generation strategy is selected to generate an answer to the query question.

[0043] In one embodiment of the present application, the context-aware multidimensional feature extraction model includes:

[0044] A dynamic context fusion module, used to fuse the query question with the historical interaction record through a bidirectional attention mechanism to generate context information;

[0045] A multi-dimensional feature adaptive extraction module, used to extract different types of features according to the context information to obtain a candidate multi-dimensional feature set, wherein the multi-dimensional features include grammatical features, semantic features and retrieval information features;

[0046] A hierarchical complexity attention mechanism, for calculating attention weights at different complexity levels through multiple complexity-aware attention heads, and adjusting the weights of each feature in the candidate multidimensional feature set according to the attention weights;

[0047] The adaptive feature enhancement layer is used to perform feature transformation, filtering, reconstruction and completion processing on the candidate multidimensional feature set to obtain a multidimensional feature set.

[0048] To achieve the above-mentioned purpose, the third aspect of an embodiment of the present application proposes an electronic device, which includes a memory and a processor, the memory stores a computer program, and the processor implements the method described in any embodiment of the present application when executing the computer program.

[0049] In the technical solution provided by an embodiment of the present application, multidimensional feature extraction is first performed on the query question and the corresponding historical interaction record to obtain a multidimensional feature set. By combining the corresponding historical interaction record for feature extraction, the accuracy of understanding the query question can be improved. Through multidimensional feature extraction, features under different dimensions can be extracted for subsequent complexity classification judgment, laying the foundation for the accurate classification of subsequent complexity. Then, each feature in the multidimensional feature set is analyzed, the core features are identified, and a core feature set is constructed, wherein the core features are the features that best reflect the complexity of the question. That is, by selecting the core features that best reflect the complexity of the question from the multidimensional feature set, the accuracy of complexity classification can be improved, the complexity of subsequent processing can be reduced, and the efficiency of subsequent processing can be improved. Then, an adaptive aggregation strategy is used to aggregate the core features in the core feature set to obtain an aggregated feature vector. Among them, the obtained aggregated feature vector not only contains the semantic and grammatical features of the query question, but also implies the prejudgment and preference of the complexity of the query question, which is convenient for subsequent complexity classification. Then the aggregated feature vector is input into the hierarchical complexity splitter for processing, and the complexity level corresponding to the query question can be obtained. Among them, the hierarchical complexity splitter includes a lightweight classification layer and a deep classification layer. When the lightweight classification layer cannot determine the complexity level of the query question, the deep classification layer uses a multi-head complexity decoder to determine the complexity level corresponding to the query question. That is, for query questions with a lower complexity level (i.e., simpler questions), the lightweight classification layer can directly output the complexity level, while for query questions with a higher complexity level (i.e., more complex questions), the deep classification layer can output the complexity level, which can ensure the accuracy of complexity classification while improving the efficiency of complexity classification. Finally, according to the complexity level corresponding to the query question, the corresponding retrieval and answer generation strategy is selected to generate the answer to the query question, which can improve the flexibility and efficiency of answer generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flow chart of an answer generation method provided by an embodiment of the present application;

[0051] Figure 2 A flowchart of the steps of extracting multidimensional features from query questions and corresponding historical interaction records to obtain a multidimensional feature set provided by an embodiment of the present application;

[0052] Figure 3It is a flowchart of the steps of analyzing each feature in a multidimensional feature set, identifying core features, and constructing a core feature set provided by an embodiment of the present application;

[0053] Figure 4 This is a flowchart of the steps of using an adaptive aggregation strategy to aggregate each core feature in a core feature set to obtain an aggregated feature vector, provided by an embodiment of the present application;

[0054] Figure 5 A flowchart of the steps of inputting an aggregated feature vector into a hierarchical complexity splitter for processing to obtain a complexity level corresponding to a query question provided by an embodiment of the present application;

[0055] Figure 6 The embodiment of the present application provides a flowchart of steps for selecting a corresponding retrieval and answer generation strategy according to the complexity level corresponding to the query question to generate an answer to the query question;

[0056] Figure 7 is an architectural block diagram of a knowledge question answering system provided by an embodiment of the present application;

[0057] Figure 8 It is an architectural block diagram of a context-aware multidimensional feature extraction model provided in an embodiment of the present application;

[0058] Fig. 9 It is an architectural block diagram of a multi-level complexity-aware fusion network provided in an embodiment of the present application;

[0059] Fig.10 It is a schematic diagram of the hardware structure of an electronic device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0060] 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.

[0061] It should be noted that, although the functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first", "second", etc. in the specification, claims and the above drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0062] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0063] Terminology explanation:

[0064] The improved Retrieval-Augmented Generation (RAG) model is a natural language processing technology that combines information retrieval and generation models. It uses the retrieval system to find relevant information from a large number of documents, and then passes this information to the generation model to generate more accurate and detailed text answers. Retrieval-Augmented Generation technology combines information retrieval and generation models, and uses the retrieved relevant information to enhance the answering ability of the generation model, thereby improving the overall performance of the natural language processing system.

[0065] Traditional RAG models usually use fixed strategies when processing queries of different complexity and cannot be adjusted dynamically. In fact, users' queries vary in complexity, some queries are very simple, and some require complex multi-step reasoning. Models using fixed strategies have difficulty performing well in queries of different complexity. For example, for simple queries, using a multi-step approach will result in unnecessary computational overhead, while for complex queries, using a simple approach may not provide accurate answers.

[0066] Based on this, an embodiment of the present application provides an answer generation method that can automatically select appropriate retrieval and answer generation strategies to generate answers based on the complexity of the query question, thereby improving the flexibility and efficiency of answer generation.

[0067] Reference Figure 1 , Figure 1 It is a flowchart of the answer generation method provided in one embodiment of the present application, including but not limited to steps S110 to S150.

[0068] Step S110, obtaining the user's query question and the corresponding historical interaction record, and performing multidimensional feature extraction on the query question and the corresponding historical interaction record to obtain a multidimensional feature set.

[0069] In the embodiment of the present application, when a user makes a query, he or she will first enter a query question. For more complex questions, there may also be interaction with the machine side to form a historical interaction record. The historical interaction record is crucial to the understanding of the query question. For more complex questions, the corresponding historical interaction record is needed to truly understand the meaning of the query question. While obtaining the user's query question, the embodiment of the present application also obtains the corresponding historical interaction record to improve the accuracy of understanding the query question.

[0070] After obtaining the user's query questions and historical interaction records, multi-dimensional feature extraction can be further performed on the query questions and the corresponding historical interaction records to obtain a multi-dimensional feature set. For example, the grammatical features and semantic features corresponding to the query questions can be extracted for subsequent complexity analysis.

[0071] Reference Figure 2 , Figure 2 It is a flowchart of the steps of extracting multidimensional features from query questions and corresponding historical interaction records to obtain a multidimensional feature set, including but not limited to steps S210 to S240, provided by an embodiment of the present application.

[0072] Step S210, using a bidirectional attention mechanism, the query question is integrated with the historical interaction records to generate context information;

[0073] Step S220, extracting different types of features according to the context information to obtain a candidate multidimensional feature set, wherein the candidate multidimensional feature set includes grammatical features, semantic features, and retrieval information features;

[0074] Step S230, calculating the attention weights at different complexity levels through multiple complexity-aware attention heads, and adjusting the weights of each feature in the candidate multidimensional feature set according to the attention weights;

[0075] Step S240, performing feature transformation, filtering, reconstruction and completion processing on the candidate multidimensional feature set to obtain a multidimensional feature set.

[0076] In an embodiment of the present application, the input query question can be preprocessed first, noise characters can be removed, and the core semantics can be retained. Then, through a two-way attention mechanism, the query question is fused with the historical interaction record to generate context information. Specifically, for the historical interaction record, the time-aware memory unit can be used to store and update in chronological order. By introducing a time encoding mechanism, the temporal dynamics of historical interactions can be captured, and the priority of the latest interactive information can be ensured. Among them, the historical interaction record not only includes historical text records, but also user behavior data (such as clicks, dwell time) and emotional feedback (emoticons, emoticons), etc. The historical interaction record is converted into a vector representation through an embedding layer, and then through a two-way attention mechanism, the current query question is fused with the historical interaction record represented by the vector to generate context information. Through the two-way attention mechanism, features can be captured from two directions (such as the forward and backward directions of the time series) at the same time, and the global and local features of the sequence can also be taken into account at the same time, which can improve the accuracy of semantic understanding.

[0077] Next, according to the context information, different types of features are extracted to obtain a candidate multidimensional feature set, wherein the candidate multidimensional feature set includes grammatical features, semantic features, and retrieval information features. Specifically, a special grammatical parsing network can be used to extract grammatical features such as the number of clauses and nesting levels in the context. The keyword density in the query question can also be identified through a keyword recognition network. An adaptive semantic encoder can also be used to generate the semantic embedding of the query question, and the similarity is calculated by comparing it with the standard question template to determine the semantic features of the query question. A reference resolution network can also be applied to detect fuzzy references and multi-round dialogue requirements in the query question to determine the retrieval information features corresponding to the query question. A multi-task learning framework that shares the underlying representation can be used to design special processing channels for different types of features to ensure the optimal extraction and expression of various features, obtain corresponding features under different dimensions, and construct a candidate multidimensional feature set.

[0078] Taking into account the different importance of different features to complexity classification, the embodiment of the present application uses multiple complexity-aware attention heads to calculate the attention weights at different complexity levels, and adjusts the weights of each feature in the candidate multidimensional feature set according to the attention weights. Specifically, a multi-level attention mechanism can be designed to allocate attention weights in a hierarchical manner according to the complexity requirements of the features to ensure the priority of important features. Among them, each attention head can focus on features at different complexity levels to enhance the sensitivity and responsiveness of the model to complexity. The weight allocation of attention heads at each layer can also be dynamically adjusted according to the actual complexity of the query problem to achieve dynamic priority sorting of features. Through multiple complexity-aware attention heads, the attention weights at different complexity levels are calculated respectively, and then the weights of each feature in the candidate multidimensional feature set are adjusted according to the attention weights, so that the weights corresponding to each feature in the candidate multidimensional feature set can be dynamically adjusted, which can ensure that the feature selection is highly matched with the complexity requirements.

[0079] Considering that some features may be missing or incomplete in the candidate multidimensional feature set, and some features may be irrelevant to the complexity classification, the embodiment of the present application also performs feature transformation, filtering, reconstruction and completion processing on the candidate multidimensional feature set to obtain a multidimensional feature set. Specifically, the generative adversarial network (GAN) technology can be used to reconstruct and complete the missing or incomplete features to improve the integrity of the features. A noise filtering mechanism can be applied to remove the features in the candidate multidimensional feature set that are irrelevant to the complexity classification, and the features in the candidate multidimensional feature set can also be subjected to multi-layer nonlinear transformations, thereby enhancing the key features in the candidate multidimensional feature set and improving the richness and discrimination ability of feature expression.

[0080] The embodiment of the present application uses multi-dimensional feature extraction to extract features in different dimensions for subsequent complexity classification judgment, laying the foundation for subsequent accurate complexity classification.

[0081] Step S120 , analyzing each feature in the multidimensional feature set, identifying core features, and constructing a core feature set, wherein the core features are the features that best reflect the complexity of the problem.

[0082] In the embodiment of the present application, after obtaining a multidimensional feature set based on the context information of the constructed query question, it is considered that the multidimensional feature set contains different types of features, and different types of features contribute differently to the complexity classification. Therefore, in order to improve the accuracy and efficiency of the complexity level classification, it is necessary to analyze each feature in the multidimensional feature set, identify the core features that best reflect the complexity of the question, and construct a core feature set.

[0083] Reference Figure 3 , Figure 3 It is a flowchart of the steps of analyzing each feature in a multidimensional feature set, identifying core features, and constructing a core feature set provided by an embodiment of the present application, including but not limited to steps S310 to S330.

[0084] Step S310, constructing a complexity anchor point, which is a benchmark for measuring complexity;

[0085] Step S320, comparing each feature in the multidimensional feature set with the complexity anchor point to screen out the core features;

[0086] Step S330, gathering the screened core features to obtain a core feature set.

[0087] In the embodiment of the present application, in order to select the core features that best reflect the complexity of the problem from the multidimensional feature set and avoid feature redundancy, it is necessary to first construct a complexity anchor point, which is a benchmark for measuring complexity. Specifically, a set of "complexity anchor vectors" (Complexity Anchors) can be learned from the training process. These anchor points are trainable embeddings used to represent the concept space of different complexity dimensions (such as simple, medium, complex, and knowledge verification). Unlike the traditional attention mechanism that only allocates weights between features, these complexity anchor points can provide the model with a benchmark representation of complexity, so that subsequent feature aggregation is not only a choice of "feature importance", but also an optimization of "feature and complexity concept matching". Then, by comparing each feature in the multidimensional feature set with the complexity anchor point, the core features can be screened out. The comparison process is similar to repositioning the features in the "complexity benchmark coordinate system", thereby automatically highlighting those features that are critical to complexity classification. In this way, subtle language structures or low-frequency keywords that are discriminating for complexity classification can be identified, and irrelevant or redundant information can be removed. Then all the filtered core features are combined to obtain the core feature set.

[0088] In an embodiment of the present application, by selecting the core features that best reflect the complexity of the problem from a multidimensional feature set to construct a core feature set, redundant features can be further removed, the accuracy of complexity classification can be improved, and the complexity of subsequent processing can be reduced, thereby improving the efficiency of subsequent processing.

[0089] Step S130 , using an adaptive aggregation strategy, aggregates the core features in the core feature set to obtain an aggregated feature vector.

[0090] In the embodiment of the present application, after the core feature set is constructed, an adaptive aggregation strategy can be used to combine the core feature set into an optimized aggregate feature vector. The aggregate feature vector not only contains the semantic and grammatical features of the query question, but also implies the prediction and preference of the complexity of the query question, which is convenient for subsequent complexity classification.

[0091] Reference Figure 4 , Figure 4 It is a flowchart of the steps of using an adaptive aggregation strategy to aggregate the core features in the core feature set to obtain an aggregated feature vector, including but not limited to steps S410 to S420, provided in an embodiment of the present application.

[0092] Step S410, determining the weight corresponding to each core feature according to the relative importance of each core feature in the core feature set and the matching degree between the core feature and the complexity anchor point;

[0093] Step S420 , according to the weights corresponding to the core features, an adaptive aggregation strategy is used to aggregate the core features to obtain an aggregated feature vector.

[0094] In the embodiment of the present application, considering that the matching degree between different core features and complexity anchor points is different in the core feature set, the relative importance between different core features is also different. Therefore, the weight corresponding to each core feature can be determined according to the relative importance between each core feature in the core feature set and the matching degree between the core features and the complexity anchor point, such as assigning a higher weight to the core features with a higher matching degree between the core features and the complexity anchor point and a higher relative importance between the core features, and assigning a lower weight to the core features with a lower matching degree between the core features and the complexity anchor point and a lower relative importance between the core features. Then, according to the weight corresponding to each core feature, an adaptive aggregation strategy is used to aggregate each core feature to obtain an aggregated feature vector. For example, the feature integration strategy can be dynamically adjusted through a complexity-aware gating unit. The complexity-aware gating unit can be guided by the complexity anchor point, and decides in real time which core features should be strengthened and which core features should be suppressed according to the complexity tendency, so that the finally generated aggregated feature vector is more adaptable and discriminative for subsequent complexity classification.

[0095] Step S140, input the aggregated feature vector into a hierarchical complexity splitter for processing to obtain a complexity level corresponding to the query question, wherein the hierarchical complexity splitter includes a lightweight classification layer and a deep classification layer. When the lightweight classification layer cannot determine the complexity level of the query question, the deep classification layer uses a multi-head complexity decoder to determine the complexity level corresponding to the query question.

[0096] In an embodiment of the present application, after obtaining the aggregated feature vector, the aggregated feature vector can be directly input into a hierarchical complexity splitter for processing. The hierarchical complexity splitter includes a lightweight classification layer and a deep classification layer. For query problems with a lower complexity level (i.e., simpler problems), the lightweight classification layer can directly output the complexity level, while for query problems with a higher complexity level (i.e., more complex problems), the deep classification layer outputs the complexity level, which can ensure the accuracy of complexity classification while improving the efficiency of complexity classification.

[0097] Reference Figure 5 , Figure 5 It is a flowchart of the steps of inputting the aggregated feature vector into the hierarchical complexity splitter for processing to obtain the complexity level corresponding to the query question, provided by an embodiment of the present application, including but not limited to steps S510 to S530.

[0098] Step S510, inputting the aggregated feature vector into the hierarchical complexity splitter, obtaining the aggregated feature vector by the lightweight classification layer, and determining the complexity of the query question;

[0099] Step S520: If the lightweight classification layer cannot determine the complexity of the query question, the deep classification layer uses a multi-head complexity decoder to determine the complexity level corresponding to the query question and output it;

[0100] Step S530, using the deep classification layer to predict the number of sub-questions, domain knowledge call marks and knowledge verification marks.

[0101] In an embodiment of the present application, the aggregated feature vector can be first input into the lightweight classification layer in the hierarchical complexity splitter. The lightweight classification layer is like an experienced receptionist who can quickly make decisions on simple questions and directly output the results. If the lightweight classification layer determines that the query question has a high complexity uncertainty, the aggregated feature vector can be delivered to the deep classification layer, which uses a multi-head complexity decoder to determine the complexity level corresponding to the query question and output it. Among them, the multi-head complexity decoder can be a group of parallel sub-network branches, each branch can be jointly trained to predict multi-dimensional outputs, that is, in addition to outputting the complexity level of the query question, it can also output the structural information corresponding to the query question, the structural information includes the number of sub-questions, domain knowledge call tags and knowledge verification tags, etc. Among them, the structural information can be used to verify the subsequent generation of answers.

[0102] In an embodiment of the present application, by inputting the aggregated feature vector into a hierarchical complexity splitter, the complexity level can be directly output by the lightweight classification layer for query problems with a lower complexity level (i.e., simpler problems), while the complexity level is output by the deep classification layer for query problems with a higher complexity level (i.e., more complex problems). This can ensure the accuracy of complexity classification while improving the efficiency of complexity classification.

[0103] Step S150, according to the complexity level corresponding to the query question, select the corresponding retrieval and answer generation strategy to generate an answer to the query question.

[0104] In the embodiment of the present application, after determining the complexity level of the query question, a corresponding retrieval and answer generation strategy can be selected according to the complexity level of the query question to generate an answer to the query question. For example, for simple questions, the answer can be generated directly; for complex questions, the accuracy of the answer can be ensured through a multi-step retrieval and generation process.

[0105] Reference Figure 6 , Figure 6It is a flowchart of steps provided by an embodiment of the present application for selecting corresponding retrieval and answer generation strategies according to the complexity level corresponding to the query question to generate an answer to the query question, including but not limited to steps S610 to S640.

[0106] Step S610, if the complexity level corresponding to the query question is the first level, the internal knowledge of the large language model is used to generate a corresponding answer;

[0107] Step S620: if the complexity corresponding to the query question is the second level, documents related to the query question are retrieved from the external knowledge base, and the query question and the retrieved documents are input into the large language model to generate an answer to the query question;

[0108] Step S630: if the complexity corresponding to the query question is the third level, the query question is decomposed into multiple sub-questions, and each sub-question is searched and a sub-answer of each sub-question is generated respectively, so as to generate an answer to the query question according to all the sub-answers;

[0109] Step S640: If the complexity corresponding to the query question is the fourth level, then multi-source information related to the query question is retrieved, and an answer to the query question is generated based on the multi-source information.

[0110] In the embodiment of the present application, the complexity level may include the first level, the second level, the third level and the fourth level, wherein the complexity represented by the first level, the second level, the third level and the fourth level increases in sequence. When the complexity level of the query question is the first level, the internal knowledge of the large language model (LLM) can be directly used to generate the answer. This method is not only efficient, but also can quickly respond to the user's simple query. When the complexity level of the query question is the second level, a single-step retrieval and generation strategy is adopted. That is, the retrieval module is used to retrieve documents related to the query question from the external knowledge base, and then the query question and the retrieved documents are input into the LLM model to generate the answer. When the complexity level of the query question is the third level, a multi-step retrieval and generation strategy is adopted to ensure the accuracy of the answer through multiple iterations. Specifically, it includes preliminary retrieval of relevant documents and judgment of their relevance to the question. If the relevance of the documents initially retrieved is low, the query question is decomposed into multiple sub-questions using the question decomposition module, and each sub-question is retrieved and the sub-answers corresponding to each sub-question are generated respectively, and finally all the sub-answers are integrated to generate the answer corresponding to the query question. When the complexity level of the query is the fourth level, multi-source information related to the query is retrieved and verified using a hybrid expert model. Then, through the reward mechanism of the reinforcement learning model, the system can verify the accuracy of this information to ensure that the knowledge in the knowledge base is reliable and up-to-date. Finally, the system generates the answer to the query based on the multi-source information. In this way, according to the complexity level corresponding to the query, the corresponding retrieval and answer generation strategy is selected to generate the answer to the query, which can improve the flexibility and efficiency of answer generation.

[0111] In one embodiment of the present application, after selecting the corresponding retrieval and answer generation strategy according to the complexity level corresponding to the query question to generate the answer to the query question, user feedback can also be obtained, and the retrieval and answer generation strategy can be optimized based on the user feedback. Among them, user feedback includes user feedback on the answers to the generated query questions. For example, the retrieval and answer generation strategy can be continuously optimized through user feedback through reinforcement learning algorithms. After the system generates the answer, the user's feedback information is recorded, and the strategy is updated using a reinforcement learning algorithm (such as Deep Q-Network) to maximize the expected cumulative return. Through repeated training and iteration, the system can be continuously improved in actual use to improve overall performance.

[0112] Reference Figure 7 , Figure 7 is a block diagram of the knowledge question answering system provided by an embodiment of the present application. Figure 7As shown, the knowledge question answering system 700 includes a context-aware multi-dimensional feature extraction model 710, a multi-level complexity-aware fusion network 720, and an improved retrieval-based enhanced generation model 730. Among them:

[0113] The context-aware multi-dimensional feature extraction model 710 is used to perform the following steps:

[0114] Obtaining the user's query questions and corresponding historical interaction records, and performing multidimensional feature extraction on the query questions and the corresponding historical interaction records to obtain a multidimensional feature set;

[0115] The multi-level complexity-aware fusion network 720 is used to perform the following steps:

[0116] Analyze each feature in the multidimensional feature set, identify the core features, and construct a core feature set, where the core features are the features that best reflect the complexity of the problem;

[0117] Use adaptive aggregation strategy to aggregate each core feature in the core feature set to obtain an aggregated feature vector;

[0118] The aggregated feature vector is input into a hierarchical complexity splitter for processing to obtain the complexity level corresponding to the query problem, wherein the hierarchical complexity splitter includes a lightweight classification layer and a deep classification layer. When the lightweight classification layer cannot determine the complexity level of the query problem, the deep classification layer uses a multi-head complexity decoder to determine the complexity level corresponding to the query problem;

[0119] The improved search formula enhancement generation model 730 is used to perform the following steps:

[0120] According to the complexity level corresponding to the query question, the corresponding retrieval and answer generation strategy is selected to generate the answer to the query question.

[0121] In the embodiment of the present application, the context-aware multidimensional feature extraction model 710 is used to perform Figure 1 In step S110, the multi-level complexity-aware fusion network 720 is used to perform Figure 1 The improved search formula enhancement generation model 730 is used to perform Figure 1 The specific processing process of step S150 is the same as that described above and will not be repeated here.

[0122] Continue to refer to Figure 7 and Figure 8 , Figure 8 Context-aware multidimensional feature extraction model 710 includes:

[0123] A dynamic context fusion module 711 is used to fuse the query question with the historical interaction record through a bidirectional attention mechanism to generate context information;

[0124] A multi-dimensional feature adaptive extraction module 712 is used to extract different types of features according to context information to obtain a candidate multi-dimensional feature set, wherein the multi-dimensional features include grammatical features, semantic features and retrieval information features;

[0125] A hierarchical complexity attention mechanism 713, which is used to calculate the attention weights at different complexity levels through multiple complexity-aware attention heads, and adjust the weights of each feature in the candidate multidimensional feature set according to the attention weights;

[0126] The adaptive feature enhancement layer 714 is used to perform feature transformation, filtering, reconstruction and completion processing on the candidate multi-dimensional feature set to obtain a multi-dimensional feature set.

[0127] The dynamic context fusion module 711 mainly includes a historical information encoding unit, a time-aware memory update unit and a context fusion unit. The historical information encoding unit is used to convert the historical interaction records into vector representations through an embedding layer. The time-aware memory update unit is used to store and update historical information in chronological order. The context fusion unit is used to fuse the current query question with the historical interaction records represented by the vector through a bidirectional attention mechanism to generate context information.

[0128] The multi-dimensional feature adaptive extraction module 712 is mainly used to simultaneously perform tasks such as grammatical complexity analysis, keyword density calculation, semantic embedding generation, and fuzzy reference detection, using a multi-task learning framework that shares underlying representations. At the same time, special processing channels can be designed for different types of features to ensure the optimal extraction and expression of various features.

[0129] Hierarchical Complexity Attention Mechanism 713 ensures the priority of important features by designing a multi-level attention mechanism and allocating attention weights hierarchically according to the complexity requirements of the features. Each attention head focuses on features at different complexity levels, which can improve the model's sensitivity and responsiveness to complexity. According to the actual complexity of the query problem, the weight allocation of attention heads at each layer is dynamically adjusted to achieve dynamic priority sorting of features.

[0130] The adaptive feature enhancement layer 714 can remove irrelevant features mainly through noise filtering and feature reconstruction, and complete the missing information through the feature reconstruction network.

[0131] In the embodiment of the present application, the context-aware multidimensional feature extraction model 710 is used to perform Figure 2 The specific processing procedures of the steps shown are the same as those described above and will not be repeated here.

[0132] Reference Figure 7 and Fig. 9 , Fig. 9 7 is a block diagram of the architecture of a multi-level complexity-aware fusion network provided in an embodiment of the present application. The multi-level complexity-aware fusion network 720 includes:

[0133] A complexity anchor point generation module 721 is used to construct a complexity anchor point, which is a benchmark for measuring complexity;

[0134] A dynamic complexity matching attention module 722 is used to compare each feature in the multidimensional feature set with the complexity anchor point to filter out the core features and construct a core feature set;

[0135] The adaptive aggregation and complexity-aware gating module 723 is used to aggregate the core features in the core feature set using an adaptive aggregation strategy to obtain an aggregated feature vector;

[0136] The hierarchical complexity splitter 724 is used to obtain the aggregated feature vector to output the complexity level corresponding to the query question, wherein the hierarchical complexity splitter 724 includes a lightweight classification layer and a deep classification layer. When the lightweight classification layer cannot determine the complexity level of the query question, the deep classification layer uses a multi-head complexity decoder to determine the complexity level corresponding to the query question.

[0137] In the embodiment of the present application, the multi-level complexity perception fusion network 720 is used to perform Figure 3-Figure 5 The specific processing procedures of the steps shown are the same as those described above and will not be repeated here.

[0138] It should be noted that after the complexity level corresponding to the query problem is output through the multi-level complexity-aware fusion network 720, the user's feedback on the complexity level of the output query problem can be obtained, and the multi-level complexity-aware fusion network 720 can be optimized according to the user feedback, and the complexity anchor point, complexity-aware gating parameters, and parameters of the multi-head complexity decoder can be optimized. Specifically, when the user gives feedback, the user feedback can be used as a reward signal, and the reinforcement learning module can be used to guide the multi-level complexity-aware fusion network 720 to reallocate attention weights, adjust the complexity anchor point position, or correct the complexity-aware gating parameters and parameters of the multi-head complexity decoder in the next round of interaction according to the reward signal, so that the multi-level complexity-aware fusion network 720 can more accurately extract core features and make correct complexity classification when facing similar problems.

[0139] The embodiment of the present application also provides an electronic device, the electronic device includes a memory and a processor, the memory stores a computer program, and the processor implements the above-mentioned answer generation method when executing the computer program. The electronic device can be any intelligent terminal including a tablet computer, a car computer, etc.

[0140] See also Fig.10 , Fig.10 : is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present application, the electronic device includes:

[0141] The processor 1001 may be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present application;

[0142] The memory 1002 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 1002 can store an operating system and other applications. When the technical solution provided in the embodiments of this specification is implemented by software or firmware, the relevant program code is stored in the memory 1002, and the processor 1001 calls and executes the answer generation method of the embodiment of this application;

[0143] Input / output interface 1003, used to implement information input and output;

[0144] The communication interface 1004 is used to realize the communication interaction between the device and other devices. The communication can be realized through a wired manner (such as USB, network cable, etc.) or a wireless manner (such as mobile network, WIFI, Bluetooth, etc.);

[0145] A bus 1005 , which transmits information between various components of the device (e.g., the processor 1001 , the memory 1002 , the input / output interface 1003 , and the communication interface 1004 );

[0146] The processor 1001 , the memory 1002 , the input / output interface 1003 and the communication interface 1004 are connected to each other in communication within the device via the bus 1005 .

[0147] The embodiments described in the embodiments of the present application are intended to more clearly illustrate the technical solutions of the embodiments of the present application, and do not constitute a limitation on the technical solutions provided in the embodiments of the present application. Those skilled in the art will appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0148] Those skilled in the art will appreciate that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and may include more or fewer steps than shown in the figures, or a combination of certain steps, or different steps.

[0149] The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separated, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0150] Those skilled in the art will appreciate that all or some of the steps in the methods disclosed above, and the functional modules / units in the systems and devices may be implemented as software, firmware, hardware, or a suitable combination thereof.

[0151] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0152] It should be understood that in the present application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0153] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the above units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0154] The units described above as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0155] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0156] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including multiple instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, referred to as ROM), random access memory (Random Access Memory, referred to as RAM), disk or optical disk and other media that can store programs.

[0157] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but the scope of the rights of the present invention is not limited thereto. Any modification, equivalent substitution and improvement made by a person skilled in the art without departing from the scope and substance of the present invention should be within the scope of the rights of the present invention.

Claims

1. A method for generating an answer, characterized in that: The method comprises: Obtaining a user's query question and a corresponding historical interaction record, and performing multidimensional feature extraction on the query question and the corresponding historical interaction record to obtain a multidimensional feature set; Analyze each feature in the multidimensional feature set, identify core features, and construct a core feature set, wherein the core features are the features that best reflect the complexity of the problem; Using an adaptive aggregation strategy, aggregating each core feature in the core feature set to obtain an aggregated feature vector; Inputting the aggregated feature vector into a hierarchical complexity splitter for processing to obtain a complexity level corresponding to the query question, wherein the hierarchical complexity splitter includes a lightweight classification layer and a deep classification layer, and when the lightweight classification layer cannot determine the complexity level of the query question, the deep classification layer uses a multi-head complexity decoder to determine the complexity level corresponding to the query question; According to the complexity level corresponding to the query question, a corresponding retrieval and answer generation strategy is selected to generate an answer to the query question.

2. The method according to claim 1, characterized in that Perform multidimensional feature extraction on the query question and the corresponding historical interaction record to obtain a multidimensional feature set, including: The query question is integrated with the historical interaction record through a bidirectional attention mechanism to generate context information; Extracting different types of features according to the context information to obtain a candidate multidimensional feature set, wherein the candidate multidimensional feature set includes grammatical features, semantic features, and retrieval information features; By using a plurality of complexity-aware attention heads, attention weights at different complexity levels are calculated respectively, and the weight of each feature in the candidate multidimensional feature set is adjusted according to the attention weights; The candidate multidimensional feature set is subjected to feature transformation, filtering, reconstruction and completion processing to obtain a multidimensional feature set.

3. The method according to claim 1, characterized in that The analyzing each feature in the multidimensional feature set, identifying the core features, and constructing the core feature set includes: Constructing a complexity anchor point, wherein the complexity anchor point is a benchmark for measuring complexity; Comparing each feature in the multidimensional feature set with the complexity anchor point to screen out core features; The filtered core features are collected to obtain a core feature set.

4. The method according to claim 3, characterized in that The adaptive aggregation strategy is used to aggregate the core features in the core feature set to obtain an aggregated feature vector including: Determine the weight corresponding to each core feature according to the relative importance of each core feature in the core feature set and the matching degree between the core feature and the complexity anchor point; According to the weights corresponding to each core feature, an adaptive aggregation strategy is used to aggregate each core feature to obtain an aggregated feature vector.

5. The method according to claim 1, characterized in that The step of inputting the aggregated feature vector into a hierarchical complexity splitter for processing to obtain a complexity level corresponding to the query question includes: Inputting the aggregated feature vector into a hierarchical complexity splitter, obtaining the aggregated feature vector by the lightweight classification layer, and determining the complexity of the query question; If the lightweight classification layer cannot determine the complexity of the query question, the deep classification layer uses a multi-head complexity decoder to determine the complexity level corresponding to the query question and output it; The deep classification layer is used to predict the number of sub-questions, domain knowledge call tags, and knowledge verification tags.

6. The method according to claim 1, characterized in that The complexity levels include a first level, a second level, a third level and a fourth level, wherein the complexity represented by the first level, the second level, the third level and the fourth level increases in sequence, and the corresponding retrieval and answer generation strategy is selected according to the complexity level corresponding to the query question to generate an answer to the query question, including: If the complexity level corresponding to the query question is the first level, the internal knowledge of the large language model is used to generate a corresponding answer; If the complexity corresponding to the query question is the second level, documents related to the query question are retrieved from an external knowledge base, and the query question and the retrieved documents are input into a large language model to generate an answer to the query question; If the complexity corresponding to the query question is the third level, the query question is decomposed into a plurality of sub-questions, and each of the sub-questions is searched and a sub-answer of each sub-question is generated respectively, so as to generate an answer to the query question according to all the sub-answers; If the complexity corresponding to the query question is the fourth level, multi-source information related to the query question is retrieved, and an answer to the query question is generated according to the multi-source information.

7. The method according to claim 1, characterized in that After selecting a corresponding retrieval and answer generation strategy according to the complexity level corresponding to the query question to generate an answer to the query question, the method further includes: Obtaining user feedback, wherein the user feedback includes user feedback on the generated answer to the query question; Retrieval and answer generation strategies are optimized based on the user feedback.

8. A knowledge question answering system, characterized in that: It includes a context-aware multidimensional feature extraction model, a multi-level complexity-aware fusion network, and an improved retrieval-based enhanced generation model, among which: The context-aware multidimensional feature extraction model is used to perform the following steps: Obtaining a user's query question and a corresponding historical interaction record, and performing multidimensional feature extraction on the query question and the corresponding historical interaction record to obtain a multidimensional feature set; The multi-level complexity-aware fusion network is used to perform the following steps: Analyze each feature in the multidimensional feature set, identify core features, and construct a core feature set, wherein the core features are the features that best reflect the complexity of the problem; Using an adaptive aggregation strategy, aggregating each core feature in the core feature set to obtain an aggregated feature vector; Inputting the aggregated feature vector into a hierarchical complexity splitter for processing to obtain a complexity level corresponding to the query question, wherein the hierarchical complexity splitter includes a lightweight classification layer and a deep classification layer, and when the lightweight classification layer cannot determine the complexity level of the query question, the deep classification layer uses a multi-head complexity decoder to determine the complexity level corresponding to the query question; The improved search-based enhanced generation model is used to perform the following steps: According to the complexity level corresponding to the query question, a corresponding retrieval and answer generation strategy is selected to generate an answer to the query question.

9. The system according to claim 8, characterized in that The context-aware multidimensional feature extraction model includes: A dynamic context fusion module, used to fuse the query question with the historical interaction record through a bidirectional attention mechanism to generate context information; A multi-dimensional feature adaptive extraction module, used to extract different types of features according to the context information to obtain a candidate multi-dimensional feature set, wherein the multi-dimensional features include grammatical features, semantic features and retrieval information features; A hierarchical complexity attention mechanism, for calculating attention weights at different complexity levels through multiple complexity-aware attention heads, and adjusting the weights of each feature in the candidate multidimensional feature set according to the attention weights; The adaptive feature enhancement layer is used to perform feature transformation, filtering, reconstruction and completion processing on the candidate multidimensional feature set to obtain a multidimensional feature set.

10. An electronic device, characterized in that: The electronic device comprises a memory and a processor, the memory stores a computer program, and the processor implements the method according to any one of claims 1 to 7 when executing the computer program.

Images