Complex question and answer method and system based on adaptive task deconstruction and multi-modal evidence aggregation

Abstract

The invention provides a complex question and answer method and system based on adaptive task deconstruction and multi-modal evidence aggregation, belongs to the technical field of natural language processing, and designs a dynamic Few-shot prompt construction method based on dependency syntax fingerprints to ensure that a prompt template is matched with a question structure; the invention discloses a dynamic problem deconstruction method based on confidence evaluation and auto-reflection. The method comprises the following steps: recursively decomposing a problem tree by using a large language model; converting the problem tree into a standardized linear task execution sequence by a problem tree context dependence specification and task sequence generation method; obtaining a high-correlation evidence set of each task based on an evidence generation method of two-way recall and cross encoder rearrangement; and the task sequence is reasoned and dynamically optimized by a question answer extraction method based on double-strategy aggregation reasoning. According to the method, the accurate complex question and answer result can be provided on the premise of ensuring the question disassembling quality, restraining error propagation and comprehensively recalling evidences.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing, in particular to a complex question answering method and system based on adaptive task decomposition and multi-modal evidence aggregation. BACKGROUND

[0002] Multi-hop complex question answering is an important challenge in the field of natural language processing, which requires integrating information from multiple documents or knowledge sources and conducting multi-step reasoning to answer a single complex question. Traditional retrieval-augmentation generation methods usually directly use the original question for evidence retrieval, but this is not effective for multi-hop questions where information is scattered in different documents. Question decomposition has been proven to be an effective strategy to solve the retrieval problem of multi-hop questions, that is, first decompose the complex question into independent sub-questions, then retrieve relevant evidence and integrate reasoning.

[0003] Early decomposition methods rely on rules or dependency syntax analysis to hard-decompose complex questions into simple questions, which has poor generalization ability and is difficult to handle the diversity and dynamics of language expression. In recent years, the enhancement of large language models (LLM) provides a new way for question decomposition, and its powerful language understanding and generation capabilities are used to decompose complex questions and provide a basis for multi-step reasoning. However, without additional mechanisms, the sub-questions generated by LLM often lack necessary entity references or fail to fully capture the original intent, resulting in incomplete semantics. This defect in initial decomposition directly affects the quality and relevance of evidence retrieval. Multi-hop question reasoning also has the vulnerability of error propagation. If an error occurs at any intermediate step of decomposition, sub-question answering, or evidence integration, the error will propagate along the reasoning chain and be amplified, significantly reducing the accuracy and reliability of the final answer. In addition, existing retrieval methods rely on a single modality, making it difficult to fully recall all evidence scattered in multiple documents or requiring precise keyword and detailed semantic understanding. Therefore, how to ensure high-quality question decomposition, suppress error propagation, and achieve comprehensive evidence retrieval while providing users with accurate, robust, and explainable complex question answering results has become a key problem that needs to be solved. Multi-hop complex question answering requires integrating information from multiple documents or knowledge sources for multi-step reasoning. However, existing methods generally have problems such as inaccurate question decomposition, unclear context dependence, low retrieval efficiency, unstable answer quality, and high risk of error propagation. Therefore, there is an urgent need for an efficient, accurate, and reliable intelligent question answering method that can fully represent the semantic link between sub-tasks in multi-hop questions. SUMMARY

[0004] The present application aims to provide a complex question answering method and system based on adaptive task decomposition and multi-modal evidence aggregation to solve at least one of the technical problems in the background art.

[0005] To achieve the above object, the present application adopts the following technical solutions:

[0006] In a first aspect, the present application provides a complex question answering method based on adaptive task decomposition and multi-modal evidence aggregation, comprising: extracting the syntactic fingerprint of the question based on dependency syntax analysis technology and matching it with a pre-built multi-hop question disassembly example library, combining the Few-shot example obtained by matching with a fixed template to generate a question decomposition prompt template; driving the LLM to recursively decompose the question based on dynamic Few-shot prompts to form a preliminary question tree, triggering a reconstruction mechanism for low-quality question trees through confidence evaluation and normative verification until a logically rigorous question tree is generated; post-processing the question tree to standardize the reference relationship between nodes to accurately define sub-task dependencies, and converting the question tree into a linear task execution sequence using a post-order traversal algorithm; using dense and sparse retrievers to perform parallel two-way recall on each sub-task in the task sequence, using a cross-encoder to perform secondary rearrangement on the merged candidate documents to generate a Top-K evidence set with high relevance; reasoning and answer extraction for each task from bottom to top according to the task sequence, for non-leaf nodes, performing two strategies of direct extraction and aggregated reasoning in parallel and selecting the optimal answer by comparing confidence, and finally obtaining the complete answer of the question tree.

[0007] In a second aspect, the present application provides a complex question answering system based on adaptive task decomposition and multi-modal evidence aggregation, comprising: a matching module for extracting the syntactic fingerprint of the question based on dependency syntax analysis technology and matching it with a pre-built multi-hop question disassembly example library, combining the Few-shot example obtained by matching with a fixed template to generate a question decomposition prompt template; a decomposition module for driving the LLM to recursively decompose the question based on dynamic Few-shot prompts to form a preliminary question tree, triggering a reconstruction mechanism for low-quality question trees through confidence evaluation and normative verification until a logically rigorous question tree is generated; a conversion module for post-processing the question tree to standardize the reference relationship between nodes to accurately define sub-task dependencies, and converting the question tree into a linear task execution sequence using a post-order traversal algorithm; a recall module for using dense and sparse retrievers to perform parallel two-way recall on each sub-task in the task sequence, using a cross-encoder to perform secondary rearrangement on the merged candidate documents to generate a Top-K evidence set with high relevance; an extraction and aggregation module for reasoning and answer extraction for each task from bottom to top according to the task sequence, for non-leaf nodes, performing two strategies of direct extraction and aggregated reasoning in parallel and selecting the optimal answer by comparing confidence, and finally obtaining the complete answer of the question tree.

[0008] In a third aspect, the present application provides a non-transitory computer readable storage medium for storing computer instructions, which, when executed by a processor, implement the complex question answering method based on adaptive task decomposition and multi-modal evidence aggregation as described in the first aspect.

[0009] In a fourth aspect, the present application provides a computer device comprising a memory and a processor, the processor and the memory being in communication with each other, the memory storing program instructions executable by the processor, and the processor invoking the program instructions to execute the complex question answering method based on adaptive task decomposition and multi-modal evidence aggregation as described in the first aspect.

[0010] In a fifth aspect, the present application provides an electronic device comprising a processor, a memory and a computer program, wherein the processor is connected to the memory, and the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to make the electronic device execute instructions for implementing the complex question answering method based on adaptive task decomposition and multi-modal evidence aggregation as described in the first aspect.

[0011] The present application has the following advantages: a dynamic Few-shot prompt based on dependency syntax fingerprints can be constructed, ensuring that the prompt template is highly matched with the syntax structure of new questions, solving the problem of insufficient generalization ability of traditional methods, and providing high-quality input for subsequent LLM decomposition tasks; the initial question tree generated by the LLM is automatically evaluated and optimized in quality through a confidence evaluation and self-reflection mechanism, effectively preventing error propagation caused by inaccurate initial decomposition, and ensuring the logical rigor of the question tree; the anaphora placeholders in the question tree are standardized, the context dependency relationship between sub-tasks is accurately defined, and the question tree is converted into a linear task execution sequence, providing a strict execution order and logical basis for subsequent evidence retrieval, fusion and answer extraction; a hybrid multi-level retrieval strategy of double-path recall and cross-encoder rearrangement is adopted to realize comprehensive recall and high-precision filtering of multi-source evidence, and improve the quality and relevance of the evidence set of each sub-task; the task sequence is inferred from bottom to top, and the two strategies of direct extraction and aggregation reasoning are executed in parallel, and the optimal answer is dynamically selected by combining confidence evaluation, thereby enhancing the accuracy and robustness of the final answer.

[0012] The advantages of the additional aspects of the present application will be more apparent from the following description section or will be understood through the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0014] Figure 1 The flowchart of the complex question and answer method based on adaptive task decomposition and multi-modal evidence aggregation according to the embodiments of the present application.

[0015] Figure 2 The flowchart of the dynamic question decomposition method based on confidence evaluation and self-reflection according to the embodiments of the present application.

[0016] Figure 3 The flowchart of the question answer extraction method based on double-strategy aggregation reasoning according to the embodiments of the present application. DETAILED DESCRIPTION

[0017] The embodiments of the present application will be described in detail below, and the examples of the embodiments are shown in the drawings, wherein the same or similar notations represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below through the drawings are exemplary and are only used to explain the present application, and cannot be interpreted as a limitation on the present application.

[0018] The application provides a complex question answering method based on adaptive task decomposition and multi-modal evidence aggregation. First, a multi-hop question disassembly example library is constructed by manual annotation, and the core syntactic structure features of the question are extracted from multiple dimensions such as interrogative words, dependency paths, dependency relationship sequences, part-of-speech sequences, clause boundaries, and functional word modification based on dependency syntax analysis technology to construct a syntactic fingerprint. Further, the syntactic fingerprint of the new input original question is generated, and a similarity function is constructed to dynamically search the example library, select the top K disassembly examples with the highest similarity scores, and splice them with the fixed prompt template to form a complete Few-shot prompt. Second, the LLM is driven by the dynamic Few-shot prompt to recursively decompose the original question and obtain an initial question tree. On this basis, the quality of the preliminary decomposition result generated by the LLM is automatically evaluated, and when the confidence of any question group in the question tree is lower than the preset threshold or a normative violation is detected, the self-reflection and reconstruction mechanism is triggered. The mechanism takes the low-quality question tree as a counterexample, combines the reflection prompt to re-drive the LLM for decomposition until a stable and rigorous logical question tree is generated. Third, the generated question tree is post-processed, and a breadth-first traversal method is used to correct the reference placeholders in each sub-question text. If the placeholder is an inherited reference, the number of placeholders is adjusted to be consistent with the parent question. If it is a newly generated reference, it is adjusted according to its depth. On this basis, a post-order traversal algorithm is used to convert the hierarchical tree structure into a linear task execution sequence, ensuring that each sub-task in the sequence appears before its dependent parent task. Then, the sparse retrieval based on the BM25 retriever and the dense retrieval based on the dense retriever are performed in parallel to obtain a comprehensive candidate document set. Further, the cross-encoder is used to rearrange the candidate document set to improve the relevance of the candidate paragraphs and sub-tasks, and finally generate a Top-K evidence set with high relevance. Finally, according to the post-order execution sequence of the task sequence, knowledge fusion and answer extraction are performed for each task from bottom to top. For leaf nodes without sub-task dependency, the evidence set is directly used to extract the answer by LLM. For non-leaf nodes with cross-task dependency, the direct extraction strategy and the aggregation reasoning strategy are executed in parallel, and the optimal answer is selected by comparing the answer confidence scores generated by the two strategies and combining the effectiveness verification. This iteration continues until the answer extraction of the root node is completed, and finally the complete answer to the original question is obtained.

[0019] The dependency syntax fingerprint-based dynamic Few-shot prompt construction method of the application dynamically selects the most similar example to the new question structure from the example library through syntax fingerprint matching, constructs a high-quality prompt template, effectively solves the problem of insufficient generalization ability of the fixed prompt template in the traditional method, and provides high-quality and targeted input for subsequent large language model problem decomposition tasks. A dynamic question decomposition method based on confidence evaluation and self-reflection is designed, which realizes the automatic evaluation and optimization of the quality of question decomposition by recursive decomposition and the introduction of self-reflection and reconstruction mechanism, effectively suppresses the error propagation caused by initial decomposition defects, and ensures the logical rigor of the question tree. A question tree context dependency specification and task sequence generation method is designed, which standardizes the reference placeholders between nodes, accurately defines the context dependency relationship between subtasks, and converts the hierarchical tree structure into a linear task execution sequence, providing a rigorous execution order and logical basis for the subsequent evidence aggregation-based reasoning process. An evidence generation method based on double-path recall and cross-encoder rearrangement is designed, which realizes the full recall and high-precision filtering of the evidence by parallel execution of sparse retrieval and dense retrieval, and fine rearrangement by using a cross-encoder, and improves the quality of the evidence set of each subtask. A question answer extraction method based on double-strategy aggregation reasoning is designed, which dynamically selects the optimal answer by executing the direct extraction strategy on the leaf nodes and executing the direct extraction and aggregation reasoning strategies in parallel on the non-leaf nodes and combining the confidence evaluation, enhancing the accuracy and robustness of the final answer.

[0020] Example 1

[0021] In this embodiment 1, first, a complex question answering system based on adaptive task decomposition and multi-modal evidence aggregation is provided, including: a matching module for extracting the syntactic fingerprint of the question based on dependency syntax analysis technology and matching with a pre-built multi-hop question disassembly example library, combining the matched Few-shot example with a fixed template to generate a question decomposition prompt template. The decomposition module is used to drive the LLM to recursively decompose the question based on dynamic Few-shot prompts, form a preliminary question tree, trigger a reconstruction mechanism through confidence evaluation and normative verification for low-quality question trees, until a logically rigorous question tree is generated. The conversion module is used to post-process the question tree, standardize the reference relationship between nodes to accurately define sub-task dependencies, and use a post-order traversal algorithm to convert the question tree into a linear task execution sequence. The recall module is used to perform parallel double-channel recall on each sub-task in the task sequence using dense and sparse retrievers, and use a cross-encoder to perform secondary rearrangement on the merged candidate documents to generate a Top-K evidence set with high relevance. The extraction and aggregation module is used to infer and extract answers for each task from bottom to top according to the task sequence, and for non-leaf nodes, two strategies of direct extraction and aggregated reasoning are executed in parallel and the optimal answer is selected by comparing the confidence, and finally the complete answer of the question tree is obtained.

[0022] In this embodiment, based on the above system, a complex question answering method based on adaptive task decomposition and multi-modal evidence aggregation is implemented, including: extracting the syntactic fingerprint of the question based on dependency syntax analysis technology and matching with a pre-built multi-hop question disassembly example library, combining the matched Few-shot example with a fixed template to generate a question decomposition prompt template; driving the LLM to recursively decompose the question based on dynamic Few-shot prompts, forming a preliminary question tree, triggering a reconstruction mechanism through confidence evaluation and normative verification for low-quality question trees, until a logically rigorous question tree is generated; post-processing the question tree, standardizing the reference relationship between nodes to accurately define sub-task dependencies, and using a post-order traversal algorithm to convert the question tree into a linear task execution sequence; using dense and sparse retrievers to perform parallel double-channel recall on each sub-task in the task sequence, and using a cross-encoder to perform secondary rearrangement on the merged candidate documents to generate a Top-K evidence set with high relevance; inferring and extracting answers for each task from bottom to top according to the task sequence, and for non-leaf nodes, two strategies of direct extraction and aggregated reasoning are executed in parallel and the optimal answer is selected by comparing the confidence, and finally the complete answer of the question tree is obtained.

[0023] The syntax fingerprint of the question is extracted based on the dependency syntax analysis technology, and is matched with a pre-constructed multi-hop question disassembly example library. The few-shot example matched is combined with a fixed template to generate a question disassembly prompt template, which specifically includes: a multi-hop question disassembly example library is constructed by manual annotation, each example in the example library includes an original question and its corresponding decomposition sequence; for each original question, the core syntax structure features are extracted based on the dependency syntax analysis technology, and the syntax fingerprint is constructed, which is a tuple containing multi-dimensional features; for a new input original multi-hop question, its corresponding syntax fingerprint is also generated; a similarity function is constructed to measure the similarity between the syntax fingerprint of the new input original multi-hop question and any syntax fingerprint in the example library; the example library is dynamically searched based on the similarity function, and the K example sets with the highest similarity scores are selected; the K similar disassembly example sets searched are sequentially spliced with the template in the format of and The new question is appended to the end in the format of and to form a complete few-shot prompt as the input of the subsequent LLM to perform the question disassembly task.

[0024] The LLM is driven by the dynamic few-shot prompt to recursively disassemble the question to form a preliminary question tree, and a reconstruction mechanism is triggered for low-quality question trees through confidence evaluation and normative verification until a logically rigorous question tree is generated, which specifically includes: the generated dynamic few-shot prompt drives the LLM to preliminarily recursively disassemble a series of original questions to obtain an initial question tree set ; the preliminary disassembly process converts each original question into a structured JSON object in the form of ; at the same time, the text response generated by the LLM and its corresponding word token probability sequence are recorded; the preliminary disassembly result generated by the LLM is automatically evaluated, and the evaluation is independently performed for each question group in the question tree.

[0025] The question tree is post-processed to standardize the reference relationship between nodes to accurately define the sub-task dependency, and a post-order traversal algorithm is used to convert the question tree into a linear task execution sequence, which specifically includes: the question text of each node in the question tree is identified and parsed to identify the reference placeholder ​; the placeholder adopts a specific format, i.e. a number of # symbols followed by a number, and the parsing logic of this format is as follows: the meaning of the number of # symbols in the placeholder: the number of # symbols is used to indicate the reference level; the meaning of the number in the placeholder: the number part represents the position index of the question in the task list at the specified level, counting from 1; the problem tree is processed hierarchically using the breadth-first traversal method, and the reference placeholder in the text of each sub-question is processed in the traversal process correction is performed.

[0026] Parallel dual recall is performed on each sub-task in the task sequence using dense and sparse retrievers, and the merged candidate document set is rearranged again using a cross encoder to generate a Top-K evidence set with high relevance, which specifically includes: the output linear task sequence is traversed, where represents the th task node in the sequence, and each task node corresponds to a sub-task to be processed, and the question text is recorded as ; dual parallel recall is performed to obtain a preliminary candidate document set ; sparse retrieval: using a BM25 retriever based on keyword matching, Top-K documents are recalled from the pre-constructed corpus using as the query text; dense retrieval: using a dense retriever based on semantic vectors, encoding into vector , and all documents in the corpus have been pre-encoded into a vector set ; by calculating the vector cosine similarity, Top-K documents that are most relevant in semantics are recalled; the merged candidate document set is rearranged again using a cross encoder; the cross encoder receives the question text and each candidate document as input pairs, calculates the fine relevance score between them; the is sorted in descending order according to the rearrangement score , and the Top-K documents with the highest scores are selected as the final evidence set of the sub-task , which is used as input for the subsequent answer extraction and reasoning process.

[0027] The tasks are inferred and answers are extracted from the task sequence from bottom to top, and for non-leaf nodes, direct extraction and aggregated reasoning strategies are executed in parallel and the optimal answer is selected by comparing the confidence, and finally the complete answer to the problem tree is obtained, which specifically includes: the generated linear task sequence And following the post-order traversal of the sequence, perform a bottom-up process on each task node. Process; for leaf nodes in the sequence, extract the question text from that leaf node. and the Top-K evidence set rearranged by the cross encoder in step 4. Input to LLM; LLM based on The strategy of directly extracting information to generate answers And calculate its average log probability as the confidence score. For non-leaf nodes in the sequence that have cross-task dependencies, the answer extraction process employs a dual-strategy parallel execution; the question text for this non-leaf node is denoted as... Its subtask ID list is Due to the bottom-up nature of the sequence, all its subtasks The answer All have been generated and stored.

[0028] Example 2

[0029] This embodiment 2 discloses a complex question-answering method based on adaptive task deconstruction and multimodal evidence aggregation. First, a dynamic few-shot hint construction method based on dependency syntactic fingerprinting is designed to ensure that the hint template matches the question structure. Second, a dynamic question deconstruction method based on confidence assessment and self-reflection is designed, recursively decomposing the question tree using a large language model. Third, a question tree context dependency specification and task sequence generation method is designed to transform the question tree into a standardized linear task execution sequence. Then, an evidence generation method based on dual-path recall and cross-encoder rearrangement is designed to obtain highly relevant evidence sets for each task. Finally, a question answer extraction method based on dual-strategy aggregation reasoning is designed to reason about the task sequence and dynamically select the best answer. This method can provide accurate complex question-answering results while ensuring the quality of question deconstruction, curbing error propagation, and achieving comprehensive evidence recall.

[0030] This method specifically includes the following steps:

[0031] Step 1: Extract the syntactic fingerprint of the problem based on dependency parsing technology and match it with a pre-built multi-hop problem decomposition example library. Combine the matched Few-shot examples with a fixed template to generate a problem decomposition hint template.

[0032] Step 2: Based on dynamic Few-shot hints, drive LLM to recursively deconstruct the problem to form a preliminary problem tree. Through confidence assessment and normative verification, trigger the reconstruction mechanism for low-quality problem trees until a logically rigorous problem tree is generated.

[0033] Step 3: Post-processing of the problem tree, standardizing the reference relationship between nodes to accurately define the sub-task dependency, and converting the problem tree into a linear task execution sequence using a post-order traversal algorithm.

[0034] Step 4: Using dense and sparse retrievers for parallel two-way recall of each sub-task in the task sequence, and using a cross-encoder for secondary rearrangement of the merged candidate documents to generate a Top-K evidence set with high relevance.

[0035] Step 5: Reasoning and answer extraction for each task from bottom to top according to the task sequence, parallel execution of direct extraction and aggregated reasoning strategies for non-leaf nodes, and selection of the optimal answer by comparing confidence levels to obtain the complete answer of the problem tree.

[0036] In step 1, based on dependency syntax analysis technology, the syntax fingerprint of the question is extracted and matched with the pre-built multi-hop question disassembly example library. The matched Few-shot example is combined with the fixed template to generate a question deconstruction prompt template, which includes:

[0037] A multi-hop question disassembly example library is constructed through manual annotation . Each example in contains an original question and its corresponding decomposition sequence .

[0038] For each question in , based on dependency syntax analysis technology, the core syntactic structure features are extracted to construct the syntax fingerprint , which is a tuple containing multiple features, defined as follows:

[0039]

[0040] Among them, is the interrogative word feature, representing the set of interrogative words in the question ; is the dependency path feature, capturing the set of dependency paths from the key information word to the main sentence root node; is the dependency relation sequence, recording the dependency relation type sequence of all words in the question , is the part-of-speech sequence, recording the part-of-speech sequence of all words in the question , is the subordinate clause boundary feature, identifying and quantifying the types of subordinate clauses contained in the question; is the functional word modification feature, identifying and counting whether there are prepositions or subordinate conjunctions in the dependency path.

[0041] For the original multi-hop problem of new input Similarly, its corresponding syntactic fingerprint is generated. .

[0042] Constructing a similarity function ,measure With any syntactic fingerprint in the example library The similarity is calculated using a weighted sum of features across all dimensions.

[0043]

[0044] in, Representing feature dimension, For the weights of each dimension, This is the similarity measurement function for the corresponding dimension. For set-type features, The function is Jaccard similarity. For path features, a customized path similarity function is used. ,in, Represents the set of paths to the new problem; Represents the set of paths for the example problem. It calculates two paths. and A function of Jaccard similarity between them; and These are the sizes of the two path sets, respectively.

[0045] Based on similarity function For example libraries Perform dynamic retrieval and select the K example sets with the highest similarity scores. , where K is a preset integer.

[0046] The design includes a fixed prompt template containing instructions to guide the LLM in performing tasks. and define and The structured format.

[0047] The retrieved set of K similar decomposition examples ,according to and The format is the same as the template. The pieces are then assembled. Ultimately, this addresses the new problem. by and The format is appended to the end to form a complete Few-shot prompt. As input for subsequent LLM execution problem deconstruction tasks.

[0048] In step 2, the dynamic Few-shot prompt drives the LLM to recursively deconstruct the problem, forming a preliminary problem tree. Through confidence evaluation and normative verification, the low-quality problem tree triggers a reconstruction mechanism until a logically rigorous problem tree is generated. Specifically, it includes:

[0049] Using the dynamic Few-shot prompt generated in step 1 Drive LLM, a series of original questions Preliminary recursive deconstruction to obtain the initial problem tree set .

[0050] The preliminary deconstruction process converts each original question into a structured JSON object in the form of . At the same time, record the text response generated by LLM and its corresponding word token log probability sequence.

[0051] Automated quality assessment of the preliminary deconstruction results generated by LLM. This assessment is performed independently for each question group in the problem tree (i.e. a parent question and its directly decomposed child question list), including the following two aspects:

[0052] (1) Confidence evaluation: For each question group in the problem tree , calculate the average log probability of its decomposition sequence as the confidence score . The calculation method of the score is as follows: 1) Identify the word span in the JSON structure representing the sequence of sub-questions. 2) Extract all word log probability sequences within this span. 3) Calculate the average log probability score : ; 4) Compare the score with the preset threshold .

[0053] (2) Normative verification: Perform multiple normative checks on each question group to identify structural errors in the deconstruction: 1) repeated sub-question violation: detect whether there is a sub-question in the decomposition sequence that is the same as the parent question. 2) insufficient decomposition violation: detect whether any question is decomposed into less than two sub-questions.

[0054] Design a self-reflection and reconstruction mechanism to correct low-quality preliminary deconstruction results. This mechanism is based on the following logic for loop iteration:

[0055] (1) When the confidence score of any question group in the problem tree < Or when any of the above normative violations are detected, the reconstruction process is triggered.

[0056] (2) The current low-quality problem tree decomposition result JSON string is taken as a counterexample The dynamic part of the reflection prompt template is constructed, and the instructions for re-emphasizing the decomposition rules and checking the counterexample are taken as the static part of the reflection prompt template. The two parts together constitute a complete reflection prompt template , guiding the LLM to reflect on and correct the root problem of .

[0057] (3) The constructed reflection prompt is input into the LLM. The LLM self-corrects the counterexample and generates a revised new decomposition result .

[0058] (4) The newly generated problem tree is again subjected to confidence assessment and normative verification. If it still does not meet all quality standards (i.e., the confidence score of any problem group is less than the pre-set threshold, or normative violations are detected), the next reflection reconstruction cycle is continued until the pre-set maximum number of reconstructions K is reached.

[0059] (5) When the latest problem tree generated by the LLM meets all quality standards (i.e., the confidence score of all problem groups ≥ and there are no normative violations), the cycle is terminated, and the problem tree is adopted as the final decomposition result. If the cycle is terminated due to reaching the pre-set maximum number of reconstructions K, the problem tree obtained by the last (Kth) reconstruction is adopted as the final decomposition result.

[0060] In step 3, the problem tree is post-processed to precisely define the subtask dependencies by standardizing the reference relationships between nodes. The problem tree is converted into a linear task execution sequence using a post-order traversal algorithm, which specifically includes:

[0061] Identify and parse the reference placeholders in the problem text of each node in the problem tree that represent context dependency relationships. The placeholders use a specific format, i.e., a number of # symbols followed by a number (e.g., #1, ##2). The parsing logic for this format is as follows:

[0062] (1) The meaning of the number of # symbols in the placeholder: The number of # symbols indicates the reference level.

[0063] 1) When the number of # is equal to the depth of the current problem node in the problem tree, the placeholder is used to refer to the task at the same level, i.e. a certain task in the list of sub-tasks of the parent problem to which the current problem belongs.

[0064] 2) When the number of # is less than the depth of the current problem node, the placeholder is used to refer to the task at the upper level, i.e. a certain task in the corresponding problem group at the upper level, whose parent problem is a sub-task.

[0065] (2) The meaning of the number in the placeholder: the number part represents the position index of the problem in the task list at the specified level, counting from 1.

[0066] The breadth-first traversal method is used to process the problem tree hierarchically, and the reference placeholders in the text of each sub-problem are corrected during the traversal process. The correction process follows the following double logic:

[0067] (1) Inheritance reference judgment: when processing any problem node, first identify and extract all reference placeholders and their local text context before and after from the parent problem text, and construct a reference set containing (placeholder, before and after text) tuples. Then, traverse the text of the current sub-problem, identify the placeholders in it, and extract their corresponding local text context. If the local text context of the placeholder in the sub-problem matches the before and after text of any tuple in the reference set, it is determined that the placeholder is directly inherited from the parent problem. At this time, the placeholder in the sub-problem is corrected so that the number of # is exactly the same as the corresponding placeholder in the parent problem.

[0068] (2) New generation reference correction: if the placeholder is not inherited from the parent problem, it is considered as a newly generated reference at the current level. At this time, according to the depth d of the current problem node (the root node depth is 0), the number of # is uniformly adjusted to d+1.

[0069] Convert the standardized problem tree data structure generated by the above steps into a decomposition mapping convenient for traversal , whose structure is . The mapping takes each parent problem node as the key and its direct sub-problem list as the value. Remove any self-reference that may cause a cycle (i.e. the case where the sub-problem is exactly the same as its parent problem).

[0070] Post-order traversal and serialization: use the post-order traversal algorithm to perform a depth-first traversal of the problem tree. The algorithm starts from the root problem and recursively visits all nodes in the order of .

[0071] During the traversal process, each visited problem node and its direct parent node form a tuple ​and appended to the linear list in turn . a preliminary task sequence that strictly follows the contextual dependency relationship, in which all subtasks appear before their parent task.

[0072] metadata construction is performed on each task node in the sequence . First, each task node is assigned an incremental unique integer ID according to the traversal order in , and an ID mapping is established.

[0073] The problem text of each task node is parsed, and the reference placeholders therein are mapped to the corresponding task IDs according to the aforementioned placeholder resolution rules, to construct a contextual dependency ID list .

[0074] The direct subtask list of the current task is obtained according to , and is mapped to a subtask ID list . Information including the unique ID, normalized problem text, contextual dependency ID list , and subtask ID list is encapsulated as metadata, and is assembled into the final linear task sequence .

[0075] In step 4, dense and sparse retrievers are used to perform parallel two-way recall on each subtask in the task sequence, and a cross encoder is used to perform secondary rearrangement on the merged candidate documents to generate a Top-K evidence set with high relevance. Specifically, the steps include:

[0076] The linear task sequence output in step 3 is traversed, where represents the th task node in the sequence. Each task node corresponds to a subtask to be processed, and its problem text is denoted as . This method performs two-way parallel recall to obtain a preliminary candidate document set

[0077] (1) Sparse retrieval: a BM25 retriever based on keyword matching is used to retrieve Top-K documents from a pre-constructed corpus with as the query text. This process can be formalized as:

[0078]

[0079] where is the number of candidate documents for sparse retrieval. ​

[0080] (2) Dense retrieval: adopt a dense retriever based on semantic vectors, which encodes into vectors , while all documents in the corpus have been pre-encoded into a set of vectors . By computing the cosine similarity of vectors, recall the Top-K most semantically relevant documents. This process can be formalized as:

[0081]

[0082] where is the number of candidate documents for dense retrieval hyperparameter.

[0083] Take as the base set, then traverse each document in , if its unique identifier (the combination of title and text snippet) does not appear in , add it to , merge and deduplicate the sparse retrieval results and dense retrieval results to form a comprehensive candidate document set .

[0084] Use the cross encoder to rearrange the merged candidate document set again. The cross encoder receives the question text and each candidate document as input pairs, calculates the fine-grained relevance score between them. This rearrangement process can be represented as:

[0085]

[0086] According to the rearrangement score , sort in descending order, and select the Top-K documents with the highest scores as the final evidence set of this subtask , as the input of the subsequent answer extraction and reasoning process.

[0087] In step 5, the bottom-up reasoning and answer extraction of each task according to the task sequence is performed, and the direct extraction and aggregated reasoning strategies are executed in parallel for non-leaf nodes, and the optimal answer is selected by comparing the confidence, and finally the complete answer of the question tree is obtained, which specifically includes:

[0088] Use the linear task sequence generated in step 3 , and according to the postorder traversal order of the sequence, process each task node from bottom to top.

[0089] For leaf nodes in the sequence (i.e., nodes without any subtask dependency), the leaf node’s question text and its Top-K evidence set rearranged by the cross encoder in step 4 are input to the LLM. The LLM directly extracts the answer based on the information in and computes its average log probability as the confidence score . This process can be represented as:

[0090] The answer and its confidence score are stored for later reference by the upper-level tasks.

[0091] For non-leaf nodes in the sequence with cross-task dependency, the answer extraction process employs a two-strategy parallel execution. The non-leaf node’s question text is denoted as and its list of subtask IDs is . Due to the bottom-up nature of the sequence, the answers of all its subtasks have already been generated and stored.

[0092] (1) Direct extraction strategy: This strategy ignores the answers of the subtasks and only relies on the current non-leaf node’s own evidence set . The pair is input to the LLM, and the answer is directly extracted. The average log probability is also computed as the confidence score .

[0093] (2) Aggregated reasoning strategy: The answers of the subtasks are used as structured context to guide the LLM to perform aggregated reasoning. The stored subtask answers are used to fill the placeholder references in , resulting in the filled question text . Meanwhile, a structured context containing all the subtask-question-answer pairs is constructed. The pair is input to the LLM, and the answer is generated through aggregated reasoning. The confidence score is also computed. This process can be formalized as:

[0094] The answer ​​An effectiveness check is performed. If there is obvious negative or invalid information in the answers of the sub-tasks in the aggregation context, it is directly determined that the aggregation result is invalid, and the result of the direct extraction strategy is selected.

[0095] Under the premise that the aggregation result is valid, the scores of the two strategies are compared and Since the aggregation reasoning result integrates multi-step information and is usually more reliable, when the scores are similar, this strategy realizes the preferential selection of the aggregation answer by slightly punishing .

[0096]

[0097] wherein, is a penalty factor.

[0098] The selected optimal answer and its score are stored. This iteration is performed until the entire task sequence is traversed, and the final root node answer is the complete solution to the original question.

[0099] Embodiment 3

[0100] In this embodiment, a complex question answering method based on adaptive task decomposition and multi-modal evidence aggregation is provided. It mainly includes a dynamic Few-shot prompt construction method based on dependency syntax fingerprints, a dynamic question decomposition method based on confidence evaluation and self-reflection, a question tree context dependency specification and task sequence generation method, an evidence generation method based on double-recall and cross-encoder rearrangement, and a question answer extraction method based on double-strategy aggregation reasoning.

[0101] The dynamic Few-shot prompt construction method based on dependency syntax fingerprints first constructs a multi-hop question disassembly example library through artificial annotation, then extracts the core syntax structure features of the question from the aspects of subject-predicate-object relationship, modifier structure and question type based on dependency syntax analysis technology to form syntax fingerprints. Thirdly, the syntax fingerprints of the new input original multi-hop question are generated and matched with the constructed feature library, and the K most similar disassembly examples to the new question structure are dynamically retrieved and selected. Further, the matched examples are spliced with fixed templates to construct complete Few-shot prompt templates, which are used as the input of the subsequent LLM to perform the question decomposition task.

[0102] Based on the confidence evaluation and self-reflection question dynamic deconstruction method, the dynamic Few-shot prompt template generated in step 1 is used to drive the LLM to perform recursive task decomposition on the original question, forming a preliminary question tree. On this basis, the self-reflection and reconstruction mechanism is introduced to ensure the decomposition quality: first, the average log probability (logprob) of the sub-questions in each question group is calculated as the confidence score of the question group, and the decomposition results are also subjected to standardization check. When the confidence of any question group is lower than the preset threshold or a violation is detected, the reconstruction process will be triggered, taking the current low-quality question tree as an example, and driving the LLM to perform task decomposition again with the reflection prompt until a stable and logical question tree is generated.

[0103] The question tree context dependency specification and task sequence generation method part processes the question tree generated in step 2. First, the referential placeholders in the text of each node question in the tree are subjected to multi-level normalization processing to precisely define the context dependency relationship between sub-tasks. On this basis, a post-order traversal algorithm is designed to convert the hierarchical tree structure into a linear task execution sequence. This linear sequence ensures that all sub-tasks are processed before their dependent parent tasks, providing a strict execution order and logical basis for the subsequent evidence aggregation-based reasoning process.

[0104] The evidence generation method part based on dual-path recall and cross-encoder rearrangement first uses a semantic vector-based dense retriever and a keyword-based sparse retriever to perform parallel candidate document recall for each sub-task in the linear task sequence generated in step 3. Second, a cross-encoder (Cross-Encoder) is used to rearrange the candidate documents merged from the two retrieval results, improving the relevance of the candidate paragraphs to the sub-task. Finally, the Top-K documents after rearrangement are taken as the initial evidence for the sub-task, forming the evidence set for answer extraction.

[0105] The question answer extraction method part based on dual-strategy aggregation reasoning performs knowledge fusion and answer extraction for each task from bottom to top according to the post-order execution sequence of the task sequence. For leaf nodes without sub-task dependency, the high-precision evidence set generated in step 4 is directly used to extract answers through LLM. For non-leaf nodes with cross-task dependency, two answer generation strategies are executed in parallel. The direct extraction strategy: the retrieval evidence set of the non-leaf node itself is input into LLM to directly extract answers; the aggregation reasoning strategy: the non-leaf node is taken as a "sub-question-answer" pair of the question group where the parent question is located as structured context, and the node question is input into LLM for reasoning and aggregation. By comparing the answer confidence scores generated by the two strategies and combining effectiveness verification, the optimal answer is selected, and this iteration is repeated until the answer extraction of the root node is completed, finally obtaining the complete answer of the entire question tree.

[0106] In combination with theFigure 1 As shown, the method specifically includes the following flow steps:

[0107] Step 1: Dynamic Few-shot prompt construction based on dependency syntax fingerprints. This embodiment constructs a multi-hop question disassembly example library based on manual annotation, extracts core syntax fingerprints based on dependency syntax technology, and finally dynamically matches and generates a Few-shot prompt template for new input questions.

[0108] The step 1 is dynamic Few-shot prompt construction based on dependency syntax fingerprints.

[0109] A multi-hop question disassembly example library is constructed by manual annotation .Each example in the library contains an original question and its corresponding decomposition sequence , that is . .

[0110] For each question in the library , based on dependency syntax analysis technology, its core syntax structure features are extracted to construct a syntax fingerprint , is a tuple containing multiple features, defined as follows:

[0111]

[0112] Wherein, the definition and role of each feature dimension are shown in Table 1:

[0113] Table 1 Definition and role of syntax fingerprint feature dimensions

[0114]

[0115] For a new input original multi-hop question , the same method as the above step is performed to generate its corresponding syntax fingerprint .

[0116] A similarity function is constructed to measure the similarity between and any syntax fingerprint in the example library . The similarity function is calculated by the weighted sum of each dimension feature:

[0117]

[0118] Wherein, represents the feature dimension, is the weight of each dimension, This is the similarity measurement function for the corresponding dimension. For set-type features, The function is Jaccard similarity. For path features, a customized path similarity function is used. ,in, Represents the set of paths to the new problem; Represents the set of paths for the example problem. It calculates two paths. and A function of Jaccard similarity between them; and These are the sizes of the two path sets, respectively.

[0119] Based on similarity function For example libraries Perform dynamic retrieval and select the K example sets with the highest similarity scores. , where K is a preset integer.

[0120] The design includes a fixed prompt template containing instructions to guide the LLM in performing tasks. and define and The structured format.

[0121] The retrieved set of K similar decomposition examples ,according to and The format is the same as the template. The pieces are then assembled. Ultimately, this addresses the new problem. by and The format is appended to the end to form a complete Few-shot prompt. As input for subsequent LLM execution problem deconstruction tasks.

[0122] Step 2: Dynamic Problem Deconstruction Based on Confidence Assessment and Self-Reflection. This embodiment designs a dynamic problem deconstruction method based on confidence assessment and self-reflection. It utilizes LLM to recursively decompose complex multi-hop problems, quantitatively assesses the confidence of the deconstruction results, and integrates an iterative self-correction mechanism to achieve precise control and optimization of the problem deconstruction quality. The implementation process is shown in the appendix. Figure 2 As shown.

[0123] Step 2 is based on confidence assessment and self-reflection-based dynamic problem deconstruction.

[0124] Using the dynamic Few-shot prompt generated in step 1 Driving LLM, addressing a series of primal problems Perform initial recursive deconstruction to obtain an initial set of question trees .

[0125] The initial deconstruction process converts each original question into a structured JSON object in the form of . At the same time, record the text response generated by the LLM and its corresponding logprobs sequence.

[0126] Perform automated quality assessment on the initial deconstruction results generated by the LLM. This assessment is performed independently for each question group in the question tree (i.e., a parent question and its directly decomposed child question list), including confidence evaluation and specification verification.

[0127] Design a self-reflection and reconstruction mechanism to correct low-quality initial deconstruction results. This mechanism is based on the following logic for iterative cycles.

[0128] Step 3: Question tree context dependency specification and task sequence generation method construction. This embodiment designs a question tree context dependency specification and task sequence generation method to post-process the question tree generated in step 2, ensuring that the logical relationship between each sub-task is precisely defined, and finally outputting a linear task sequence that can be executed subsequently.

[0129] The step 3 question tree context dependency specification and task sequence generation method construction.

[0130] Identify and parse the reference placeholders in the question text of each node in the question tree that represent context dependency relationships. The placeholders use a specific format, i.e., a number of # symbols followed by a number (e.g., #1, ##2). The parsing logic for this format is as follows:

[0131] (1) The meaning of the number of # symbols in the placeholder: the number of # symbols indicates the reference level.

[0132] 1) When the number of # is equal to the depth of the current question node in the question tree, the placeholder refers to a task at the same level, i.e., a task in the list of sub-tasks decomposed by the parent question of the current question.

[0133] 2) When the number of # is less than the depth of the current question node, the placeholder refers to an upper-level task, i.e., its parent question as a sub-task, corresponding to a certain task in the upper-level question group.

[0134] (2) The meaning of the number in the placeholder: the number part represents the position index of the question in the specified level task list, counting from 1.

[0135] The breadth-first search method is used to hierarchically process the problem tree, and in the traversal process, the referential placeholders in the text of each sub-problem are identified correction. The correction process follows the following double logic:

[0136] (1) Inheritance of reference judgment: when processing any problem node, first identify and extract all referential placeholders and their local text context before and after from the parent problem text, and construct a reference set containing (placeholder, pre and post text) tuples. Then, traverse the current sub-problem text, identify the placeholders in it, and extract their corresponding local text context. If the local text context of the placeholder in the sub-problem matches the pre and post text of any tuple in the reference set, it is determined that the placeholder is directly inherited from the parent problem. At this time, the number of placeholders in the sub-problem is corrected to be exactly the same as the corresponding placeholder in the parent problem, ensuring the stability and logical correctness of cross-level references.

[0137] (2) New generation of reference correction: if the placeholder is not inherited from the parent problem, it is considered as a newly generated reference at the current level. At this time, according to the depth d (root node depth is 0) of the current problem node, the number of placeholders is uniformly adjusted to d+1, ensuring that the level depth represented by the placeholder matches the level depth of the current sub-problem, eliminating the ambiguity of references caused by level nesting.

[0138] Convert the standardized problem tree data structure generated by the above steps into a decomposition map convenient for traversal , the structure is . The map takes each parent problem node as the key and its direct sub-problem list as the value. Remove any self-references that may cause cycles (i.e. the case where the sub-problem is exactly the same as its parent problem), to ensure the correctness of the traversal process.

[0139] Post-order traversal and serialization: use the post-order traversal algorithm to perform a depth-first traversal of the problem tree. The algorithm starts from the root problem , recursively visits all nodes in the order of .

[0140] In the traversal process, each visited problem node and its direct parent node form a tuple , and is appended to the linear list in turn. to strictly follow the context-dependent relationship of the preliminary task sequence, in which all sub-tasks appear before their parent tasks.

[0141] Construct metadata for each task node in the sequence . First, according to traversal order in the graph, assign an incremental unique integer ID to each task node, and build an ID mapping .

[0142] Parse the question text of each task node, and according to the aforementioned placeholder resolution rules, map the referential placeholders in it to the corresponding task IDs, and construct a context-dependent ID list .

[0143] According to get the direct subtask list of the current task and map it to the subtask ID list .

[0144] Encapsulate information including unique ID, normalized question text, context-dependent ID list and subtask ID list , etc. into metadata, and collect into the final linearized task sequence , which not only ensures the execution order of the tasks, but also provides complete context and dependency information for each task, providing accurate input for the subsequent evidence generation and aggregation reasoning process.

[0145] Step 4: Construction of evidence generation method based on dual recall and cross-encoder rearrangement. This embodiment designs an evidence generation method based on dual recall and cross-encoder rearrangement to provide a high-relevance evidence set for the linearized subtask sequence generated in step 3. This method generates a high-quality and comprehensive evidence set by performing parallel sparse and dense retrieval, and combining a fine rearrangement model.

[0146] The step 4 is based on the evidence generation method of dual recall and cross-encoder rearrangement.

[0147] traverse the linear task sequence output by step 3 , where represents the th task node in the sequence. Each task node corresponds to a subtask to be processed, and its question text is denoted as . This method performs dual parallel recall to obtain a preliminary candidate document set

[0148] (1) Sparse retrieval: use the BM25 retriever based on keyword matching to query from the pre-constructed corpus Top-K documents. This process can be formalized as: , where is the number of candidate documents for sparse retrieval.

[0149] (2) Dense retrieval: A semantic vector-based dense retrieval system is used to... Encode as a vector Meanwhile, all documents in the corpus have been pre-encoded as vector sets. By calculating the cosine similarity of vectors, the top-K semantically most relevant documents are retrieved. This process can be formalized as follows: ,in, This is a hyperparameter for the number of candidate documents in dense retrieval.

[0150] Will Use it as the base set, and then iterate through it. Each document in the document, if its unique identifier (the combination of title and text fragment) is not in If it appears in the text, add it to the list. In the context of sparse search results With dense search results Merge and deduplicate documents to form a comprehensive candidate document set. .

[0151] Using a cross encoder to process the merged candidate document set Perform a second rearrangement. The cross encoder receives the problem text. With each candidate document As input pairs, calculate the fine correlation score between them. The rearrangement process can be represented as:

[0152] Based on rearrangement score right Sort the documents in descending order and select the top-K documents with the highest scores as this subtask. The final set of evidence This serves as the input for subsequent answer extraction and reasoning processes.

[0153] Step 5: Construction of a Question Answer Extraction Method Based on Dual-Strategy Aggregation Reasoning. This embodiment designs a question answer extraction method based on dual-strategy aggregation reasoning. The method flow is shown in the appendix. Figure 3 As shown, the method performs bottom-up reasoning on the task sequence and its evidence set generated in step 4 to obtain the complete answer to the original multi-hop problem. This method executes two strategies in parallel: direct extraction and aggregate reasoning, for non-leaf nodes with cross-task dependencies. It also dynamically selects the optimal answer through confidence evaluation and validity verification, ensuring the accuracy and robustness of the reasoning chain.

[0154] Step 5 is constructed based on the question answer extraction method of dual-strategy aggregation reasoning.

[0155] Using the linear task sequence generated in step 3 And following the post-order traversal of the sequence, perform a bottom-up process on each task node. Process it.

[0156] For leaf nodes in the sequence (i.e., nodes that have no subtask dependencies), the question text of that leaf node is... and the Top-K evidence set rearranged by the cross encoder in step 4. Input to LLM. LLM is based on The strategy of directly extracting information to generate answers And calculate its average log probability as the confidence score. This process can be represented as:

[0157] This answer and its confidence score Store it for later use by higher-level tasks.

[0158] For non-leaf nodes in the sequence that have cross-task dependencies, the answer extraction process employs a dual-strategy parallel execution. The question text for this non-leaf node is denoted as... Its subtask ID list is Due to the bottom-up nature of the sequence, all its subtasks (in ) All have been generated and stored.

[0159] (1) Direct extraction strategy: This strategy ignores the answers to subtasks and relies only on the evidence set of the current non-leaf node itself. .Will and Enter LLM to directly extract the answer. Similarly, the average log probability is calculated as the confidence score. .

[0160] (2) Aggregate Reasoning Strategy: Using the answers to subtasks as structured context, the LLM is guided to perform aggregate reasoning. The stored subtask answers are utilized. filling The placeholder in the text is used to obtain the filled problem text. Simultaneously, a structured context containing question-answer pairs for all subtasks is constructed. .Will and Input LLM together, perform reasoning aggregation, and generate the answer. And calculate its confidence score. This process can be formalized as follows:

[0161] The answer generated by the aggregation reasoning An effectiveness check is performed. If there is obvious negative or invalid information in the sub-task answers in the aggregation context, the aggregation result is directly determined to be invalid, and the result of the direct extraction strategy is selected.

[0162] Under the premise that the aggregation result is valid, the scores of the two strategies are compared and Since the aggregation reasoning result integrates multi-step information, it is usually more reliable, so when the scores are similar, this strategy realizes the preferential selection of the aggregation answer by slightly punishing .

[0163]

[0164] wherein, is a penalty factor.

[0165] The selected optimal answer and its score are stored. This iteration is performed until the entire task sequence is traversed, and the final root node answer is the complete solution to the original question.

[0166] Embodiment 4

[0167] Embodiment 4 provides a non-transitory computer readable storage medium for storing computer instructions, which, when executed by a processor, implement the complex question answering method based on adaptive task decomposition and multi-modal evidence aggregation as described above.

[0168] Embodiment 5

[0169] Embodiment 5 provides a computer device comprising a memory and a processor, the processor and the memory being in communication with each other, the memory storing program instructions executable by the processor, and the processor invoking the program instructions to execute the complex question answering method based on adaptive task decomposition and multi-modal evidence aggregation as described above.

[0170] Embodiment 6

[0171] Embodiment 6 provides an electronic device comprising a processor, a memory, and a computer program; wherein the processor is connected with the memory, and the computer program is stored in the memory; when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes the instructions for implementing the complex question answering method based on adaptive task decomposition and multi-modal evidence aggregation as described above.

[0172] The above describes the specific embodiments of the present application in combination with the drawings, but is not a limitation on the protection scope of the present application. Those skilled in the art should understand that various modifications or variations made by those skilled in the art on the basis of the disclosed technical solutions of the present application, without creative labor, should be covered in the protection scope of the present application.

Claims

1. A complex question answering method based on adaptive task deconstruction and multi-modal evidence aggregation, characterized in that, Comprising: Based on the dependency syntax analysis technology to extract the syntax fingerprint of the question and match with the pre-built multi-hop question disassembly example library, combine the Few-shot example matched with the fixed template to generate the question deconstruction prompt template; Based on the dynamic Few-shot prompt to drive the LLM to recursively deconstruct the question, form a preliminary question tree, trigger the reconstruction mechanism through confidence evaluation and normative check for low-quality question trees, until a logically rigorous question tree is generated; Post-processing the question tree to standardize the reference relationship between nodes to accurately define the sub-task dependency, and converting the question tree into a linear task execution sequence using a post-order traversal algorithm; Using dense and sparse retriever to perform parallel double recall on each sub-task in the task sequence, and using a cross-encoder to perform secondary rearrangement on the merged candidate documents to generate a Top-K evidence set with high relevance; According to the task sequence, infer and extract the answer from each task from bottom to top, for non-leaf nodes, perform direct extraction and aggregated reasoning strategies in parallel and select the optimal answer by comparing the confidence, and finally obtain the complete answer of the question tree.

2. The complex question answering method based on adaptive task deconstruction and multi-modal evidence aggregation according to claim 1, characterized in that, The problem's syntactic fingerprint is extracted using dependency parsing and matched with a pre-built multi-hop problem decomposition example library. The matched few-shot examples are then combined with a fixed template to generate a problem decomposition hint template. Specifically, this involves: constructing a multi-hop problem decomposition example library through manual annotation, where each example contains an original problem and its corresponding decomposition sequence; parsing each original problem using dependency parsing to extract its core syntactic structural features and constructing a syntactic fingerprint, which is a tuple containing multi-dimensional features; generating a corresponding syntactic fingerprint for each new input original multi-hop problem; constructing a similarity function to measure the similarity between the syntactic fingerprint of a new input original multi-hop problem and any syntactic fingerprint in the example library; dynamically searching the example library based on the similarity function and selecting the K sets of examples with the highest similarity scores; and then further refining the retrieved K sets of similar decomposition examples according to... and The format is the same as the template. To piece together; to address new problems by and The format is appended to the end to form a complete Few-shot prompt. As input for subsequent LLM execution problem deconstruction tasks.

3. The complex question answering method based on adaptive task decomposition and multi-modal evidence aggregation according to claim 2, characterized in that, Based on dynamic Few-shot prompt driving LLM to recursively deconstruct the problem, form a preliminary problem tree, through confidence evaluation and normative check to trigger the reconstruction mechanism for low-quality problem trees, until a logically rigorous problem tree is generated, including: using the generated dynamic Few-shot prompt to drive LLM to preliminarily recursively deconstruct a series of original problems to obtain an initial problem tree set ; the preliminary deconstruction process converts each original problem into a structured JSON object in the form of ; at the same time, records the text response generated by LLM and its corresponding word token probability sequence; automatically evaluates the quality of the preliminary deconstruction results generated by LLM, which is independently executed for each problem group in the problem tree.

4. The complex question answering method based on adaptive task decomposition and multi-modal evidence aggregation according to claim 3, characterized in that, Post-processing is performed on the problem tree to standardize the referential relationships between nodes to precisely define subtask dependencies. A post-order traversal algorithm is then used to transform the problem tree into a linear task execution sequence. Specifically, this includes: identifying and parsing the referential placeholders used to represent contextual dependencies in the problem text of each node in the problem tree. The placeholders are represented in a specific format: several # symbols followed by a number. The parsing logic for this format is as follows: The number of # symbols in the placeholder indicates the reference level; the number in the placeholder represents the index of the problem's location in the task list at the specified level, counting from 1; a breadth-first traversal method is used to hierarchically process the problem tree, and the placeholders in each sub-problem text are checked during the traversal. Perform corrections.

5. The complex question answering method based on adaptive task decomposition and multi-modal evidence aggregation according to claim 4, characterized in that, A parallel dual-path recall process is employed, using both dense and sparse retrieval algorithms to retrieve each subtask from the task sequence. A cross-encoder is then used to perform a secondary reordering of the merged candidate documents, generating a highly relevant Top-K evidence set. Specifically, this includes processing the output linear task sequence... Perform a traversal, where Represents the first in the sequence Each task node For each subtask to be processed, the problem text is denoted as... Perform a dual-path parallel recall to obtain an initial candidate document set. Sparse retrieval: Employs the BM25 search engine based on keyword matching. To query text, from a pre-built corpus Recall the Top-K documents; Dense retrieval: Employ a semantic vector-based dense retrieval system to retrieve the top-K documents. Encode as a vector Meanwhile, all documents in the corpus have been pre-encoded as vector sets. By calculating the cosine similarity of vectors, the top-K semantically most relevant documents are retrieved; a cross-encoder is used to process the merged candidate document set. Perform secondary rearrangement; cross encoder receives problem text. With each candidate document As input pairs, calculate the fine correlation score between them. Based on the rearranged score right Sort the documents in descending order and select the top-K documents with the highest scores as this subtask. The final set of evidence This serves as the input for subsequent answer extraction and reasoning processes.

6. The complex question answering method based on adaptive task deconstruction and multi-modal evidence aggregation according to claim 5, characterized in that, According to the task sequence, each task is inferred and answer is extracted from bottom to top. For non-leaf nodes, two strategies of direct extraction and aggregated reasoning are executed in parallel, and the optimal answer is selected by comparing the confidence. Finally, the complete answer of the question tree is obtained, which specifically includes: using the generated linear task sequence , and according to the postorder traversal order of the sequence, each task node is processed from bottom to top; for the leaf nodes in the sequence, the question text of the leaf node and its Top-K evidence set rearranged by the cross encoder in step 4 are input to the LLM; the LLM generates an answer based on the information in direct extraction strategy , and calculates the average logarithmic probability as the confidence score ; for non-leaf nodes with cross-task dependencies in the sequence, the answer extraction process adopts a double-strategy parallel execution; the question text of this non-leaf node is recorded as , and the list of its subtask IDs is ; due to the bottom-up nature of the sequence, the answers of all its subtasks have been generated and stored.

7. A complex question answering system based on adaptive task deconstruction and multimodal evidence aggregation, characterized in that, Comprising: A matching module for extracting the syntax fingerprint of the question based on the dependency syntax analysis technology and matching with the pre-built multi-hop question disassembly example library, combining the Few-shot example matched with the fixed template to generate the question deconstruction prompt template; A deconstruction module for recursively deconstructing the question based on the dynamic Few-shot prompt to drive the LLM, forming a preliminary question tree, triggering the reconstruction mechanism through confidence evaluation and normative check for low-quality question trees, until a logically rigorous question tree is generated; A conversion module for post-processing the question tree to standardize the reference relationship between nodes to accurately define the sub-task dependency, and converting the question tree into a linear task execution sequence using a post-order traversal algorithm; A recall module for using dense and sparse retriever to perform parallel double recall on each sub-task in the task sequence, and using a cross-encoder to perform secondary rearrangement on the merged candidate documents to generate a Top-K evidence set with high relevance; An extraction and aggregation module for inferring and extracting the answer from each task from bottom to top according to the task sequence, for non-leaf nodes, performing direct extraction and aggregated reasoning strategies in parallel and selecting the optimal answer by comparing the confidence, and finally obtaining the complete answer of the question tree.

8. A non-transitory computer-readable storage medium, comprising: The non-transitory computer readable storage medium is used to store computer instructions, which are executed by the processor to implement the complex question and answer method based on adaptive task deconstruction and multi-modal evidence aggregation of any one of claims 1-6.

9. A computer device, comprising: Comprising a memory and a processor, the processor and the memory communicate with each other, the memory stores program instructions executable by the processor, and the processor invokes the program instructions to execute the complex question and answer method based on adaptive task deconstruction and multi-modal evidence aggregation of any one of claims 1-6.

10. An electronic device, comprising: Comprising: A processor, a memory and a computer program; wherein the processor is connected with the memory, and the computer program is stored in the memory; when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes the instructions for realizing the method for complex question answering based on adaptive task decomposition and multi-modal evidence aggregation according to any one of claims 1-6.

Images