A Retrieval Enhancement Method and System Based on Neural Symbol Collaboration
By transforming text into a symbolic knowledge structure through a neural symbolic collaborative architecture and combining it with neural semantic embedding to construct a hybrid index, the efficiency and accuracy issues of existing information retrieval systems in complex semantic queries are solved. This enables multi-hop reasoning and dynamic sorting, thereby improving retrieval performance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA NANHU ACAD OF ELECTRONICS & INFORMATION TECH
- Filing Date
- 2026-01-22
- Publication Date
- 2026-05-26
AI Technical Summary
Existing information retrieval systems lack an efficient and unified index structure that combines symbolic knowledge and neural representation when processing complex semantic queries. This makes it difficult to perform multi-hop relational reasoning and result ranking without context awareness, resulting in insufficient retrieval efficiency and accuracy.
By using a neuro-symbolic collaborative architecture, text is transformed into a symbolic knowledge structure and jointly modeled with neuro-semantic embeddings to construct a hybrid memory index. A large language model is then used for query semantic parsing and multi-hop information propagation to achieve explicit logical reasoning and dynamic sorting.
It improves the retrieval accuracy and interpretability of complex semantic queries, maintains the efficient semantic representation capability of neural networks, and enhances the contextual relevance of retrieval results by utilizing symbolic structures for multi-hop reasoning and dynamic propagation.
Smart Images
Figure CN122087076A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence, and specifically to a retrieval enhancement method and system based on neural symbol collaboration. Background Technology
[0002] With the rapid development of artificial intelligence technology, information retrieval systems have been widely used in fields such as question answering, recommendation, and knowledge management. Existing retrieval systems mainly include keyword-based retrieval methods and semantic-based retrieval methods.
[0003] Keyword-based retrieval methods typically rely on inverted index structures, returning results by matching user-input keywords with terms in documents. While these methods are simple to implement and fast, they cannot effectively understand semantic relationships in natural language and struggle to handle complex semantic issues such as synonyms and contextual dependencies.
[0004] Semantic-based retrieval methods generate text representations using word vectors, sentence vectors, or large language models (LMs) to capture the latent semantic similarity between queries and documents. While these methods significantly improve semantic matching capabilities, they still suffer from two main problems: first, pure neural network models often lack interpretability and logical reasoning capabilities, making it difficult to handle complex reasoning queries; second, the computational cost of these models is high, hindering large-scale, real-time retrieval.
[0005] In recent years, neuro-symbolic AI has become an important direction for improving the reasoning capabilities of intelligent systems. This method combines the representation learning ability of neural networks with the logical reasoning ability of symbolic systems, achieving synergy between knowledge representation and reasoning. However, existing neuro-symbolic systems still face several challenges in information retrieval scenarios:
[0006] The lack of an efficient and unified hybrid index structure between symbolic knowledge and neural representation limits retrieval efficiency; query semantic parsing remains at the shallow entity extraction and matching stage, making it difficult to handle complex logical queries and multi-hop relational reasoning; the retrieval result ranking mechanism lacks context awareness and dynamic propagation capabilities, and cannot fully utilize the complementary information of symbolic relations and semantic similarity.
[0007] Therefore, there is an urgent need for a new retrieval enhancement method that can establish a collaborative mechanism between symbolic knowledge and neurosemantics, which can maintain the efficient semantic representation capability of neural networks and realize explicit logical reasoning using symbolic structures, thereby achieving high-quality retrieval and reasoning in an open-domain knowledge environment. Summary of the Invention
[0008] To address the problems of existing technologies, this invention proposes a retrieval enhancement method and system based on neural symbol collaboration. By introducing a neural symbol collaboration architecture, textual knowledge is transformed into a symbolic knowledge structure, and joint modeling is performed with neural semantic embedding to form a hybrid memory index that can be used for efficient reasoning and retrieval.
[0009] The purpose of this invention is to overcome the problems of insufficient semantic understanding, limited logical reasoning ability, and lack of contextual association in retrieval ranking in existing semantic retrieval methods, and to propose a retrieval enhancement method based on neural symbol collaboration. This method can achieve multi-hop reasoning and dynamic propagation by utilizing the structured logical relationships of the symbol layer while maintaining the semantic representation ability of the neural network, significantly improving the retrieval accuracy and interpretability in complex semantic query scenarios.
[0010] To achieve the above objectives, the present invention provides a retrieval enhancement method based on neural symbol collaboration, comprising the following steps:
[0011] S1. Transform text into symbolic knowledge through a neural symbolic collaborative architecture to construct a hybrid memory index;
[0012] S2. Query semantic parsing and constructing a symbolic query structure;
[0013] S3. Perform a search, sort the results, and output the search results.
[0014] Preferably, the neural symbol co-architecture in step S1 includes a symbol extraction module, a symbol map construction module, and a neural association module.
[0015] Preferably, in the symbol extraction module, open-domain triples are extracted from the text using a large language model to form symbolic knowledge units; in the symbolic graph construction module, the symbolic knowledge units are constructed into a knowledge graph structure; in the neural association module, the similarity between entity nodes is calculated using an embedding model, and supplementary similarity relationship edges are generated based on the similarity to form a neural association layer, which is then merged to form an indexable graph structure to construct a hybrid memory index.
[0016] Preferably, the query semantic parsing and symbolic query structure construction in step S2 is achieved by extracting query entities and their semantic relationships from natural language queries to generate a symbolic query structure.
[0017] Preferably, for complex queries, the inductive reasoning ability of the large language model is used to decompose them into sub-problems that can be reasoned about, so as to clearly describe semantic relationships.
[0018] Preferably, in step S3, the similarity between the query entity and the node is calculated, the query entity is mapped to the graph structure node, the initial activated node is determined, and the context-based multi-hop information propagation is performed based on the symbolic relationship and similarity relationship between the nodes to expand the candidate set, and the candidate set is sorted by probability or score.
[0019] Preferably, the context-based multi-hop information propagation is implemented through personalized PageRank, weighted graph propagation, or other multi-hop propagation algorithms, and is used to calculate the node activation probability or propagation score on the graph structure.
[0020] Compared with existing technologies, the retrieval enhancement method based on neural symbol collaboration of the present invention has the following beneficial effects:
[0021] 1) By unifying structured symbolic knowledge and neurosemantic embedding into the same index structure, the richness of semantic expression is maintained while the interpretability of explicit logical relationships is also achieved;
[0022] 2) Introducing a semantic decomposition mechanism based on a large language model can decompose complex natural language queries into sub-problems that can be symbolically reasoned, thereby improving the ability to handle multi-relation and multi-hop queries.
[0023] 3) By executing a context-based multi-hop propagation algorithm in the hybrid index graph, dynamic reasoning with query context awareness is achieved, making the result sorting more consistent with semantic association and knowledge logic.
[0024] Another aspect of the present invention provides a retrieval enhancement system based on neural symbol collaboration, including a processor and a memory storing a plurality of computer instructions, characterized in that the computer instructions, when executed by the processor, implement the steps of the retrieval enhancement method based on neural symbol collaboration. Attached Figure Description
[0025] Figure 1 This is a schematic diagram of the retrieval enhancement method based on neural symbol collaboration according to an embodiment of the present invention;
[0026] Figure 2 This is a schematic diagram of neural symbol retrieval and dynamic reasoning in an embodiment of the present invention. Detailed Implementation
[0027] Example 1
[0028] like Figure 1 As shown, this embodiment provides a retrieval enhancement method based on neural symbol collaboration, including the following steps:
[0029] S1. Text symbolization and hybrid memory index construction.
[0030] Open-domain unstructured text is expected to be transformed into symbolic knowledge units, and a hybrid memory index containing neural association layers is constructed.
[0031] Step S1 specifically includes the following steps:
[0032] S11. Perform the symbol extraction step to extract symbol-level knowledge units from the original text.
[0033] In practical implementation, large language models (such as Qwen and DeepSeek) can be used to perform open and information extraction tasks, identifying and outputting knowledge representations in the form of triples (subject, predicate, object) from the input corpus.
[0034] For example, you could type: "Einstein proposed the theory of relativity."
[0035] The possible triplet is: (Einstein, proposed the theory of relativity).
[0036] Each triple is treated as a symbolic knowledge unit and stored in a structured knowledge table.
[0037] S12. Perform the symbolic graph construction steps to organize symbolic knowledge units into a graph structure.
[0038] After knowledge extraction is completed in the preceding steps, the discrete symbolic knowledge units need to be organized into a computable and reasonable structured form. The extracted entities, paragraphs, facts, and various edges (facts, synonyms, paragraphs, etc.) are mapped into a graph structure. The storage method of the generated graph structure can be selected according to the task requirements: in the case of structured query and complex relational reasoning scenarios, a graph database can be used for storage and indexing; in the case of tasks that require high-dimensional representation, such as semantic similarity calculation and graph vector retrieval, the graph nodes and edges can be represented as vectorized structures and stored in a graph vector library.
[0039] S13. Perform the neural connection step and supplement similar edges.
[0040] Although symbolic graphs possess explicit relational structures, semantic drift still exists between different expressions (such as "relativity" and "special relativity") in open-domain scenarios.
[0041] Therefore, a semantic encoding of nodes is performed using an embedding model, and the cosine similarity between nodes is calculated. When the similarity exceeds a preset threshold, similarity edges are added to the graph to generate a supplementary set of relationships, forming a neural association layer. This layer is embedded in the symbolic graph, constituting the final neural symbolic hybrid index structure. This index structure simultaneously supports symbolic-level logical jumps and semantic-level fuzzy associations, achieving an integrated representation of information through "explicit logic + implicit semantics".
[0042] S2. Query semantic parsing and construct symbol query structure.
[0043] It receives natural language queries, uses a large language model to identify named entities, intents, times, etc., and performs semantic parsing and structured representation of the query content.
[0044] For example: Search for: What is the capital of the country of the author of the Harry Potter series?
[0045] Retrieve query entities: ["Harry Potter", "Author", "Country", "Capital"]
[0046] Query structure: text embedding of the query and entity text embedding.
[0047] S3, Neural Symbol Retrieval and Dynamic Reasoning.
[0048] S31. Parallel execution of Dense Paragraph Retrieval (DPR) and direct matching process with Knowledge Graph (KG).
[0049] Dense Paragraph Retrieval (DPR) retrieves the top-k paragraphs from a vector database by generating vector embeddings for the query or its sub-items, thus obtaining the DPR_score. Knowledge Graph (KG) direct matching matches the query entity with KG nodes, obtaining a list of matching nodes and symbol confidence.
[0050] S32. Merge the recalled paragraphs and graph nodes into a candidate set, and assign an initial value to each node:
[0051]
[0052] in, This represents the normalized similarity score between the paragraph and the query. The confidence level of the symbol.
[0053] S33. On the graph, a personalized PageRank propagation algorithm is executed with the candidate set or query mapping nodes as the initial distribution. During the propagation process, the edge weights (symbolic relationship credibility, similarity) and the node reset probability (further normalized based on init_score) are used as the initial distribution of PPR.
[0054] S34. Score nodes / paragraphs based on post-propagation scores. Secondary fusion with the initial retrieval score:
[0055]
[0056] in, The weights of the graph propagation score. The larger the scale, the more emphasis is placed on the propagation of structural logic; It is the weight of the initial search score; It is a weighting factor that adds to the credibility of the source; Graph propagation score is the activation probability or importance score of a node after multiple hops in the graph structure; `provenance_boost` is a weighted coefficient related to the "source credibility" of a paragraph or knowledge node, used to adjust the importance of nodes in the final ranking score. The top-K paragraphs are taken as the results. If a natural language answer needs to be generated, these paragraphs are used as context input to the generation model to output the final answer.
[0057] Example 2
[0058] like Figure 2 As shown, this embodiment provides a retrieval and dynamic reasoning process based on neural symbols. To further improve the accuracy and interpretability of the neural symbol retrieval stage, the specific steps of neural symbol retrieval and dynamic reasoning are as follows:
[0059] S21. Analyze the set of facts in the knowledge graph. Each fact is usually represented as a triple:
[0060] Fact=(subject,Ration,Object)
[0061] The subject and object are collectively referred to as phrase nodes. For each phrase node, a weight is calculated based on the average score of the fact to which it belongs.
[0062]
[0063] in This indicates the number of facts containing the phrase. This represents the average semantic score of the fact. This step is used to measure the global importance of each phrase node in the current query context.
[0064] S22. Based on the weight values obtained in step S21, sort all phrase nodes and select the Top-K phrase nodes with the highest weights. These phrase nodes serve as "anchor nodes" for subsequent graph search, limiting the initial propagation range of inference and thus effectively controlling search complexity.
[0065] For each selected phrase node, a dense paragraph retrieval (DPR) operation is performed to retrieve the most semantically similar paragraphs from the knowledge corpus and calculate paragraph-level similarity scores. For each paragraph node, its weight is defined as follows:
[0066]
[0067] Where DPR_score(d) is the normalized similarity score between the paragraph and the query, and λpara is the paragraph node weight adjustment coefficient, which reflects the direct semantic relevance between the paragraph node and the query.
[0068] S23. Combine the phrase node weights and paragraph node weights to obtain the comprehensive "Reset Probability" for each node in the graph, which is defined as:
[0069]
[0070] in This is the fusion coefficient, typically around 0.5. This fusion weight is used as the initial probability distribution in the personalized PageRank algorithm. , where represents the probability that each node is selected as the starting point in a random walk.
[0071] S24. Based on the comprehensive reset probability, execute the personalized PageRank (PPR) propagation algorithm on the graph. The propagation update formula is:
[0072]
[0073] Where W is the adjacency matrix of the graph (after normalization). Let be the node distribution in the t-th iteration. Set the reset probability (e.g., 0.15). The algorithm iterates until convergence. Output the final importance score for each node. These scores represent the importance of the node in the propagation path related to the query semantics.
[0074] S25. Sort the paragraph nodes in descending order according to the PPR score, and output the Top-N paragraphs and their corresponding scores.
[0075] Example 3
[0076] A retrieval enhancement system based on neural symbol collaboration is provided, including a processor and a memory storing a number of computer instructions, which, when executed by the processor, implement the steps of the neuron encoding method for distributed brain-like simulation.
[0077] For specific limitations on retrieval enhancement systems based on neural symbol collaboration, please refer to the limitations of a retrieval enhancement method based on neural symbol collaboration mentioned above, which will not be repeated here.
[0078] The memory and processor are electrically connected directly or indirectly to enable data transmission or interaction. For example, these components can be electrically connected to each other via one or more communication buses or signal lines. The memory stores a computer program that can run on the processor, which implements the method in the embodiments of the present invention by running the computer program stored in the memory.
[0079] The memory may be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The memory stores the program, and the processor executes the program upon receiving an execution instruction.
[0080] The processor may be an integrated circuit chip with data processing capabilities. The aforementioned processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor.
[0081] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0082] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.
Claims
1. A retrieval enhancement method based on neural symbol collaboration, characterized in that, Includes the following steps: S1. Transform text into symbolic knowledge through a neural symbolic collaborative architecture to construct a hybrid memory index; S2. Query semantic parsing and constructing a symbolic query structure; S3. Perform a search, sort the results, and output the search results.
2. The method according to claim 1, characterized in that, The neural symbol co-architecture in step S1 includes symbol extraction, symbol map construction, and neural association.
3. The method according to claim 2, characterized in that, In the symbol extraction module, open-domain triples are extracted from the text using a large language model to form symbolic knowledge units. In the symbolic graph construction module, the symbolic knowledge units are constructed into a knowledge graph structure. In the neural association module, the similarity between entity nodes is calculated using an embedding model, and supplementary similarity relationship edges are generated based on the similarity to form a neural association layer. These layers are then merged to form an indexable graph structure, thus constructing a hybrid memory index.
4. The method according to claim 1, characterized in that, The query semantic parsing and symbolic query structure construction in step S2 is achieved by extracting query entities and their semantic relationships from natural language queries to generate a symbolic query structure.
5. The method according to claim 4, characterized in that, For complex queries, the inductive reasoning ability of the large language model is used to decompose them into sub-problems that can be reasoned about, so that semantic relationships can be clearly described.
6. The method according to claim 1, characterized in that, In step S3, the similarity between the query entity and the node is calculated, the query entity is mapped to the graph structure node, the initial activation node is determined, and the context-based multi-hop information propagation is performed based on the symbolic relationship and similarity relationship between the nodes to expand the candidate set, and the candidate set is sorted by probability or score.
7. The method according to claim 6, characterized in that, The context-based multi-hop information propagation is implemented through personalized PageRank, weighted graph propagation, or other multi-hop propagation algorithms, and is used to calculate the node activation probability or propagation score on the graph structure.
8. A retrieval enhancement method and system based on neural symbol collaboration, comprising a processor and a memory storing a plurality of computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of solving the dynamic programming problem based on neural symbol collaboration as described in any one of claims 1 to 7.