Biological knowledge graph question answering system based on multi-hop bridge reasoning

Abstract

The application relates to the field of knowledge graphs, and specifically discloses a biological knowledge graph question-answering system based on multi-hop bridge reasoning, which comprises the following modules: a question acquisition and entity recognition module, which is used for acquiring a biomedical question in a natural language form, inputting the biomedical question into a field language encoder for feature extraction, and identifying biomedical entity segments in the biomedical question and a predicted entity type corresponding to each biomedical entity segment through a sequence labeling algorithm; an entity disambiguation and linking module, which is used for constructing a candidate entity set according to the identified biomedical entity segments; a bridge entity retrieval module; a multi-hop reasoning path generation module; a candidate path sorting module; an evidence-enhanced answer generation module; and an answer consistency verification module. The technical scheme of the application can eliminate ambiguity, has high reasoning efficiency, avoids factitious hallucinations, and has result interpretability.

Description

Technical Field

[0001] This invention relates to the field of knowledge graphs, and in particular to a biological knowledge graph question-answering system based on multi-hop bridge reasoning. Background Technology

[0002] Biomedical knowledge graphs are used to organize and express entities such as genes, proteins, gene mutations, drugs, diseases, biological pathways, biological processes, phenotypes, cells, tissues, and species, as well as the relationships between entities, including coding, function, association, regulation, participation, homology, and influence. They are typically stored in structured triples of "head entity—relationship—tail entity." Question answering systems based on biological knowledge graphs can convert user-input natural language questions into knowledge retrieval or reasoning tasks, providing auxiliary information for biomedical research, drug mechanism analysis, disease risk association analysis, and cross-species functional comparisons.

[0003] Existing biomedical question-answering technologies mainly include keyword retrieval, fixed query templates, database queries, single-hop knowledge graph queries, pre-trained language model question answering, graph neural network reasoning, and retrieval enhancement generation. For questions such as "Which diseases is TP53 associated with?" and "What targets does metformin act on?", existing single-hop query technologies can obtain answers from the direct neighbors of the question entity.

[0004] However, for questions such as "which gene mutations lead to the loss of TP53 function and thus increase the risk of cancer", "what pathway does metformin use to lower blood sugar", or "why BRCA1 mutant patients may be sensitive to PARP inhibitors", there are usually intermediate entities such as gene mutations, proteins, targets, biological pathways or biological processes between the question entity and the answer entity. It is necessary to combine reasoning along multiple directional relationships, and it is difficult to obtain a complete answer by simply relying on direct neighbor queries.

[0005] The reason is that existing biomedical question-answering technologies mainly have the following problems:

[0006] First, entity recognition and entity linking are prone to ambiguity, leading to incorrect starting points for subsequent reasoning. The biomedical field contains standard names, historical names, aliases, abbreviations, and entities with the same name across species. The same text may also involve different types such as genes, proteins, drugs, diseases, and mutations. If knowledge graph entities are determined solely based on string similarity or precise name matching, it becomes difficult to comprehensively utilize the question context, entity type, species information, and graph structure relationships between other entities in the question. This can easily lead to confusion between genes and proteins, human genes and genes from other species, and incorrect linking of drug names. Therefore, even with subsequent multi-hop retrieval, incorrect entity links will cause the system to expand along incorrect knowledge graph nodes, failing to obtain reasoning results corresponding to the question.

[0007] Second, the number of candidate paths in multi-hop reasoning expands rapidly with the increase in reasoning layers and the number of entity neighbors. Existing methods that directly traverse the adjacent nodes of the knowledge graph generate significant computational costs and long response times, while also introducing a large number of entities and relationships irrelevant to the question's semantics, making it difficult to filter out the correct path from the candidate paths. On the other hand, selecting paths solely based on graph structure distance or single similarity cannot adequately determine whether the entity types in the candidate paths match the expected answer, whether the relationship direction is correct, whether the species information is compatible, and whether the factual sources and confidence levels are reliable. Consequently, it is easy to form reasoning chains with incorrect relationship directions, misuse of causal relationships, or incorrect cross-species transfer.

[0008] Third, existing generative question-answering technologies are prone to generating content lacking supporting knowledge evidence after completing multi-hop reasoning. Some systems only input the question text or rough search results into the language generation model, failing to explicitly constrain the knowledge triples, relational directions, species information, and factual sources corresponding to candidate reasoning paths. This leads to generated answers that may contain entities, relations, or conclusions not found in the knowledge graph, or that express relationships such as "related" or "participated" as deterministic causal relationships. Furthermore, existing systems typically only output natural language answers, lacking bridging entities, hop-by-hop relations, and knowledge sources, making it difficult for users to verify the answers.

[0009] Therefore, there is a need for a biological knowledge graph question-answering system based on multi-hop bridge reasoning that can eliminate ambiguity, has high reasoning efficiency, avoids factual illusions, and has interpretable results. Summary of the Invention

[0010] This invention provides a biological knowledge graph question-answering system based on multi-hop bridge reasoning, which can eliminate ambiguity, has high reasoning efficiency, avoids factual illusions, and has interpretable results.

[0011] To solve the above-mentioned technical problems, this application provides the following technical solution:

[0012] A biological knowledge graph question-answering system based on multi-hop bridge reasoning includes:

[0013] The question acquisition and entity recognition module is used to acquire biomedical questions in natural language form, input them into the domain language encoder for feature extraction, and identify biomedical entity fragments in the biomedical questions and the predicted entity type corresponding to each biomedical entity fragment through a sequence labeling algorithm.

[0014] The entity disambiguation and linking module is used to construct a candidate entity set based on the identified biomedical entity fragments, and to calculate the comprehensive similarity score of each candidate entity based on string similarity, contextual semantic similarity, matching degree between candidate entity type and predicted entity type, species information in the question, and global graph connectivity of the knowledge graph. The biomedical entity fragments are then mapped to the target standard entity with the highest score in the knowledge graph as the starting entity for reasoning.

[0015] The bridging entity retrieval module is used to perform a dual-channel retrieval in the knowledge graph based on the starting entity, utilizing a retrieval channel based on word matching and a semantic retrieval channel based on graph representation learning, to recall bridging entities that are semantically associated with the biomedical question.

[0016] The multi-hop reasoning path generation module is used to perform multi-step iterative expansion from the starting entity to the bridging entity in the knowledge graph, and execute security logic constraints including entity type constraints, relationship direction constraints, species compatibility constraints and fact confidence filtering at each stage of path expansion, so as to eliminate biologically non-compliant topological structures and generate structured candidate reasoning paths.

[0017] The candidate path ranking module is used to perform serialization transformation on the candidate reasoning paths while retaining the factual information of the original graph, extract the structured combination features of each candidate reasoning path, and combine the serialization results with the deep cross-attention features of the biomedical question to calculate a comprehensive ranking score for each candidate reasoning path, thereby selecting high-confidence candidate reasoning paths.

[0018] The evidence-enhanced answer generation module is used to encapsulate the biomedical question, the high-confidence candidate reasoning path, and the corresponding knowledge graph triples into contextual evidence with source annotations, and input it into an autoregressive generative model that incorporates a replication mechanism, and output a natural language answer containing traceable knowledge sources.

[0019] The answer consistency verification module is used to extract the answer entities and predicted generation relationships in the natural language answer, and compare them with the topological logic of the corresponding high-confidence candidate inference path to verify the entity reachability and semantic causal relationship direction of the two. If the verification is passed, the natural language answer, the high-confidence candidate inference path and the corresponding knowledge source are output to form the question answering reasoning result.

[0020] The basic principle and beneficial effects of this invention are as follows: This invention solves the problem through the collaborative work of a question acquisition and entity recognition module and an entity disambiguation and linking module. First, it utilizes a domain-pre-trained language model to identify entities and their predicted types, eliminating interference from conventional vocabulary. Then, in the entity disambiguation and linking process, this invention overcomes the limitations of traditional techniques that rely solely on simple text similarity matching. It comprehensively utilizes five dimensions of information—string similarity, contextual semantic similarity, predicted entity type matching degree, species information, and global graph connectivity—to disambiguate and score candidate entities.

[0021] By introducing "global graph connectivity," the system can analyze the accessibility and semantic tightness of candidate entities and other entities identified in the problem on the knowledge graph topology network (i.e., determine whether they have potential connections or short-step paths in the graph), thereby utilizing the structured prior constraints of the graph to link entities. Through multi-dimensional fusion and proactive disambiguation, it can fundamentally correct ambiguity drift caused by synonyms, historical naming, or cross-species homonyms, ensuring the accuracy of the starting point for inference.

[0022] This invention overcomes this challenge through an efficient dual-convergence design of the bridging entity retrieval module and the multi-hop inference path generation module. The bridging entity retrieval module constructs a dual-channel retrieval mechanism combining lexical matching and graph representation learning semantic retrieval. Specifically, graph representation learning employs graph neural networks to perform multi-layer propagation and aggregation of entities and relationships within the vector space, ensuring that the recalled bridging entities simultaneously embody both textual concepts and graph spatial structure semantics.

[0023] After recalling high-confidence bridging entities, the inference path generation module no longer blindly performs a full-graph blind search, but instead iteratively and directionally approaches these identified bridging entity targets. At each hop of the multi-step expansion, the path expansion controller strictly executes multi-dimensional logical hard-constraint pruning, including entity type, relation direction, species compatibility, and factual confidence: paths that do not conform to the healthy type of the entity chain are pruned; paths that confuse inhibition and activation direction relationships are filtered; paths that cross irrelevant species without compatibility proof are discarded; and paths below the confidence threshold are screened out.

[0024] The combined effect of bidirectional convergence and logical hard constraints cuts off the blind growth of invalid paths at the source, suppresses the exponential path search space into a very small optimal subspace, and greatly shortens the system response time. This not only avoids system paralysis caused by the surge in memory and computation, but also protects core high-quality paths from being buried by messy information.

[0025] This invention constructs a closed-loop barrier from probability ranking and constraint generation to reverse verification through a candidate path ranking module, an evidence-enhanced answer generation module, and an answer consistency verification module. This invention performs feature serialization on multi-hop inference paths and calculates a ranking score that integrates semantic matching, intent alignment, and confidence aggregation to select the optimal path subgraph.

[0026] In answer generation, a probability-gated copying mechanism is constructed to dynamically balance the self-generation probability and the copying probability. This forces the generator to directly copy entity names and relational representation symbols from verifiable high-confidence path triples into the response text, eliminating the illusion of fact caused by the model's free-flowing approach.

[0027] Finally, the answer consistency verification module performs a second verification, which reverses the generated natural language into structured response triples and compares whether the reachability of the entities and the direction of the relationships corresponding to these triples conform to the topological logic of the searched path.

[0028] Outputs that fail consistency checks or whose inference confidence is insufficient are intercepted and downgraded to safe statements, rejecting "random generation." This closed loop minimizes the risk of illusion, and the system outputs a complete visual path diagram, bridging nodes, and verifiable factual sources. This significantly improves the interpretability, transparency, and practical value of the answers.

[0029] Meanwhile, thanks to the "centripetal search" mechanism formed by dual-channel bridging recall and hard pruning control, the algorithm complexity of multi-hop relationship mining is reduced from exponential to near linear, ensuring that multi-hop (especially 3-5 hop) inference can run stably and smoothly in a high-concurrency environment at the millisecond level in ultra-large-scale biological graphs containing millions of nodes and tens of millions of relationships.

[0030] Because entity recognition, graph coding, and path sequencing have a deep understanding of biological semantic relationships, this invention can adapt well to diverse unknown biological question answering, ranging from reasoning about the mechanisms of newly emerging epidemic viruses to the analysis of potential targets for newly synthesized drugs, without modifying the core architecture, and achieves flexible generalization across mechanisms and entities.

[0031] In practical engineering applications, even with a very limited number of standard questions and answer annotation samples, this invention still demonstrates high determinism comparable to a model fine-tuned with massive amounts of data, thanks to the deterministic facts of the base map and the physical isolation of the output consistency verification. This significantly reduces the cold start and annotation costs of the question-answering system and has extremely high commercial and scientific research transformation potential.

[0032] In summary, this invention achieves the effects of eliminating ambiguity, high reasoning efficiency, avoiding factual illusions, and providing interpretable results.

[0033] Furthermore, the problem acquisition and entity recognition module is specifically used for:

[0034] The acquired biomedical questions undergo word segmentation, sub-word splitting, and normalization preprocessing to identify abbreviations in the questions. Simultaneously, during entity recognition, a bidirectional encoder representation model is used to extract contextual representation vectors of the word sequences from the biomedical questions. These contextual representation vectors are then input into a bidirectional long short-term memory network to extract sequence structure features. A conditional random field is used to globally label the sequence symbols based on the prediction transition matrix, thereby identifying biomedical entity fragments belonging to one or more of the following categories: genes, proteins, drugs, diseases, biological pathways, biological processes, phenotypes, cells, tissues, species, and biomarkers. The question acquisition and entity recognition module also utilizes a question intent recognition structure to classify the biomedical questions, predicting the expected answer entity type and relational intent type of the biomedical questions.

[0035] Furthermore, the entity disambiguation and linking module is configured to perform calculations including the following sub-steps when calculating the comprehensive similarity score:

[0036] Calculate the edit distance and character similarity between the biomedical entity fragment and the standard name of the candidate entity in the knowledge graph, and calculate the string matching score;

[0037] The cosine similarity score is calculated by comparing the text vector of the biomedical question with the description text and relational attribute vector of the candidate entity.

[0038] Based on the nearest neighbor relationships in the knowledge graph, and combined with the interconnection degree of other entities detected in the question, the graph connection path density between the candidate entity and the question entity is calculated within a set graph depth threshold to obtain the graph connectivity score.

[0039] The string matching score, the context semantic similarity score, the graph connectivity score, the matching score between the candidate entity type and the predicted entity type, and the matching score between the species information of the candidate entity label and the target species information corresponding to the biomedical question are weighted and summed to obtain the corresponding comprehensive similarity score. The candidate entity with the highest score and the potential related entities within a set degree in its neighborhood are used as the disambiguation mapping result.

[0040] Furthermore, the bridging entity retrieval module is configured to execute the following dual-channel fusion retrieval logic:

[0041] The first channel is vocabulary retrieval. The BM25 retrieval algorithm is used to perform weighted retrieval of the text content of the biomedical question in the entity inverted index of the knowledge graph, matching entity names and aliases that match the question words, and outputting the word meaning retrieval score.

[0042] The second channel is semantic retrieval, which uses graph convolutional neural networks or graph attention networks to aggregate and encode entity nodes and relation edges in the knowledge graph. It fuses the text attribute representation of the node itself with the topological association representation of its multi-hop neighbor nodes, calculates the dot product or cosine of the angle between the question representation vector and the high-dimensional graph entity representation vector in the entity vector space, and outputs the graph semantic retrieval score.

[0043] The word retrieval score and the graph semantic retrieval score are normalized by the maximum absolute value. The normalized scores of the two channels are dynamically weighted by the fusion coefficient. Combined with the path hop count from the starting entity to the candidate bridging entity, the joint recall probability of the bridging entity is determined, and the top-K priority entity nodes are output as candidate bridging entities.

[0044] Furthermore, the multi-hop inference path generation module includes a path expansion controller, which performs real-time pruning of the sub-paths expanded in each generation according to the following filtering rules:

[0045] Path length out-of-bounds pruning: Filters out invalid paths that exceed the set hop count limit or contain the same entity node in a circular repeating pattern;

[0046] Entity type consistency constraint: Based on biomedical logic, perform chain matching on the intermediate types of the starting entity, the bridging entity, and the expected answer entity, and filter out paths where the intermediate node type and the answer endpoint type do not match the predicted entity type and the expected answer entity type;

[0047] Species information span constraint: When the biomedical problem contains a specific species indication, determine whether each jump triple contains other species information that contradicts the specific species. Cross-species reasoning paths must contain explicitly declared species homology mapping relationships; otherwise, they are filtered out.

[0048] Confidence and factual relation constraints: Filter out inference paths that contain triples with factual confidence scores below a set confidence threshold in any jump, and prohibit the mixing or rewriting of weakly correlated or unidentified directional relations into strong causal inference chains.

[0049] Furthermore, the candidate path ranking module is configured to perform the following when calculating the comprehensive ranking score:

[0050] The natural language question and the candidate reasoning path to be evaluated are concatenated into the same sequence as input, and the self-attention mechanism in the cross encoder is used to obtain the cross global feature representation of the question word and the triple entity in the path.

[0051] The overall semantic matching score of the sequence is calculated based on the cross-global feature representation. The following are incorporated into the score: the matching score between the end entity of the candidate reasoning path and the expected answer category, the alignment score between the path relationship sequence and the intended reasoning relationship of the question, and the cumulative or weighted average of the confidence scores of all graph facts experienced in the path.

[0052] Based on the number of hops in the sequence, a damping coefficient is designed. A ranking penalty is applied to paths with longer hop counts. The scores after the damping penalty are weighted and fused to obtain the comprehensive ranking score of the corresponding candidate reasoning paths. The top-N candidate reasoning paths with the highest scores are selected as the optimal reasoning subgraphs and sent to the answer generation stage.

[0053] Furthermore, the evidence-enhanced answer generation module executes the following control logic when generating an autoregressive answer based on contextual evidence:

[0054] The autoregressive generation model is equipped with a probability gating module. At each time step of word decoding and generation, a binary classification generation probability weight gating index is calculated using the encoder hidden state.

[0055] The gating index is used to dynamically balance the generation probability of autonomously predicting new words from the global vocabulary with the copying probability of directly copying words from the prompt text composed of the input biomedical entity fragments, high-confidence candidate reasoning paths and corresponding knowledge graph triples, so that the model can directly extract and output fact-specific words or relation identifiers with medical evidence in the path with a high probability.

[0056] When the system determines that it cannot find a candidate reasoning path that meets the confidence requirement, or that the overall ranking score of the optimal path that meets the requirement is lower than the threshold of alternative answers, it automatically downgrades the answer generator to generate uncertainty warning text to declare the lack of knowledge and refuses to generate mismatched prospects.

[0057] Furthermore, the answer consistency verification module is configured to perform fact verification based on knowledge graph facts, which includes:

[0058] Named entities and declarative semantic nodes containing subject-verb-object features are extracted from the generated natural language answers to construct a sequence of response triples for generating the response;

[0059] The response triple sequence is compared with the high-confidence candidate inference path by entity intersection and topological path logic to verify whether the response entity exists in the entity tree of the inference path and whether the directionality of the response relationship is consistent with the factual relationship in the inference path and the original definition of the knowledge graph.

[0060] The comparison calculation generates a comprehensive consistency check score for the answer based on entity coverage, relation consistency, type matching, and reliability of fact sources. When the consistency check score is lower than a set check threshold, the current natural language answer is intercepted, and the multi-hop reasoning path generation module and the candidate path sorting module are automatically triggered to re-retrieve and replace candidate paths or forcibly remove unfounded entities lacking evidence in the generation process for downgraded output.

[0061] Furthermore, it also includes:

[0062] A dynamic adaptive inference path caching and online local evolution update module is used to establish a structured cache library of frequently asked questions and related high-confidence candidate inference paths in the biomedical field. When the background biological knowledge graph is updated, it detects and identifies dirty data triples that have been added, deleted, or have changed confidence, and extracts the entity identifiers and relation types involved in the dirty data triples. It then traverses the cache library, locates all affected cache paths containing the dirty data triples, and modifies the status identifier of the affected cache paths to a state of pending reconstruction.

[0063] When a new question is received that is identical to the historical question corresponding to the state to be reconstructed, the dynamic adaptive inference path cache and online local evolution update module intercepts the new question, extracts only the local topology subgraph caused by the data update, and controls the multi-hop inference path generation module to perform incremental fine-tuning inference on the specific hop or specific local topology branch affected by the update, while directly reusing the original cache path fragments in the linked entity cache intervals at both ends that are not affected by the update.

[0064] By using incremental fine-tuning of the inference-generated corrected sub-paths to replace the corresponding deprecated cache fragments, an updated target inference path is generated to quickly respond to questions and answers, avoiding recalculation of the global inference network. This achieves second-level updates of multi-hop paths and low-latency correction of data links under the dynamic evolution of the biomedical knowledge graph.

[0065] Furthermore, it also includes:

[0066] The proactive error correction feedback and self-healing graph update module is used to collect failed question-and-answer sessions that are blocked due to the calculated answer consistency verification score being lower than the verification threshold, as well as multiple candidate inference paths that are semantically contradictory to each other generated during the multi-hop process.

[0067] The intercepted failed question-and-answer sessions and unsolvable multi-hop path sequences are packaged into feedback samples, and local dependency analysis is used to identify the bottleneck breakpoint nodes and missing relation edges that cause inference interruption or verification failure.

[0068] Actively trigger the external knowledge base association retrieval interface, and conduct targeted literature mining and rule association reasoning for the bottleneck breakpoint nodes in open and compliant external biological academic databases, and calculate the confidence and credibility level of supplementary relationships in external data;

[0069] When a harmonic and complementary relation for splicing knowledge gaps is successfully mined from an external database, the new triple containing the complementary relation is written into the corresponding biological knowledge graph.

[0070] If no objective evidence is found, the confidence scores of the internal fact triples with semantic or confidence conflicts will be adjusted by a smooth decay function. At the same time, the topology of the topic region involved in the breakpoint will be sent to the human-machine review queue for expert proofreading and updating, so as to realize the self-closed-loop correction of the underlying knowledge resources of the entire question-answering system and the self-healing upgrade of the multi-hop reasoning network. Attached Figure Description

[0071] Figure 1 The logical block diagram of a biological knowledge graph question-answering system based on multi-hop bridge reasoning;

[0072] Figure 2 This is a flowchart of the sub-workflow of the problem acquisition and entity recognition module;

[0073] Figure 3 Flowchart for bridging entity dual-channel retrieval. Detailed Implementation

[0074] The following detailed description illustrates the specific implementation method:

[0075] Biological knowledge graph question answering system based on multi-hop bridge reasoning (e.g.) Figure 1 (as shown), including:

[0076] The question acquisition and entity recognition module is used to acquire biomedical questions in natural language form, input them into the domain language encoder for feature extraction, and identify biomedical entity fragments in the biomedical questions and the predicted entity type corresponding to each biomedical entity fragment through a sequence labeling algorithm.

[0077] The entity disambiguation and linking module is used to construct a candidate entity set based on the identified biomedical entity fragments, and to calculate the comprehensive similarity score of each candidate entity based on string similarity, contextual semantic similarity, matching degree between candidate entity type and predicted entity type, species information in the question, and global graph connectivity of the knowledge graph. The biomedical entity fragments are then mapped to the target standard entity with the highest score in the knowledge graph as the starting entity for reasoning.

[0078] The bridging entity retrieval module is used to perform a dual-channel retrieval in the knowledge graph based on the starting entity, utilizing a retrieval channel based on word matching and a semantic retrieval channel based on graph representation learning, to recall bridging entities that are semantically associated with the biomedical question.

[0079] The multi-hop reasoning path generation module is used to perform multi-step iterative expansion from the starting entity to the bridging entity in the knowledge graph, and execute security logic constraints including entity type constraints, relationship direction constraints, species compatibility constraints and fact confidence filtering at each stage of path expansion, so as to eliminate biologically non-compliant topological structures and generate structured candidate reasoning paths.

[0080] The candidate path ranking module is used to perform serialization transformation on the candidate reasoning paths while retaining the factual information of the original graph, extract the structured combination features of each candidate reasoning path, and combine the serialization results with the deep cross-attention features of the biomedical question to calculate a comprehensive ranking score for each candidate reasoning path, thereby selecting high-confidence candidate reasoning paths.

[0081] The evidence-enhanced answer generation module is used to encapsulate the biomedical question, the high-confidence candidate reasoning path, and the corresponding knowledge graph triples into contextual evidence with source annotations, and input it into an autoregressive generative model that incorporates a replication mechanism, and output a natural language answer containing traceable knowledge sources.

[0082] The answer consistency verification module is used to extract the answer entities and predicted generation relationships in the natural language answer, and compare them with the topological logic of the corresponding high-confidence candidate inference path to verify the entity reachability and semantic causal relationship direction of the two. If the verification is passed, the natural language answer, the high-confidence candidate inference path and the corresponding knowledge source are output to form the question answering reasoning result.

[0083] In practical use, it is mainly deployed in intelligent computing servers equipped with graphics processing units (GPUs, such as NVIDIA A100) and high-speed memory, and provides natural language question answering interfaces to the outside world through a microservice architecture to handle complex biomedical question answering tasks.

[0084] The problem acquisition and entity recognition module specifically includes the following sub-workflows (such as...). Figure 2 (as shown)

[0085] First, the acquired natural language questions undergo standardization and cleaning, including converting full-width characters to half-width characters, standardizing Chinese and English symbols, and removing punctuation marks that do not affect semantics. Then, sub-word segmentation tools such as Douduen or SentencePiece are used to obtain word sequences. ,in The number of lexical units is used. A pre-read dictionary of commonly used abbreviations in the biomedical field is used to restore the original words from the identified abbreviations, and mapping markers are inserted into the lexical sequence.

[0086] The processed word sequence The input is placed into the pre-trained domain-specific language model. In this embodiment, the domain language encoder uses a BioBERT or PubMedBERT model built on the Transformer architecture. Bidirectional feature extraction is performed through a multi-layer self-attention mechanism in the model to obtain the context representation vector of each word.

[0087]

[0088] in, This represents the feature dimension of the hidden layer, typically set to 768.

[0089] Next, the context representation vector sequence The input is fed into a Bidirectional Long Short-Term Memory (BiLSTM) network to further extract temporal and bidirectional long-range sequence structure features, obtaining a feature representation after concatenation of the forward and backward passes:

[0090]

[0091] in, The hidden layer number of the unidirectional LSTM is set to 256 in this embodiment, therefore the bidirectional stitching dimension is... .

[0092] Next, the feature representation vector The input is fed into a Conditional Random Field (CRF) layer. The CRF operates based on the state transition probability matrix. We perform global modeling of the label assignment for the entire sequence and use the Viterbi algorithm to calculate the optimal entity label sequence that maximizes the global probability. :

[0093]

[0094] in, It is for tags The projection weights of the fully connected layer; Indicates by label Move to label The transition probability score is used to determine the biomedical entity fragment based on the output of the above model. The start and end positions and classifications. In this embodiment, the specific biomedical entity fragment types identified include at least:

[0095] Gene: Gene;

[0096] Protein: Protein;

[0097] Mutation: Gene mutation;

[0098] Drug: medicine;

[0099] Disease: illness;

[0100] Pathway: biological pathway;

[0101] Biological Process;

[0102] Phenotype: Phenotype;

[0103] Cell: a cell;

[0104] Tissue: organization;

[0105] Species: species;

[0106] Biomarker: a biomarker.

[0107] The question acquisition and entity recognition module aggregates features from natural language questions through an external intent recognition multi-label classifier, thereby predicting the type of the expected answer entity implied by the question. (e.g., pathways) and relational intent types (e.g., facilitating, inhibiting, or binding relationships). This intent classifier employs a single-layer MLP structure external to the top layer of BioBERT, using the sigmoid function to output the probabilities of each class.

[0108] In this embodiment, the training phase of the entity recognition model (BioBERT-BiLSTM-CRF) is performed on publicly labeled datasets (such as JNLPBA, BC5CDR, and NCBI-Disease) that annotate gene, drug, and disease entities. Model optimization uses the negative log-likelihood loss function as the training loss value, and the AdamW optimizer is used to update parameters. The initial learning rate is set to... After 20 epochs of iteration, the F1-Score on the validation set stabilized at above 0.91.

[0109] The specific calculation logic of the entity disambiguation and linking module includes the following steps:

[0110] For a biomedical entity fragment identified by the pre-stage (e.g., "TP53"), firstly, fuzzy matching and inverted index retrieval are performed in the candidate alias table and synonym table of the biomedical knowledge graph, and the top-ranked terms are returned. (In this embodiment) Entities with a value of 30 or higher constitute a candidate entity set. .

[0111] To ensure accurate disambiguation, for each candidate entity The scores are calculated and normalized across the following five dimensions using a computational circuit:

[0112] String matching score Calculate the Jaccard distance and Levenstein edit distance between the input string and the standard name and alias of the candidate entities. The formula is as follows:

[0113]

[0114] In this embodiment, the weighting coefficient The value is 0.4.

[0115] Contextual semantic similarity score : Candidate entities The descriptive text and relational attribute features in the graph are concatenated into textual sequences and then converted into high-dimensional descriptive vectors using PubMedBERT. The original question is encoded using the same model to obtain the question vector. Calculate the cosine similarity between the two:

[0116]

[0117] Type Matching Score When candidate entities When the entity type labeled in the knowledge graph matches the predicted entity type identified in the previous stage. ,otherwise .

[0118] Species matching score If species-restricting words (such as "human" or "mouse") are extracted from the original question, and the candidate entities... If the biological species to which the species belongs matches the limiting term, then If a conflict exists, the value will be 0. If the original sentence does not explicitly mention a species, the default configuration is human, with a matching score of 0.8.

[0119] Graph connectivity score To prevent local fragmentation caused by independent entity linking, for all other pre-linked entity sets extracted simultaneously in this question... Calculate candidate entities in a knowledge graph and The topological path density between entity nodes in the graph. Specifically, the graph depth threshold is calculated using the following formula. Within the jump, the sum of the reciprocals of the number of paths between the candidate entity and each entity:

[0120]

[0121] in, Represents a node and The number of jumps between them is The number of reachable paths. In this embodiment, if no path exists, the score for this item is set to 0.

[0122] Finally, calculate the overall similarity score. :

[0123]

[0124] In this embodiment, each weight coefficient is set to , , , , The entity linking module in the candidate entity set The candidate entity with the highest comprehensive score is selected as the target standard entity for disambiguation and mapping. This entity node is used as the starting node for multi-hop inference.

[0125] The bridging entity retrieval module is used to perform dual-channel retrieval in the knowledge graph based on the starting entity, utilizing a lexical matching-based retrieval channel and a graph representation learning-based semantic retrieval channel to recall bridging entities that are semantically related to biomedical questions.

[0126] In multi-hop reasoning of biological maps, to avoid the exponential state explosion problem caused by full-graph search, the bridging entity retrieval module is configured to execute dual-channel retrieval logic (e.g., ...). Figure 3 (as shown)

[0127] First channel: BM25 vocabulary retrieval. This embodiment uses the segmented text of the natural language question as the retrieval item, performs word frequency retrieval in the inverted index of a large-scale knowledge graph entity, obtains graph nodes that strongly match the keywords in the question, and calculates the word meaning retrieval score accordingly. .

[0128] Second channel: Graph representation learning for semantic retrieval. This embodiment uses a Graph Convolutional Network (GCN) or a Relational Graph Convolutional Network (R-GCN) to train the entire knowledge graph offline for graph embedding. Node In the The formula for calculating the hidden state update of a layer is:

[0129]

[0130] in, Represents a node Through relationships The set of connected neighbor nodes; It is a relationship In the The linear transformation matrix of the layer; It is a normalized positive definite constant, which can be set as ; The activation function is ReLU (used in this embodiment). The node representation dimension of the final layer output is 256.

[0131] The candidate node representation vector recalled during the entity retrieval stage With the transformed question vector Perform cosine similarity calculation and output the graph semantic retrieval score. .

[0132] The scores of the two channels are converted using the maximum value normalization algorithm. After scoring, a fusion coefficient is used. Perform weighted calculations to determine the joint recall probability. :

[0133]

[0134] in, and These are the normalized scores and the dynamically adjusted fusion coefficients. Set it to 0.4; This indicates whether the path connection between the candidate node and the starting entity is related to the problem prediction. If they match, the value of that item is 1; otherwise, it is 0. (Weight) The score is 0.20. Based on the joint recall probability, nodes are sorted from high to low, low-scoring nodes are filtered out, and the top-scoring nodes are truncated. Each node is considered a candidate bridging entity to be bridged.

[0135] The multi-hop reasoning path generation module (hereinafter referred to as the "path generation module") is used to expand the knowledge graph from the starting entity to the bridging entity in multiple steps. At each stage of the path expansion, it executes safety logic constraints including entity type constraints, relationship direction constraints, species compatibility constraints, and fact confidence filtering to eliminate biologically non-compliant topological structures and generate structured candidate reasoning paths.

[0136] The starting entity and candidate bridging entities are cascaded via multi-hop paths. The path generation module executes a breadth-first search (BFS) algorithm or a bidirectional search algorithm through the path expansion controller. At each stage of the path expansion iteration, the following safety logic hard constraints are executed to narrow the search path space:

[0137] Path length overrun pruning: Limit the total number of hops to between 2 and 5 (i.e., only retain the number of hops including intermediate bridge nodes). (Topology). If the same entity identifier is repeatedly contained in the current generated path, it is considered a redundant loop and is immediately removed.

[0138] Entity type constraints: A compliant syntax rule table for biomedical entity transformation chains is pre-built in memory. For example, "drug-target-encoding gene-involved pathway" is a compliant transformation chain, while "drug-disease-organelle" is an non-compliant chain. The pruning module verifies the type of each adjacent node. If the type of a node connected to a certain intermediate stage deviates from the compliant syntax definition, the branch is immediately interrupted and discarded.

[0139] Species compatibility constraint: If the target of the preceding problem identification prediction is the human species, the species information label (Species) of the corresponding graph fact triple for each step of the multi-hop generation must be "Homo sapiens" (human) or have no specific species identifier. When a cross-species path jumps at an intermediate node (such as when mouse experimental evidence is associated with humans), it must include a "Homology_of" relationship triple that has been explicitly declared in the graph, as a legitimate transition bridge for species conversion, and splicing without logical spans is prohibited.

[0140] Fact confidence filtering: Each relation fact in the knowledge graph has a pre-stored fact confidence score calculated based on the frequency of support from authoritative literature. During multi-hop expansion, the controller extracts the confidence level of the associated facts. Once the single-step confidence level falls below a set hard threshold... The branch is then filtered out. If low-confidence triples exist along the entire path, the path is marked as unsuitable for strong causal reasoning in the future.

[0141] Relationship directionality verification: Jumps must be made in the direction of the directed graph arc defined in the fact in the knowledge graph to prevent the semantic contradiction logic of "the target point is activated by the pathway" leading to "the pathway activates the target point" due to reverse reasoning.

[0142] In this embodiment, after the path generation module performs the aforementioned security constraint pruning, each qualified path generated is stored as a topological chain containing multi-hop entities and relation sequences, serving as a set of candidate inference paths. .

[0143] The candidate path ranking module is used to serialize the candidate reasoning paths while preserving the factual information of the original graph, extract the structured combination features of each candidate reasoning path, and combine the serialization results with the deep cross-attention features of biomedical questions to calculate a comprehensive ranking score for each candidate reasoning path, thereby selecting high-confidence candidate reasoning paths.

[0144] To perform accurate semantic matching and ranking of the candidate multi-hop path set, this module first serializes the candidate reasoning paths in graph topology form. The conversion template is: "[CLS] Question Text [SEP]". [SEP]”.

[0145] This helps maintain the atomicity and structured attribute information of graph entities and relationships. For each serialized path The scoring is performed using a cross-encoder. In this embodiment, the cross-encoder uses the DeBERTa model, concatenating the question and the serialization path as a single input:

[0146]

[0147] Extract the [CLS] vector, which represents the global context classification indicator feature at the top layer, and input it into the fully connected layer for mapping to obtain the semantic similarity matching score. .

[0148] Next, a comprehensive score is calculated by combining the multi-dimensional structured features extracted from the path. The comprehensive ranking score is then determined. The expression is as follows:

[0149]

[0150] in:

[0151] This indicates the entity type of the last node in the candidate path and the entity type of the expected answer. For consistency, if the types match, the calculated value of this item is 1.0, otherwise it is 0.

[0152] The confidence scores of all graph facts traversed along the path are multiplied and summed to reflect the overall reliability of the knowledge chain:

[0153]

[0154] The path length damping penalty function is set to prevent the accumulation of low information along long paths. The calculation formula is:

[0155]

[0156] in, To set the penalty factor (0.15 in this example), This represents the total number of hops for the current path. Set to the baseline hop count (set to 2).

[0157] , , In this embodiment, the weighting parameters for the corresponding feature indicators are respectively set as follows: , , .

[0158] Sort the paths by their scores from highest to lowest, and then select the path with the highest score at the top. Article (in this embodiment) Candidate reasoning paths with a value of 5 are marked as high-confidence candidate reasoning paths and used as key evidence for answer generation.

[0159] The cross-encoder model is trained using a pairwise margin loss to adjust the ranking parameters.

[0160]

[0161] Among them, positive samples The main reasoning path that corresponds to the facts leading to the correct answer in the indicator annotation data, negative samples For randomly selected or pruned mismatched negative inference paths. Interval parameter. Set to 0.15.

[0162] The evidence-enhanced answer generation module encapsulates biomedical questions, high-confidence candidate reasoning paths, and corresponding knowledge graph triples into contextual evidence with source annotations. This evidence is then input into an autoregressive generative model that incorporates a replication mechanism, and the output is a natural language answer containing traceable knowledge sources.

[0163] The answer generator employs an encoder-decoder architecture that includes cross-attention interactions (in this embodiment, a BART-Large or T5-Large large model).

[0164] The preceding concatenation of the input text is constructed as follows:

[0165]

[0166] In the autoregressive generation decoding process, to ensure the absolute robustness and high-confidence fact reproduction of the generated medical terms and relation edge names, the decoder is configured with a fusion copy mechanism (Copy Mechanism / Pointer-Generator).

[0167] At the generation time step For each lexical element in the candidate vocabulary Calculate its mixed output probability The calculation formula is as follows:

[0168]

[0169] in:

[0170] This represents the probability of the word to be predicted calculated from the global vocabulary using an autoregressive method.

[0171] Indicates the decoding time step Input prompt words The calculated cross-attention weight allocation values ​​express the degree of tendency to extract and copy specific lexical units from the input contextual prompts.

[0172] It is an autoregressive probability gating index. Based on the current decoding state. Contextual attention feature vectors and already generated lexical guides The calculation is performed by fitting a perceptron network containing a sigmoid function:

[0173]

[0174] If the system determines that the path obtained from the path sorting module has the highest score. Less than the set threshold This indicates that the knowledge graph completely lacks the biomedical physics background to support the derivation. At this point, the answer generator will automatically downgrade, blocking the generation of regular answers and instead selecting a declarative statement from the system's pre-set safe response scripts, such as a warning text like "Due to factual limitations in the knowledge graph, no valid supporting evidence can be provided at present."

[0175] The answer consistency verification module (hereinafter referred to as the "verification module") is used to extract the answer entities and predicted generation relationships from the natural language answer, and compare them with the topological logic of the corresponding high-confidence candidate inference path. It verifies the entity reachability and semantic causal relationship direction of the two. If the verification passes, it outputs the natural language answer, the high-confidence candidate inference path, and the corresponding knowledge source to constitute the question-answering reasoning result. The answer consistency verification module is built at the output end of the answer generator. Before the final confirmation of the answer, the output semantics undergo a double verification check:

[0176] Entity reachability verification: Extract textual units from the generated natural language answer and use a named entity extraction algorithm to compare their intersection with the node set in the core top-3 inference paths. If neither of the two key diseases or therapeutic proteins claimed in the answer can be connected to the entities in the previously derived multi-hop inference paths, it is considered an error.

[0177] Semantic causal consistency verification: The cascading relationships described in the generated text segment are extracted into a "subject-relation-object" declaration expression (response triple sequence). The verification module aligns this expression with the arc direction of the candidate paths in the original graph. If the inference path displays "TP53 mutation",... lead to The text states that "loss of function" is incorrect, while the generated text claims that "TP53 deficiency led to mutation," which is considered a reversal of causal logic.

[0178] Consistency check score The quantitative calculation formula is as follows:

[0179]

[0180] in, and These represent the sets of entities and relations contained in the natural language answer, respectively. and These represent the sets of entities and directed edge relationships contained in the high-confidence candidate inference path, respectively. This refers to the number of logical causal errors that are opposite to the candidate reasoning path or cannot be supported by a source in the underlying graph. In this embodiment, each weight coefficient is set to... , , .

[0181] If the calculated lower limit of the consistency check score is lower than the set check threshold If this happens, the current autoregressive answer generation is directly intercepted. The system triggers a backtracking mechanism: instructs the multi-hop inference module to call other subsequent inference topologies from the candidate pool as evidence input to re-decode, or directly perform forced downgrading and pruning on the output string, discarding unverifiable detail node information before outputting.

[0182] In practical applications, for example, when receiving natural language input from a user. :

[0183] Which gene mutations lead to the loss of TP53 function, thereby increasing the risk of cancer?

[0184] Step 1: Entity Recognition and Disambiguation

[0185] The problem acquisition and entity recognition module Sequence segmentation was performed to identify the entity fragment "TP53" (predicted as Gene), the gene mutation-related term "loss of function" (predicted as function indicator), and the endpoint term "cancer risk" (predicted as Disease).

[0186] The entity disambiguation and linking module generates 20 candidate nodes for "TP53" in the graph index, including human [TP53] (gene ID: 7157) and mouse [Trp53]. Species matching filtering is used, and when no specific species is detected, it defaults to Homo sapiens (human). Within the graph topology two-hop, human [TP53] has a high graph connectivity score of 0.88 (having multiple connections related to mutations) compared to various potential "gene mutation" disease nodes in the sentence, while mouse [Trp53] has a graph connectivity score of 0 (because the corresponding mouse tumor node cannot be directly connected in this context). Finally, the module disambiguates and links "TP53" to the standard human gene: Gene:7157 (TP53).

[0187] Step 2: Bridging Retrieval and Multi-Hop Expansion

[0188] The bridging entity retrieval module, starting with Gene:7157 (TP53), recalled core bridging entities including “cancer” (Disease:0001), “Li-Fraumeni syndrome” (Disease:0002), and “p53 activity deficiency” (Biological Process:0055).

[0189] The multi-hop inference path generation module begins its exploration from the Gene:7157 (TP53) node. During the intermediate generation phase, the path expansion controller excludes reverse causality and matches type-compliant templates. Multi-hop topology paths are generated:

[0190] Path 1: Mutation:1234 (R273H) Protein: P04637 (p53) BiologicalProcess:0055 (DNA_Repair) Disease: 0001 (Cancer)

[0191] Path 2: Mutation:5678 (R175H) Protein: P04637 (p53) BiologicalProcess:0110 (Apoptosis) Disease: 0001 (Cancer)

[0192] Step 3: Sorting and Answer Generation Verification

[0193] The candidate path ranking module serializes multiple generated topological chains (including path 1 and path 2) into DeBERTa encoding. Path 1 has a high literature citation confidence score, resulting in a composite score of 0.89; path 2 scores 0.84. Both paths are then fed into the encoder-decoder.

[0194] The evidence-enhanced answer generation module introduces replication-gated computation, highly replicating key feature words such as "R273H", "R175H", "p53 activity", and "DNA repair function" during the lexical prediction stage. The response text obtained by autoregressive decoding is as follows:

[0195] "Mutations in the TP53 gene (such as R273H and R175H) can cause conformational changes in the p53 protein it encodes, resulting in the loss of its original DNA repair and apoptosis-promoting functions. This causes cells to lose their monitoring of damaged genes, greatly increasing the risk of cancer."

[0196] The answer consistency verification module extracts the named entities "R273H", "R175H", "p53", and "cancer" from the generated response, as well as the causal judgment words "leading to loss of function" and "increasing risk," and compares the graph structures of paths 1 and 2. Entity matching rate. The directions are consistent. The final output is this response, which is then displayed on the page as a graph topology showing the complete multi-hop inference dependency path from mutation through protein mechanisms to cancer.

[0197] For example, regarding the user's natural language problem "Through what pathway does metformin lower blood sugar?"

[0198] The system located "Metformin" (predicted type: Drug) and "Blood Glucose" (predicted endpoint phenotype / Biological Process). The entity disambiguation and linking module successfully mapped the standard chemical entity Drug:DB00331 (Metformin) in the spectrum by matching the name and spelling in the standard library.

[0199] Through semantic recall, the core genes "AMPK" (including synaptic metabolic pathway), "LKB1", and rate-limiting enzyme targets related to multi-pathway metabolic blood glucose reduction were found.

[0200] Exploring direct drug targeting. Due to the existence of type-constrained chains, proteins without association are filtered out during expansion, identifying two main derivation chains:

[0201] Path 1: Drug:Metformin Protein: AMPK complex Pathway:AMPK_signaling_pathway BiologicalProcess:Gluconeogenesis Phenotype:Blood_Glucose

[0202] Path 2: Drug:Metformin Protein: Complex I AMP / ATP_ratio Protein: AMPK complex Phenotype:Blood_Glucose

[0203] The confidence scores of each node in Path 1 were all above 0.80, with a comprehensive ranking score of 0.92. Path 2, due to its crossing of non-protein metabolic intermediates, had a damping reduction score of 0.76.

[0204] The generator synthesizes the triples of path 1 to generate the response:

[0205] "Metformin mainly lowers blood sugar by activating the AMPK complex to trigger the AMPK signaling pathway, thereby inhibiting gluconeogenesis in the liver and promoting glucose uptake in peripheral tissues."

[0206] The consistency check confirms that the entity and logical direction are completely valid, and then the output is allowed.

[0207] The above are merely embodiments of the present invention. The invention is not limited to the fields covered by these embodiments. Commonly known structures and characteristics in the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are able to access all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A biological knowledge graph question-answering system based on multi-hop bridge reasoning, characterized in that, include: The question acquisition and entity recognition module is used to acquire biomedical questions in natural language form, input them into the domain language encoder for feature extraction, and identify biomedical entity fragments in the biomedical questions and the predicted entity type corresponding to each biomedical entity fragment through a sequence labeling algorithm. The entity disambiguation and linking module is used to construct a candidate entity set based on the identified biomedical entity fragments, and to calculate the comprehensive similarity score of each candidate entity based on string similarity, contextual semantic similarity, matching degree between candidate entity type and predicted entity type, species information in the question, and global graph connectivity of the knowledge graph. The biomedical entity fragments are then mapped to the target standard entity with the highest score in the knowledge graph as the starting entity for reasoning. The bridging entity retrieval module is used to perform a dual-channel retrieval in the knowledge graph based on the starting entity, utilizing a retrieval channel based on word matching and a semantic retrieval channel based on graph representation learning, to recall bridging entities that are semantically associated with the biomedical question. The multi-hop reasoning path generation module is used to perform multi-step iterative expansion from the starting entity to the bridging entity in the knowledge graph, and execute security logic constraints including entity type constraints, relationship direction constraints, species compatibility constraints and fact confidence filtering at each stage of path expansion, so as to eliminate biologically non-compliant topological structures and generate structured candidate reasoning paths. The candidate path ranking module is used to perform serialization transformation on the candidate reasoning paths while retaining the factual information of the original graph, extract the structured combination features of each candidate reasoning path, and combine the serialization results with the deep cross-attention features of the biomedical question to calculate a comprehensive ranking score for each candidate reasoning path, thereby selecting high-confidence candidate reasoning paths. The evidence-enhanced answer generation module is used to encapsulate the biomedical question, the high-confidence candidate reasoning path, and the corresponding knowledge graph triples into contextual evidence with source annotations, and input it into an autoregressive generative model that incorporates a replication mechanism, and output a natural language answer containing traceable knowledge sources. The answer consistency verification module is used to extract the answer entities and predicted generation relationships in the natural language answer, and compare them with the topological logic of the corresponding high-confidence candidate inference path to verify the entity reachability and semantic causal relationship direction of the two. If the verification is passed, the natural language answer, the high-confidence candidate inference path and the corresponding knowledge source are output to form the question answering reasoning result.

2. The biological knowledge graph question-answering system based on multi-hop bridge reasoning according to claim 1, characterized in that, The problem acquisition and entity recognition module is specifically used for: The acquired biomedical questions undergo word segmentation, sub-word splitting, and normalization preprocessing to identify abbreviations in the questions. Simultaneously, during entity recognition, a bidirectional encoder representation model is used to extract contextual representation vectors of the word sequences from the biomedical questions. These contextual representation vectors are then input into a bidirectional long short-term memory network to extract sequence structure features. A conditional random field is used to globally label the sequence symbols based on the prediction transition matrix, thereby identifying biomedical entity fragments belonging to one or more of the following categories: genes, proteins, drugs, diseases, biological pathways, biological processes, phenotypes, cells, tissues, species, and biomarkers. The question acquisition and entity recognition module also utilizes a question intent recognition structure to classify the biomedical questions, predicting the expected answer entity type and relational intent type of the biomedical questions.

3. The biological knowledge graph question-answering system based on multi-hop bridge reasoning according to claim 2, characterized in that, The entity disambiguation and linking module is configured to perform calculations including the following sub-steps when calculating the overall similarity score: Calculate the edit distance and character similarity between the biomedical entity fragment and the standard name of the candidate entity in the knowledge graph, and calculate the string matching score; The cosine similarity score is calculated by comparing the text vector of the biomedical question with the description text and relational attribute vector of the candidate entity. Based on the nearest neighbor relationships in the knowledge graph, and combined with the interconnection degree of other entities detected in the question, the graph connection path density between the candidate entity and the question entity is calculated within a set graph depth threshold to obtain the graph connectivity score. The string matching score, the context semantic similarity score, the graph connectivity score, the matching score between the candidate entity type and the predicted entity type, and the matching score between the species information of the candidate entity label and the target species information corresponding to the biomedical question are weighted and summed to obtain the corresponding comprehensive similarity score. The candidate entity with the highest score and the potential related entities within a set degree in its neighborhood are used as the disambiguation mapping result.

4. The biological knowledge graph question-answering system based on multi-hop bridge reasoning according to claim 3, characterized in that, The bridging entity retrieval module is configured to execute the following dual-channel fusion retrieval logic: The first channel is vocabulary retrieval. The BM25 retrieval algorithm is used to perform weighted retrieval of the text content of the biomedical question in the entity inverted index of the knowledge graph, matching entity names and aliases that match the question words, and outputting the word meaning retrieval score. The second channel is semantic retrieval, which uses graph convolutional neural networks or graph attention networks to aggregate and encode entity nodes and relation edges in the knowledge graph. It fuses the text attribute representation of the node itself with the topological association representation of its multi-hop neighbor nodes, calculates the dot product or cosine of the angle between the question representation vector and the high-dimensional graph entity representation vector in the entity vector space, and outputs the graph semantic retrieval score. The word retrieval score and the graph semantic retrieval score are normalized by the maximum absolute value. The normalized scores of the two channels are dynamically weighted by the fusion coefficient. Combined with the path hop count from the starting entity to the candidate bridging entity, the joint recall probability of the bridging entity is determined, and the top-K priority entity nodes are output as candidate bridging entities.

5. The biological knowledge graph question-answering system based on multi-hop bridge reasoning according to claim 4, characterized in that, The multi-hop inference path generation module includes a path expansion controller, which performs real-time pruning of the sub-paths expanded in each generation according to the following filtering rules: Path length out-of-bounds pruning: Filters out invalid paths that exceed the set hop count limit or contain the same entity node in a circular repeating pattern; Entity type consistency constraint: Based on biomedical logic, perform chain matching on the intermediate types of the starting entity, the bridging entity, and the expected answer entity, and filter out paths where the intermediate node type and the answer endpoint type do not match the predicted entity type and the expected answer entity type; Species information span constraint: When the biomedical problem contains a specific species indication, determine whether each jump triple contains other species information that contradicts the specific species. Cross-species reasoning paths must contain explicitly declared species homology mapping relationships; otherwise, they are filtered out. Confidence and factual relation constraints: Filter out inference paths that contain triples with factual confidence scores below a set confidence threshold in any jump, and prohibit the mixing or rewriting of weakly correlated or unidentified directional relations into strong causal inference chains.

6. The biological knowledge graph question-answering system based on multi-hop bridge reasoning according to claim 5, characterized in that, The candidate path sorting module is configured to execute the following when calculating the comprehensive sorting score: The natural language question and the candidate reasoning path to be evaluated are concatenated into the same sequence as input, and the self-attention mechanism in the cross encoder is used to obtain the cross global feature representation of the question word and the triple entity in the path. The overall semantic matching score of the sequence is calculated based on the cross-global feature representation. The following are incorporated into the score: the matching score between the end entity of the candidate reasoning path and the expected answer category, the alignment score between the path relationship sequence and the intended reasoning relationship of the question, and the cumulative or weighted average of the confidence scores of all graph facts experienced in the path. Based on the number of hops in the sequence, a damping coefficient is designed. A ranking penalty is applied to paths with longer hop counts. The scores after the damping penalty are weighted and fused to obtain the comprehensive ranking score of the corresponding candidate reasoning paths. The top-N candidate reasoning paths with the highest scores are selected as the optimal reasoning subgraphs and sent to the answer generation stage.

7. The biological knowledge graph question-answering system based on multi-hop bridge reasoning according to claim 6, characterized in that, The evidence-enhanced answer generation module executes the following control logic when generating an autoregressive answer based on contextual evidence: The autoregressive generation model is equipped with a probability gating module. At each time step of word decoding and generation, a binary classification generation probability weight gating index is calculated using the encoder hidden state. The gating index is used to dynamically balance the generation probability of autonomously predicting new words from the global vocabulary with the copying probability of directly copying words from the prompt text composed of the input biomedical entity fragments, high-confidence candidate reasoning paths and corresponding knowledge graph triples, so that the model can directly extract and output fact-specific words or relation identifiers with medical evidence in the path with a high probability. When the system determines that it cannot find a candidate reasoning path that meets the confidence requirement, or that the overall ranking score of the optimal path that meets the requirement is lower than the threshold of alternative answers, it automatically downgrades the answer generator to generate uncertainty warning text to declare the lack of knowledge and refuses to generate mismatched prospects.

8. The biological knowledge graph question-answering system based on multi-hop bridge reasoning according to claim 7, characterized in that, The answer consistency verification module is configured to perform fact verification based on knowledge graph facts, including: Named entities and declarative semantic nodes containing subject-verb-object features are extracted from the generated natural language answers to construct a sequence of response triples for generating the response; The response triple sequence is compared with the high-confidence candidate inference path by entity intersection and topological path logic to verify whether the response entity exists in the entity tree of the inference path and whether the directionality of the response relationship is consistent with the factual relationship in the inference path and the original definition of the knowledge graph. The comparison calculation generates a comprehensive consistency check score for the answer based on entity coverage, relation consistency, type matching, and reliability of fact sources. When the consistency check score is lower than a set check threshold, the current natural language answer is intercepted, and the multi-hop reasoning path generation module and the candidate path sorting module are automatically triggered to re-retrieve and replace candidate paths or forcibly remove unfounded entities lacking evidence in the generation process for downgraded output.

9. The biological knowledge graph question-answering system based on multi-hop bridge reasoning according to claim 8, characterized in that, Also includes: The dynamic adaptive inference path caching and online local evolution update module is used to build a structured cache library for frequently asked questions and related high-confidence candidate inference paths in the biomedical field. When the background biological knowledge graph is updated, dirty data triples that have been added, deleted, or have changed confidence are detected and identified. The entity identifiers and relation types involved in the dirty data triples are extracted. The cache library is traversed to locate all affected cache paths that contain the dirty data triples, and the status identifier of the affected cache paths is modified to a state of pending reconstruction. When a new question is received that is identical to the historical question corresponding to the state to be reconstructed, the dynamic adaptive inference path cache and online local evolution update module intercepts the new question, extracts only the local topology subgraph caused by the data update, and controls the multi-hop inference path generation module to perform incremental fine-tuning inference on the specific hop or specific local topology branch affected by the update, while directly reusing the original cache path fragments in the linked entity cache intervals at both ends that are not affected by the update. By using incremental fine-tuning of the inference-generated corrected sub-paths to replace the corresponding deprecated cache fragments, an updated target inference path is generated to quickly respond to questions and answers, avoiding recalculation of the global inference network. This achieves second-level updates of multi-hop paths and low-latency correction of data links under the dynamic evolution of the biomedical knowledge graph.

10. The biological knowledge graph question-answering system based on multi-hop bridge reasoning according to claim 9, characterized in that, Also includes: The proactive error correction feedback and self-healing graph update module is used to collect failed question-and-answer sessions that are blocked due to the calculated answer consistency verification score being lower than the verification threshold, as well as multiple candidate inference paths that are semantically contradictory to each other generated during the multi-hop process. The intercepted failed question-and-answer sessions and unsolvable multi-hop path sequences are packaged into feedback samples, and local dependency analysis is used to identify the bottleneck breakpoint nodes and missing relation edges that cause inference interruption or verification failure. Actively trigger the external knowledge base association retrieval interface, and conduct targeted literature mining and rule association reasoning for the bottleneck breakpoint nodes in open and compliant external biological academic databases, and calculate the confidence and credibility level of supplementary relationships in external data; When a harmonic and complementary relation for splicing knowledge gaps is successfully mined from an external database, the new triple containing the complementary relation is written into the corresponding biological knowledge graph. If no objective evidence is found, the confidence scores of the internal fact triples with semantic or confidence conflicts will be adjusted by a smooth decay function. At the same time, the topology of the topic region involved in the breakpoint will be sent to the human-machine review queue for expert proofreading and updating, so as to realize the self-closed-loop correction of the underlying knowledge resources of the entire question-answering system and the self-healing upgrade of the multi-hop reasoning network.

Images