Multilingual question answering method and device based on semantic-symbol knowledge consensus

Abstract

The present application relates to a multi-language question answering method and device based on semantic-symbol knowledge consensus, and belongs to the technical field of natural language processing and knowledge graph. The present application comprises four parts: constructing a multi-language reasoning path, constructing a multi-language evidence pool, executing a multi-language voting consensus mechanism, and query statement assembly and execution. The method first diverges a single-language query into extensive multi-language parallel evidence in the semantic space and maps it to a structured candidate set; then, it uses the symbolic space to eliminate language differences through a multi-language voting mechanism, converging the collected information into a robust consensus identifier; finally, it generates an accurate structured query statement based on the consensus mechanism to obtain the final answer. The present application can effectively utilize global multi-language knowledge, significantly improving the multi-language question answering accuracy and query executability in multi-hop complex reasoning scenarios, and has important practical application value.

Description

Technical Field

[0001] This invention relates to a multilingual question-answering method and apparatus based on semantic-symbolic knowledge consensus, belonging to the fields of natural language processing and knowledge graph technology. Background Technology

[0002] Multilingual Question Answering (MQA) aims to answer user queries in different languages ​​by leveraging a wide range of information sources. With the proliferation of Large Language Models (LLMs), current techniques are primarily divided into two paradigms: semantic-based and symbolic-based. However, existing methods are generally limited by the "language confinement" problem when facing multilingual scenarios. Traditional semantic methods such as Retrieval Augmentation (RAG) suffer from language bias, their vector spaces tending to restrict queries to a subset of the source language, making it difficult to utilize the richer knowledge in other languages. Simultaneously, agentic frameworks passively reflect the user's native language during inference, missing globally available facts. On the other hand, symbolic methods (such as directly generating SPARQL) face significant difficulties in cross-language pattern alignment; directly parsing the source language into executable logical forms often leads to incorrect predicate mappings. Current models do not fully utilize consensus information between different languages ​​to eliminate noise and ambiguity introduced by retrieval, resulting in unsatisfactory retrieval performance in complex cross-language reasoning scenarios. To address this issue, this study proposes a multilingual question answering method based on semantic-symbolic knowledge consensus, aiming to improve the model's cross-lingual accurate question answering performance in multi-hop scenarios. Summary of the Invention

[0003] This invention provides a multilingual question answering method and apparatus based on semantic-symbolic knowledge consensus, in order to solve the problems of severe single-language retrieval bias, difficulty in cross-language pattern alignment, and easy execution errors in directly generating structured queries in complex reasoning scenarios in existing cross-language question answering technologies.

[0004] The technical solution of this invention is as follows: Firstly, this invention provides a multilingual question-answering method based on semantic-symbolic knowledge consensus, comprising:

[0005] Step 1: Construct multilingual reasoning paths: Use a large language model to diverge the source language query problem into parallel reasoning paths for multiple target languages, and extract the starting entities and relation sequences under each language path;

[0006] Step 2: Construct a multilingual evidence pool: Map the starting entities and relation sequences in the multilingual reasoning path to the knowledge base, retrieve and filter highly relevant candidate identifiers, and construct a multilingual entity evidence pool and a multilingual relation evidence pool.

[0007] Step 3: Execute the multilingual voting consensus mechanism: Use the voting model to perform cross-language consistency analysis on the multilingual entity evidence pool and the multilingual relationship evidence pool, eliminate language differences, and converge to obtain the optimal consensus entity identifier and consensus relationship identifier sequence;

[0008] Step 4: Query statement assembly and execution, and exception rollback: Generate a structured query template based on the source language query question and the corresponding reasoning path. Inject the consensus entity identifier and consensus relationship identifier into the structured query template to assemble an executable query statement, execute it in the knowledge base, and return the question and answer results.

[0009] Furthermore, the specific steps of Step 1 include:

[0010] Step 1.1: For a given set of target languages Using generative models Source language query problem In each target language Parallel generation of structured inference paths The reasoning path Starting entity and relation sequence Composed of various elements;

[0011] Step 1.2: Collect all structural components from the generated target language inference paths and construct a cross-language set of starting entity references. and the Jump Relationships Mentioned Sets .

[0012] Furthermore, the specific steps of Step 2 include:

[0013] Step 2.1: Utilize knowledge base retrieval tools Each of the following sets refers to the starting entity. and relational mention set Processing returns a coarse candidate set with metadata. and ;

[0014] Step 2.2: Using the selector model Evaluation of source language query issues The semantic alignment between the mention and the retrieved candidate set is used to independently select the optimal identifier and its metadata;

[0015] Step 2.3: Aggregate the output results of each language into the symbol space to form a multilingual entity evidence pool. Evidence pool for multilingual relationships .

[0016] Furthermore, the specific steps of Step 3 include:

[0017] Step 3.1: For the initial entity, address the source language query problem. and multilingual entity evidence pool Input to voting model In this process, the optimal consensus entity identifier is identified and output through cross-language consistency analysis. ;

[0018] Step 3.2: For the relation sequence, traverse each reasoning step. The corresponding multilingual relationship evidence pool Input to voting model In the process, the optimal consensus relation identifier for this hop is derived. .

[0019] Furthermore, the specific steps of Step 4 include:

[0020] Step 4.1: Utilizing the structure generation model Querying problems in the source language and source language reasoning path Generate parameterized query template functions based on the input context. ;

[0021] Step 4.2: Set the consensus entity identifier Consensus Relationship Identifier Sequence Injected into the corresponding slot of the query template, and assembled to generate a structured query statement. ;

[0022] Step 4.3: Execute the structured query statement using the execution tool. Submit to the knowledge base endpoint to retrieve the result set If the execution is successful, the result set will be... Return the result as the final answer; if an execution exception occurs or the result is empty, a fallback mechanism is triggered, utilizing a large language model to query the problem based on the source language. and source language reasoning path Generate natural language answers .

[0023] The present invention also provides a multilingual question answering device based on semantic-symbolic knowledge consensus, including a module for executing the multilingual question answering method based on semantic-symbolic knowledge consensus.

[0024] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the multilingual question answering method based on semantic-symbolic knowledge consensus.

[0025] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the multilingual question-answering method based on semantic-symbolic knowledge consensus.

[0026] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the multilingual question answering method based on semantic-symbolic knowledge consensus.

[0027] The beneficial effects of this invention are:

[0028] 1. It solves the problems of retrieval bias caused by a single language perspective and difficulty in aligning cross-language patterns due to the scarcity of labeled data in cross-language question answering;

[0029] 2. First, the monolingual query is diverged into a wide range of multilingual evidence in the semantic space and mapped to a structured candidate set. On this basis, the symbol space is used to eliminate language differences through a multilingual voting mechanism, and the collected information is converged into a robust consensus identifier, reducing the impact of language-specific noise.

[0030] 3. The method proposed in this invention can effectively utilize global multilingual knowledge to enhance the representation alignment capability for complex problems, thereby improving the performance of multi-hop cross-lingual question answering models. It has important theoretical and practical application value for cross-lingual information retrieval and knowledge question answering. Attached Figure Description

[0031] Figure 1 This is a model architecture diagram of the multilingual question-answering device based on semantic-symbolic knowledge consensus in this invention;

[0032] Figure 2 This is a flowchart. Detailed Implementation

[0033] Example 1, such as Figures 1-2 The diagram shown is a model architecture diagram and flowchart of the multilingual question answering device based on semantic-symbolic knowledge consensus in this invention. The multilingual question answering method based on semantic-symbolic knowledge consensus includes:

[0034] Step 1: Constructing multilingual reasoning paths: For source language query problems, the source language query problem is diverged into parallel reasoning paths for multiple target languages ​​using a large language model, and the starting entities and relation sequences under each language path are extracted.

[0035] In some feasible implementations, the specific steps of Step 1 are as follows:

[0036] Step 1.1: For a given set of target languages Using generative models Source language query problem In each target language Parallel generation of structured inference paths The reasoning path Starting entity and relation sequence It is composed of various elements, and the specific generation formula is as follows:

[0037]

[0038] Among them, symbols This represents a combination operator used to link sequences of relationships to form a reasoning trajectory;

[0039] Step 1.2: Collect all structural components from the generated target language inference paths and construct a cross-language set of starting entity references. and the Jump Relationships Mentioned Sets The above set is used to establish a multilingual reference base required for subsequent cross-language alignment.

[0040] Step 2: Construct a multilingual evidence pool: Map the starting entities and relation sequences in the multilingual reasoning path to the knowledge base, filter out highly relevant candidate identifiers through retrieval tools and selector models, and construct a multilingual entity evidence pool and a multilingual relation evidence pool.

[0041] In some feasible implementations, the specific steps of Step 2 are as follows:

[0042] Step 2.1: Utilize knowledge base retrieval tools The set of references to the starting entity are respectively and relational mention set Processing returns a coarse candidate set with metadata. and ;

[0043] Step 2.2: Using the selector model Evaluate the source language query problem The semantic alignment between specific mentions and the retrieved candidate set is used to independently select the optimal identifier and its metadata.

[0044]

[0045]

[0046] Step 2.3: Aggregate the output results of each language into the symbol space to form a multilingual entity evidence pool. Evidence pool for multilingual relationships .

[0047] Step 3: Execute the multilingual voting consensus mechanism: Use the voting model to perform cross-language consistency analysis on the multilingual entity evidence pool and the multilingual relationship evidence pool, eliminate language differences, and converge to obtain the optimal consensus entity identifier and consensus relationship identifier sequence;

[0048] In some feasible implementations, the specific steps of Step 3 are as follows:

[0049] Step 3.1: For the starting entity, query the source language question. and the multilingual entity evidence pool Input to voting model In this process, the optimal consensus entity identifier is identified and output through cross-language consistency analysis. :

[0050]

[0051] Step 3.2: For the relation sequence, traverse each reasoning step. The corresponding multilingual relationship evidence pool Input to voting model In the process, the optimal consensus relation identifier for this hop is derived. :

[0052]

[0053] Step 4: Query statement assembly and execution, and exception rollback: Generate a structured query template based on the source language query question and the corresponding reasoning path. Inject the consensus entity identifier and consensus relationship identifier into the structured query template to assemble an executable query statement, execute it in the knowledge base, and return the question and answer results.

[0054] In some feasible implementations, Step 4 includes:

[0055] Step 4.1: Utilizing the structure generation model Query the problem in the source language. and source language reasoning path Generate parameterized query template functions based on the input context. ;

[0056]

[0057] Step 4.2: Transfer the consensus entity identifier Consensus Relationship Identifier Sequence The data is strictly injected into the corresponding slots of the query template and assembled to generate a structured query statement. ;

[0058]

[0059] Step 4.3: Execute the structured query statement using the execution tool. Submit to the knowledge base endpoint to retrieve the result set If the execution is successful, the result set will be... Returned as the final answer; if an execution exception occurs or the result is empty, a fallback mechanism is triggered, using the model to query the question based on the source language. and source language inference paths that include explicit natural language references Generate natural language answers:

[0060]

[0061] The following are embodiments of the device of the present invention. The device of the present invention is used to execute the method implemented in the first embodiment of the present invention. For ease of explanation, only the parts related to the embodiments of the present invention are shown. For the parts not disclosed, please refer to the first embodiment of the present invention.

[0062] This invention also provides a multilingual question-answering device based on semantic-symbolic knowledge consensus, the device comprising:

[0063] The divergence and construction module is used to diverge the source language query problem into parallel reasoning paths for multiple target languages ​​using a large language model, and to extract the starting entities and relation sequences under each language path.

[0064] Evidence pool mapping module: used to map the starting entities and relation sequences in the multilingual reasoning path to the knowledge base, retrieve and filter highly relevant candidate identifiers, and construct multilingual entity evidence pool and multilingual relation evidence pool;

[0065] Voting consensus module: Used to perform cross-language consistency analysis on the multilingual entity evidence pool and multilingual relation evidence pool using a voting model, eliminate language differences, and converge to obtain the optimal consensus entity identifier and consensus relation identifier sequence;

[0066] The query execution and rollback module is used to generate a structured query template based on the source language query question and the corresponding reasoning path, inject the consensus entity identifier and consensus relationship identifier into the structured query template to assemble an executable query statement, execute it in the knowledge base and return the question and answer results.

[0067] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the multilingual question answering method based on semantic-symbolic knowledge consensus.

[0068] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the multilingual question-answering method based on semantic-symbolic knowledge consensus.

[0069] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the multilingual question answering method based on semantic-symbolic knowledge consensus.

[0070] To verify the effectiveness of the present invention, the following experiments were conducted:

[0071] The experiments in this invention are based on online API calls and open-source frameworks, using DeepSeek-V3 as the unified inference backbone network. For evaluation metrics, the multilingual knowledge evaluation benchmark dataset MLaKE is used. This dataset contains single-hop and multi-hop question-answering data for five languages: English (EN), French (FR), German (DE), Japanese (JA), and Chinese (ZH). The detailed data distribution of the experiments is shown in Table 1 below.

[0072] Table 1. Distribution of Experimental Data

[0073]

[0074] In the experiments, the model's performance was strictly evaluated according to the standard evaluation metric, Exact Match (EM). To verify the effectiveness of the proposed method in cross-language question answering scenarios, comparative experiments were conducted with existing generative large models, agent architectures, and symbolic methods. The comparison models included: DeepSeek-V3, GPT-4o, Gemini-2.5, ReAct, RAG, and SPARQL.

[0075] The comparative experimental results on the benchmark dataset are shown in Table 2:

[0076] Table 2 Comparison of experimental results for different models (mean Exact Match, unit: %)

[0077]

[0078] As can be seen from the comparative experimental results in Table 2, the question-answering performance of our method significantly outperforms all other comparative models in both single-hop and multi-hop evaluation metrics. Compared to the homologous structured generative baseline model SPARQL, our method achieves an average improvement of 5.11% in single-hop scenarios and a substantial average improvement of 10.73% in the more complex multi-hop scenarios. This is attributed to our method's ability to fully learn external knowledge information from multilingual knowledge graphs by constructing a multilingual evidence pool and a voting consensus mechanism, effectively compensating for the scarcity of non-English language labeled data and the difficulty of cross-language symbol alignment. Furthermore, the "divergence-mapping-convergence" mechanism designed in our method can significantly resist error accumulation in complex inference paths, and the backoff mechanism ensures a very high executability rate, ultimately comprehensively improving the performance of multilingual cross-language question answering.

[0079] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A multilingual question-answering method based on semantic-symbolic knowledge consensus, characterized in that, include: Step 1: Construct multilingual reasoning paths: Use a large language model to diverge the source language query problem into parallel reasoning paths for multiple target languages, and extract the starting entities and relation sequences under each language path; Step 2: Construct a multilingual evidence pool: Map the starting entities and relation sequences in the multilingual reasoning path to the knowledge base, retrieve and filter highly relevant candidate identifiers, and construct a multilingual entity evidence pool and a multilingual relation evidence pool. Step 3: Execute the multilingual voting consensus mechanism: Use the voting model to perform cross-language consistency analysis on the multilingual entity evidence pool and the multilingual relationship evidence pool, eliminate language differences, and converge to obtain the optimal consensus entity identifier and consensus relationship identifier sequence; Step 4: Query statement assembly and execution, and exception rollback: Generate a structured query template based on the source language query question and the corresponding reasoning path. Inject the consensus entity identifier and consensus relationship identifier into the structured query template to assemble an executable query statement, execute it in the knowledge base, and return the question and answer results.

2. The multilingual question-answering method based on semantic-symbolic knowledge consensus according to claim 1, characterized in that, The specific steps of Step 1 include: Step 1.1: For a given set of target languages Using generative models Source language query problem In each target language Parallel generation of structured inference paths The reasoning path Starting entity and relation sequence Composed of various elements; Step 1.2: Collect all structural components from the generated target language inference paths and construct a cross-language set of starting entity references. and the Jump Relationships Mentioned Sets .

3. The multilingual question-answering method based on semantic-symbolic knowledge consensus according to claim 1, characterized in that, The specific steps of Step 2 include: Step 2.1: Utilize knowledge base retrieval tools Each of the following sets refers to the starting entity. and relational mention set Processing returns a coarse candidate set with metadata. and ; Step 2.2: Using the selector model Evaluation of source language query issues The semantic alignment between the mention and the retrieved candidate set is used to independently select the optimal identifier and its metadata; Step 2.3: Aggregate the output results of each language into the symbol space to form a multilingual entity evidence pool. Evidence pool for multilingual relationships .

4. The multilingual question-answering method based on semantic-symbolic knowledge consensus according to claim 1, characterized in that, The specific steps of Step 3 include: Step 3.1: For the initial entity, address the source language query problem. and multilingual entity evidence pool Input to voting model In this process, the optimal consensus entity identifier is identified and output through cross-language consistency analysis. ; Step 3.2: For the relation sequence, traverse each reasoning step. The corresponding multilingual relationship evidence pool Input to voting model In the process, the optimal consensus relation identifier for this hop is derived. .

5. The multilingual question-answering method based on semantic-symbolic knowledge consensus according to claim 1, characterized in that, The specific steps of Step 4 include: Step 4.1: Utilizing the structure generation model Querying problems in the source language and source language reasoning path Generate parameterized query template functions based on the input context. ; Step 4.2: Set the consensus entity identifier Consensus Relationship Identifier Sequence Injected into the corresponding slot of the query template, and assembled to generate a structured query statement. ; Step 4.3: Execute the structured query statement using the execution tool. Submit to the knowledge base endpoint to retrieve the result set If the execution is successful, the result set will be... Return the result as the final answer; if an execution exception occurs or the result is empty, a fallback mechanism is triggered, utilizing a large language model to query the problem based on the source language. and source language reasoning path Generate natural language answers .

6. A multilingual question-answering device based on semantic-symbolic knowledge consensus, characterized in that, Includes a module for performing the multilingual question answering method based on semantic-symbolic knowledge consensus as described in any one of claims 1-5.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the multilingual question answering method based on semantic-symbolic knowledge consensus as described in any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multilingual question answering method based on semantic-symbolic knowledge consensus as described in any one of claims 1 to 5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the multilingual question answering method based on semantic-symbolic knowledge consensus as described in any one of claims 1 to 5.

Images