An open domain table question answering method and device based on evidence iterative expansion

Abstract

This invention provides an open-domain table question answering method and apparatus based on evidence iterative expansion. The method includes the following steps: encoding the original question to obtain a question encoding vector; in the first round of iteration calculation, calculating the first similarity between the question encoding vector and the evidence table vector of the corresponding evidence table in a preset database, and filtering the target evidence table based on the first similarity; in the calculation process of each iteration round of repeated iterations, concatenating the question encoding vector with the evidence table vector corresponding to the currently filtered target evidence table to form an iteration input vector, calculating the second similarity between the iteration input vector and the evidence table vector of the corresponding evidence table in the preset database, filtering the target evidence table based on the second similarity, and determining whether to end the repeated iteration based on the second similarity calculated in this iteration round; if it ends, outputting the original question and all filtered evidence tables to a large language model.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to an open-domain table question-answering method and apparatus based on evidence iterative expansion. Background Technology

[0002] Modern society has an ever-increasing demand for complex information, especially in knowledge-intensive tasks that require integrating multiple information sources to arrive at comprehensive conclusions. With the acceleration of digital transformation, massive amounts of information are stored in various databases and documents in the form of semi-structured tables and unstructured text. Users often need to synthesize information from multiple tables and text segments to find the answer to a question. For example, in the field of intelligent financial decision support, it is necessary to integrate tabular and textual data such as financial statements, supply chain contracts, and external market analysis reports; in the field of judicial adjudication, legal professionals need to perform cross-document, multi-step information deduction from vast amounts of legal provisions and historical precedents to construct complex chains of evidence; in the field of medical clinical inference, doctors often need to integrate information such as patients' electronic medical records, clinical trial results, and the latest medical guidelines to make complex diagnoses.

[0003] Tasks like these, which require integrating information from different structures across documents, typically require significant human resources for complex understanding and analysis. As a result, Open-domain Table Question Answering (OD-TQA) tasks have emerged, aiming to leverage intelligent automated systems to efficiently retrieve relevant tables and text segments from heterogeneous knowledge bases, and perform understanding and analysis to answer user questions.

[0004] Existing open-domain table question answering technologies typically include a retrieval phase and a model processing phase. In the retrieval phase, an intelligent model is needed to break down and analyze the user-inputted question. The results of this breakdown and analysis are then used to retrieve tables from the database. The retrieval results and the user-inputted question are then processed in the model processing phase, and the final question result is output through a large language model. However, existing technologies rely heavily on the accuracy of the agent's breakdown and analysis during the retrieval phase. If the accuracy is too low, it affects the retrieval results, which in turn affects the final question result. Furthermore, existing technologies consume significant computing power during the breakdown and analysis phase, resulting in substantial computational overhead. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide an open-domain table question-answering method based on evidence iterative expansion, in order to eliminate or improve one or more defects existing in the prior art.

[0006] One aspect of the present invention provides an open-domain table question-answering method based on evidence iterative expansion, the method comprising the following steps: Obtain the original question input by the user, and encode the original question to obtain a question encoding vector; In the first round of iterative calculation, the first similarity between the problem encoding vector and the evidence table vector of the corresponding evidence table in the preset database is calculated, and the target evidence table is selected based on the first similarity. In the calculation process of each iteration round of the repeated iteration, the problem encoding vector is concatenated with the evidence table vector corresponding to the evidence table of the currently selected target to form the iteration input vector. The second similarity between the iteration input vector and the evidence table vector of the corresponding evidence table in the preset database is calculated. Based on the second similarity, the evidence table of the target is selected again, and based on the second similarity calculated in this iteration round, it is determined whether to end the repeated iteration. If the repeated iterations end, the original question and the table of all filtered evidence are output to the large language model.

[0007] Using the above scheme, in the retrieval phase of open-domain table question answering technology, whether in the first iteration or subsequent iterations, this scheme only needs to calculate similarity and filter the target evidence tables based on similarity, without consuming a large amount of computing power. This scheme does not require the agent to consume computing power to analyze the original question input by the user, nor does it rely on the decomposition and analysis accuracy of the agent. Moreover, each iteration of this scheme relies on the evidence tables determined in the previous iteration as input, avoiding the errors caused by the decomposition and analysis of the agent, ensuring the relevance of the final evidence tables to the original question, and ensuring the accuracy of the question results output by the large language model.

[0008] In some embodiments of the present invention, in the step of calculating the first similarity between the question encoding vector and the evidence table vector of the corresponding evidence table in the preset database, the cosine similarity between the question encoding vector and the evidence table vector of each corresponding evidence table in the preset database is calculated.

[0009] In some embodiments of the present invention, the step of calculating a second similarity between the iterative input vector and the evidence table vector of the corresponding evidence table in a preset database, and then filtering the target evidence table again based on the second similarity, further includes: Based on the third similarity between the iterative input vector and the evidence table vector of each currently selected evidence table, an iterative decay value is calculated based on the third similarity between the iterative input vector and the evidence table vector of each currently selected evidence table. A corrected similarity is calculated based on the second similarity and the iterative decay value. The target evidence table is then selected based on the corrected similarity.

[0010] In some embodiments of the present invention, in the step of calculating the iteration decay value based on the third similarity between the iterative input vector and the evidence table vector of each currently selected evidence table, the iteration decay value is calculated using the following formula: in, Indicates the iterative decay value. This represents the preset hyperparameters. This represents the total number of evidence table vectors in the evidence set constructed from the evidence table vectors of the currently selected evidence tables. Represents any evidence table vector in the evidence set. Represents the iterative input vector. This indicates the calculation of the third similarity.

[0011] In some embodiments of the present invention, in the step of calculating the corrected similarity based on the second similarity and the iterative decay value, and screening the evidence table of the target based on the corrected similarity, the difference between the second similarity and the iterative decay value is calculated as the corrected similarity, and the evidence table corresponding to the largest preset number of corrected similarities in the current iteration round is taken as the evidence table of the target to be screened in the current round.

[0012] In some embodiments of the present invention, the repeated iteration processing step further includes, during the calculation of the first iteration round of repeated iteration, if the corrected similarity of the evidence table of any selected target is less than the average of all first similarities calculated in the first iteration round, then the original problem is determined to be a single-hop problem, the repeated iteration is terminated, and the original problem and the evidence table selected in the first iteration round are output to the large language model.

[0013] In some embodiments of the present invention, in the step of determining whether to end the repeated iteration based on the second similarity calculated in the current iteration round, during the calculation process of non-first iteration rounds of repeated iteration, the corrected similarity corresponding to the second similarity of the evidence table of each target is determined. For the evidence table of the selected target, the corrected similarity corresponding to the evidence table vector concatenated in the iterative input vector input during the calculation of the evidence table of the target is extracted, and the average value is calculated as the termination comparison value. The corrected similarity of the evidence table of the target is compared with the termination comparison value to determine whether to terminate the iteration path of the evidence table of the target.

[0014] In some embodiments of the present invention, in the step of determining whether to end the repeated iteration based on the second similarity calculated in the current iteration round, it is determined whether there is an unterminated iteration path. If there is, the repeated iteration continues; if not, the repeated iteration ends.

[0015] In some embodiments of the present invention, in the step of encoding the original question to obtain a question encoding vector, the original question is segmented into words, and each segmented word is encoded to obtain multiple word codes. The multiple word codes are input into an encoding model, which is provided with a Transformer layer, and the question encoding vector is output through the Transformer layer.

[0016] A second aspect of the present invention also provides an open-domain table question-answering apparatus based on evidence iterative expansion. The apparatus includes a computer device, the computer device including a processor and a memory, the memory storing computer instructions, and the processor executing the computer instructions stored in the memory. When the computer instructions are executed by the processor, the apparatus implements the steps of the method described above.

[0017] A third aspect of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned open-domain table question-answering method based on evidence iterative expansion.

[0018] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the text, or may be learned by practice of the invention. The objects and other advantages of the invention will become apparent from the description and the accompanying drawings.

[0019] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0020] The accompanying drawings, which are provided to further illustrate the invention and form part of this application, are not intended to limit the scope of the invention.

[0021] Figure 1 This is a schematic diagram of one implementation of the open-domain table question-answering method based on evidence iteration extension of this scheme; Figure 2 This is a schematic diagram of the processing architecture of this solution; Figure 3 This is a schematic diagram illustrating a processing example of this solution. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0023] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0024] Specifically, this solution focuses on multi-hop questions in table-based question answering. Multi-hop questions require the system to not only perform simple information retrieval but also to traverse multiple pieces of evidence, executing multiple logical reasoning steps across these corpora to arrive at the final answer. For example, answering a three-hop question might involve first extracting an entity from a text paragraph, using that entity to retrieve a table, and then associating the data in the table with other textual information for calculation or inference. This heterogeneity and multi-step nature makes traditional single-retrieval methods inadequate. Because multi-hop reasoning relies on the association and integration of cross-modal information (structured tabular data and unstructured text), it is more complex than pure text multi-hop question answering, placing stringent demands on the system's robustness and the accuracy of its reasoning paths.

[0025] Existing solutions primarily focus on optimizing the retrieval and reasoning processes. In the retrieval phase, traditional methods typically rely on semantic similarity-based retrieval, using pre-trained models as encoders to encode the question and the corpus to be retrieved, and representing relevance by calculating encoding similarity. Some methods aim to improve the cross-modal joint representation of tables and text for joint retrieval of heterogeneous data, such as preprocessing highly related tables and texts into fusion blocks, or converting table-modal information into text-modal information before encoding to enhance information interaction between the two modalities. However, for complex multi-hop problems, the original question alone is insufficient to recall all relevant evidence corpora. To guide multi-step retrieval, some methods have begun to utilize Large Language Models (LLMs) for dynamic query modification. For example, the Query2Doc method first generates a pseudo-document describing the required information based on the query, then appends it to the query for expanded retrieval, but this approach does not utilize information from the knowledge base for guidance. The MURRE method, to avoid easily retrieving similar but irrelevant tables in subsequent retrieval steps, proposes removing the portion covered by retrieved information from the original question after each retrieval step before proceeding to the next step.

[0026] In the inference phase, existing methods focus on iterative processing of tabular data. For example, ReacTable and Chain-of-Table methods use iterative modification to perform simple decomposition, merging and sorting operations on the retrieved tables using LLM, in order to make the tabular data sufficiently standardized and semantically clear, and to generate the final answer.

[0027] Current mainstream methods leverage the world knowledge of LLMs to decompose the original problem, then use the decomposed sub-problems to retrieve relevant evidence fragments, and use the LLMs to read this evidence to obtain answers to the sub-problems, thereby posing new sub-problems. By progressively answering all sub-problems, the answer to the original multi-hop problem is finally obtained. However, this decomposition method relies excessively on the internal knowledge of LLMs, while the knowledge in external corpora may differ from the knowledge learned by LLMs during pre-training, and the key procedural information required to answer multi-hop problems is often contained in external knowledge bases. If problem decomposition is performed too early and explicitly, the system may fail to recall evidence that is not directly relevant to the original problem but is crucial for multi-hop reasoning. This may ultimately cause the model to proceed along incorrect reasoning paths and fail to arrive at the correct answer. Figure 2 An example is provided: if an incorrect table is retrieved, LLMs will break down the sub-questions that the corpus cannot answer. Green arrows indicate the correct solution path.

[0028] Furthermore, semi-structured tabular data often contains various non-standard data formats. Existing methods often directly input this data with "structural noise" into LLMs for inference, causing confusion for LLMs in understanding the table structure and semantics, which seriously affects the accuracy of the final inference.

[0029] Based on the shortcomings of the prior art described above, the technical problem solved by the present invention can be summarized as follows: 1. Bridging the gap between external knowledge bases and LLMs internal knowledge to solve the problem of error propagation caused by explicit problem decomposition; 2. Standardize the data format of semi-structured tabular corpora to solve the problem of LLMs being confused due to non-standard data formats.

[0030] like Figure 1 As shown, this invention proposes an open-domain table question-answering method based on evidence iterative expansion. The steps of the method include: Step S100: Obtain the original question input by the user, and encode the original question to obtain a question encoding vector; Step S200: In the first round of iteration calculation, calculate the first similarity between the problem encoding vector and the evidence table vector of the corresponding evidence table in the preset database, and filter the target evidence table based on the first similarity. Step S300: In the calculation process of each iteration round of the repeated iteration, the problem encoding vector and the evidence table vector corresponding to the evidence table of the currently selected target are concatenated to form the iteration input vector. The second similarity between the iteration input vector and the evidence table vector of the corresponding evidence table in the preset database is calculated. Based on the second similarity, the evidence table of the target is selected again. Based on the second similarity calculated in this iteration round, it is determined whether to end the repeated iteration. In the specific implementation process, in the steps of filtering the evidence table of the target based on the first similarity and filtering the evidence table of the target again based on the second similarity, the evidence table corresponding to the first k largest first similarity or second similarity is used as the evidence table of the target.

[0031] In the specific implementation process, the method also includes table refinement processing. In the table refinement processing, the problem of mismatch between the table header and content of the evidence table of the target is solved by expanding / correcting the table header. The method of supplementing semantic annotations in the cell content (such as annotating "[nationality]" after the coach's name) or constructing a "table header-content" mapping dictionary to assist the model understanding is also within the protection scope of this solution.

[0032] Specifically, in the table refinement process, a table refiner cleans and corrects three common types of non-standard table data formats, transforming the original target evidence table into a standard database table format with a clear structure and explicit semantics. This improves the model's understanding and reasoning ability regarding the table content. The first type of problem is inconsistent data representation within columns. For example, as shown in Table 1, the "Ticket Price" column mixes British pounds and euros, or a column contains both numerical values ​​and strings. This inconsistency confuses the model's judgment of column data types, hindering effective feature extraction and numerical calculations. For instance, if the model cannot unify currency units, it cannot correctly perform price comparisons or summary calculations. An improvement method is to use LLM to identify the type and unit of all values ​​in the column and unify them to the same standard format, such as converting all currencies to the base currency or splitting mixed types into multiple semantically consistent sub-columns. The second type of problem is a mismatch between column headers and cell content. For example, the "Coach" column includes nationality information in addition to names, but the header does not explicitly state this. This ambiguity causes the model to miss key semantics when constructing the "column-value" mapping, affecting the correct establishment of multi-column relationships. One improvement is to expand or modify the table headers to accurately reflect the semantic scope of the cell content. For example, splitting "Coach" into two columns, "Coach Name" and "Coach Nationality," clarifies the semantic boundaries of each column. A third type of problem is the ineffective utilization of semantically meaningful column order. Although column order is generally irrelevant in database theory, in some unstructured document tables, there may be semantic relationships between adjacent columns, such as the "Score" column having the home team's score on the left and the away team's score on the right. If the model ignores this positional semantics, it may misunderstand the meaning of the data. An improvement method is to identify and explicitly label the semantic relationships between columns using a large language model. For example, retaining or reconstructing the semantically meaningful column order when generating sub-tables, or adding comments in the table headers explaining the dependencies between columns. Furthermore, the original table often contains a large amount of row and column information irrelevant to the current sub-problem. Directly inputting it into the model not only introduces noise and interferes with the model's focus on key information but also crowds out space for valid evidence due to context length limitations. By extracting sub-tables closely related to specific sub-problems, the system can precisely guide the model's attention to the core data, effectively shielding it from interference from irrelevant content and providing highly relevant and pure structured evidence for multi-step reasoning.

[0033] Table 1 The significance of standardized data formats in the entire system lies in their role as a bridge between retrieval and question answering, ensuring that the granularity and quality of evidence align with the needs of multi-step reasoning. By extracting sub-tables related to sub-questions, standardizing data representation, correcting table header semantics, and retaining key structural information, the system can provide accurate and clean evidence support for each step of reasoning.

[0034] In step S400, if the repeated iterations end, the original question and the table of all filtered evidence are output to the large language model.

[0035] Based on the above, this solution comprises two modules: iterative evidence expansion and table refinement. We will first introduce the iterative evidence expansion module. Given an original problem... and a corpus The corpus contains several tables and text segments, namely... ,in and These represent a table database and a text database, respectively. Open-domain table-based question answering requires a retrieval system to first extract relevant evidence from a large corpus, and then a reader to combine this external knowledge to generate the final answer. The iterative expansion retrieval method employs a tree structure for multiple rounds of retrieval to comprehensively explore possible reasoning paths (such as...). Figure 3 (As shown). First, the user's original question... As the initial search query In the Round, for each path (branch) in the tree Its input is the query generated in the previous round. It includes the original question and all evidence retrieved along that path. The retriever uses... Retrieve the top-k relevant corpora from the knowledge base. Further selection before The corpus is formatted to generate standardized evidence fragments. And assign a relevance score to each piece of evidence. , It refers to the number of branches. Each new piece of evidence... It will be related to the previous query from which it came. The data is concatenated to form a new query for the next round of retrieval. This process creates new branches on the tree. In this way, the retrieval unit explores multiple reasoning paths consisting of sequences of evidence in parallel. Finally, it selects the path with the highest average score among all paths and provides all the evidence it collects to the downstream reader to generate the answer. Using only a greedy strategy can easily lead to local optima; some evidence may start with a low score, but it might contain clues to the next crucial piece of evidence, allowing for the retrieval of highly valuable evidence later, resulting in a better overall score and final outcome for the entire path. The tree structure, by preserving and expanding multiple potential branches in parallel at each round, prevents such reasoning paths from being prematurely abandoned, thus finding better evidence combinations globally and ultimately improving the accuracy of the answer.

[0036] In traditional sub-problem decomposition processes, when the retrieved evidence is insufficient or cannot directly support the reasoning, the large language model is forced to guess based on its internally parameterized world knowledge, which is highly prone to factual errors or fabrication. This solution uniformly submits the evidence obtained in the retrieval stage to the subsequent large model for reading comprehension. This eliminates the need for pre-defined fixed sub-problem decomposition patterns and naturally adapts to complex dependencies between sub-problems. Expanding based on the actual retrieved evidence avoids generating explicit, potentially erroneous sub-problems, thereby reducing the cascading negative impact of early erroneous decisions on subsequent steps. The question-answering model, based on coherent, complete, and logically connected context, can generate more accurate and informative answers.

[0037] Using the above scheme, in the retrieval phase of open-domain table question answering technology, whether in the first iteration or subsequent iterations, this scheme only needs to calculate similarity and filter the target evidence tables based on similarity, without consuming a large amount of computing power. This scheme does not require the agent to consume computing power to analyze the original question input by the user, nor does it rely on the decomposition and analysis accuracy of the agent. Moreover, each iteration of this scheme relies on the evidence tables determined in the previous iteration as input, avoiding the errors caused by the decomposition and analysis of the agent, ensuring the relevance of the final evidence tables to the original question, and ensuring the accuracy of the question results output by the large language model.

[0038] In some embodiments of the present invention, in the step of calculating the first similarity between the question encoding vector and the evidence table vector of the corresponding evidence table in the preset database, the cosine similarity between the question encoding vector and the evidence table vector of each corresponding evidence table in the preset database is calculated.

[0039] In some embodiments of the present invention, the step of calculating a second similarity between the iterative input vector and the evidence table vector of the corresponding evidence table in a preset database, and then filtering the target evidence table again based on the second similarity, further includes: Based on the third similarity between the iterative input vector and the evidence table vector of each currently selected evidence table, an iterative decay value is calculated based on the third similarity between the iterative input vector and the evidence table vector of each currently selected evidence table. A corrected similarity is calculated based on the second similarity and the iterative decay value. The target evidence table is then selected based on the corrected similarity.

[0040] In practice, the first similarity, the second similarity, and the third similarity can all be calculated using the cosine similarity method.

[0041] In some embodiments of the present invention, in the step of calculating the iteration decay value based on the third similarity between the iterative input vector and the evidence table vector of each currently selected evidence table, the iteration decay value is calculated using the following formula: in, Indicates the iterative decay value. This represents the preset hyperparameters. This represents the total number of evidence table vectors in the evidence set constructed from the evidence table vectors of the currently selected evidence tables. Represents any evidence table vector in the evidence set. Represents the iterative input vector. This indicates the calculation of the third similarity.

[0042] Using the above scheme, in the calculation of corrected similarity, candidate corpora are required to be similar to the current query. Furthermore, by subtracting their average similarity to historical evidence, candidates that are overly similar to historical evidence but weakly related to the original question are penalized, thus forcing the retrieval engine to focus more on information relevant to the original question. This is achieved by adjusting preset hyperparameters. The balance can be precisely controlled in order to highlight the original problem.

[0043] In some embodiments of the present invention, in the step of calculating the corrected similarity based on the second similarity and the iterative decay value, and screening the evidence table of the target based on the corrected similarity, the difference between the second similarity and the iterative decay value is calculated as the corrected similarity, and the evidence table corresponding to the largest preset number of corrected similarities in the current iteration round is taken as the evidence table of the target to be screened in the current round.

[0044] In the specific implementation process, in the step of calculating the corrected similarity based on the second similarity and the iterative decay value, the difference between the second similarity and the iterative decay value is calculated as the corrected similarity.

[0045] In some embodiments of the present invention, the repeated iteration processing step further includes, during the calculation of the first iteration round of repeated iteration, if the corrected similarity of the evidence table of any selected target is less than the average of all first similarities calculated in the first iteration round, then the original problem is determined to be a single-hop problem, the repeated iteration is terminated, and the original problem and the evidence table selected in the first iteration round are output to the large language model.

[0046] In some embodiments of the present invention, in the step of determining whether to end the repeated iteration based on the second similarity calculated in the current iteration round, during the calculation process of non-first iteration rounds of repeated iteration, the corrected similarity corresponding to the second similarity of the evidence table of each target is determined. For the evidence table of the selected target, the corrected similarity corresponding to the evidence table vector concatenated in the iterative input vector input during the calculation of the evidence table of the target is extracted, and the average value is calculated as the termination comparison value. The corrected similarity of the evidence table of the target is compared with the termination comparison value to determine whether to terminate the iteration path of the evidence table of the target.

[0047] In the specific implementation process, the iterative path of the evidence table of the target is to further concatenate the evidence table vector corresponding to the evidence table of the target, and perform the extended path as in step S300.

[0048] The core idea of ​​the stopping mechanism in this scheme is that there is a significant score difference between useful and useless evidence. Therefore, if the score of new evidence drops significantly, it indicates that the information retrieved along the current path is no longer of substantial help in solving the problem and may even introduce noise. In this case, the stopping mechanism will decisively stop expanding that path, thus concentrating limited computational resources on branches with consistently valid scores. This not only effectively prevents wasting resources in the wrong direction but also avoids irrelevant information interfering with the generation of the final answer. The iterative expansion retrieval strategy of this scheme decouples the retrieval and question-answering processes, eliminating the need for subsequent question-answering models to be specifically designed or use complex prompts to coordinate with the retrieval process. This division of labor makes the system more flexible and robust overall, as improvements to the retrieval module (such as better ranking algorithms and superior stopping criteria) can directly improve the quality of the final answer.

[0049] In the iteratively expanded retrieval framework described above, each round of query consists of the original question and all retrieved evidence. As the number of rounds increases, the length of the evidence portion far exceeds the original question. For a similarity-based retrieval system, longer text typically implies richer semantic signals and greater influence. This leads the retrieval system to unconsciously focus more on the semantics of the evidence portion of the query when calculating similarity, relatively neglecting the fundamental original question. This imbalance in semantic weighting can cause a shift in the retrieval path. The system might retrieve documents highly relevant to the evidence from the previous round, but these documents are not strongly related to the core requirements of answering the original question.

[0050] In some embodiments of the present invention, in the step of determining whether to end the repeated iteration based on the second similarity calculated in the current iteration round, it is determined whether there is an unterminated iteration path. If there is, the repeated iteration continues; if not, the repeated iteration ends.

[0051] In some embodiments of the present invention, in the step of encoding the original question to obtain a question encoding vector, the original question is segmented into words, and each segmented word is encoded to obtain multiple word codes. The multiple word codes are input into an encoding model, which is provided with a Transformer layer, and the question encoding vector is output through the Transformer layer.

[0052] The beneficial effects of this plan include: 1. This proposal suggests a retrieval method based on Iterative Evidence Expansion (IEE) to achieve implicit question decomposition in open-domain table question answering. In each round of retrieval, this method iteratively concatenates the evidence fragments retrieved in the previous round into the original query to form a new query. This strategy naturally guides the retrieval engine to find the entities or contextual information needed for the next reasoning step by utilizing the retrieved evidence corpus. This evidence-driven approach replaces explicit question decomposition, effectively mitigating the risk of errors propagating along the reasoning chain.

[0053] 2. This solution summarizes and categorizes common non-standard table data formats, and uses the command-following capability of LLMs to standardize the retrieved table data, converting the data into a standard format, ensuring that the evidence provided to the final answer generation module is clean, structured, and easy to understand; 3. The stopping criteria of this scheme are based on the comparison between the score of new evidence and the average score of historical evidence. The stopping judgment method based on information gain quantification of the newly added effective information of new evidence or calculation of the proportion of the new evidence covering the key attributes of the original problem is also within the scope of protection of this patent. 4. In the refinement of tables, this solution solves the problem of mismatch between table headers and content by expanding / correcting table headers. The method of supplementing semantic annotations in cell content (such as adding "[nationality]" after the coach's name) or constructing a "table header-content" mapping dictionary to assist the model's understanding is also within the scope of protection of this patent. 5. The method of selecting corpora for formatting during retrieval and using dynamic threshold filtering based on relevance scores (such as retaining only corpora with scores higher than a preset threshold) or clustering deduplication (merging similar corpora) to screen valid evidence is also within the scope of protection of this patent.

[0054] This invention also provides an open-domain table question-answering device based on evidence iterative expansion. The device includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor executes the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps of the method described above.

[0055] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned open-domain table question-answering method based on evidence iterative expansion. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0056] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0057] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0058] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0059] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An open-domain table-based question-answering method based on iterative evidence expansion, characterized in that, The steps of the method include: Obtain the original question input by the user, and encode the original question to obtain a question encoding vector; In the first round of iterative calculation, the first similarity between the problem encoding vector and the evidence table vector of the corresponding evidence table in the preset database is calculated, and the target evidence table is selected based on the first similarity. In the calculation process of each iteration round of the repeated iteration, the problem encoding vector is concatenated with the evidence table vector corresponding to the evidence table of the currently selected target to form the iteration input vector. The second similarity between the iteration input vector and the evidence table vector of the corresponding evidence table in the preset database is calculated. Based on the second similarity, the evidence table of the target is selected again, and based on the second similarity calculated in this iteration round, it is determined whether to end the repeated iteration. If the repeated iterations end, the original question and the table of all filtered evidence are output to the large language model.

2. The open-domain table question-answering method based on evidence iterative expansion according to claim 1, characterized in that, In the step of calculating the first similarity between the question encoding vector and the evidence table vector of the corresponding evidence table in the preset database, the cosine similarity between the question encoding vector and the evidence table vector of each corresponding evidence table in the preset database is calculated.

3. The open-domain table question-answering method based on evidence iterative expansion according to claim 1 or 2, characterized in that, The step of calculating the second similarity between the iterative input vector and the evidence table vector of the corresponding evidence table in the preset database, and then filtering the target evidence table again based on the second similarity, further includes: Based on the third similarity between the iterative input vector and the evidence table vector of each currently selected evidence table, an iterative decay value is calculated based on the third similarity between the iterative input vector and the evidence table vector of each currently selected evidence table. A corrected similarity is calculated based on the second similarity and the iterative decay value. The target evidence table is then selected based on the corrected similarity.

4. The open-domain table question-answering method based on evidence iterative expansion according to claim 3, characterized in that, In the step of calculating the iteration decay value based on the third similarity between the iterative input vector and the evidence table vector of each currently selected evidence table, the iteration decay value is calculated using the following formula: in, Indicates the iterative decay value. This represents the preset hyperparameters. This represents the total number of evidence table vectors in the evidence set constructed from the evidence table vectors of the currently selected evidence tables. Represents any evidence table vector in the evidence set. Represents the iterative input vector. This indicates the calculation of the third similarity.

5. The open-domain table question-answering method based on evidence iterative expansion according to claim 3, characterized in that, In the step of calculating the corrected similarity based on the second similarity and the iterative decay value, and screening the evidence table of the target based on the corrected similarity, the difference between the second similarity and the iterative decay value is calculated as the corrected similarity, and the evidence table corresponding to the largest preset number of corrected similarities in this iteration round is taken as the evidence table of the target to be screened in this round.

6. The open-domain table question-answering method based on evidence iterative expansion according to claim 3, characterized in that, The repeated iteration processing steps also include, in the calculation process of the first iteration round of repeated iteration, if the corrected similarity of the evidence table of any selected target is less than the average of all first similarities calculated in the first iteration round, then the original problem is determined to be a single-hop problem, the repeated iteration is terminated, and the original problem and the evidence table selected in the first iteration round are output to the large language model.

7. The open-domain table question-answering method based on evidence iterative expansion according to claim 3, characterized in that, In the step of determining whether to end the repeated iteration based on the second similarity calculated in this iteration round, during the calculation process of the non-first iteration round of the repeated iteration, the corrected similarity corresponding to the second similarity of the evidence table of each target is determined. For the evidence table of the selected target, the corrected similarity corresponding to the evidence table vector concatenated in the iterative input vector input during the calculation of the evidence table of the target is extracted, and the average value is calculated as the termination comparison value. The corrected similarity of the evidence table of the target is compared with the termination comparison value to determine whether to terminate the iteration path of the evidence table of the target.

8. The open-domain table question-answering method based on evidence iterative expansion according to claim 7, characterized in that, In the step of determining whether to end the repeated iteration based on the second similarity calculated in this iteration round, it is determined whether there is an unterminated iteration path. If there is, the repeated iteration continues; if not, the repeated iteration ends.

9. The open-domain table question-answering method based on evidence iterative expansion according to claim 1, characterized in that, In the step of encoding the original question to obtain the question encoding vector, the original question is segmented into words, and each segmented word is encoded to obtain multiple word codes. The multiple word codes are input into the encoding model, which is equipped with a Transformer layer, and the question encoding vector is output through the Transformer layer.

10. An open-domain table-based question-answering device based on evidence iterative expansion, characterized in that, The device includes a computer device, which includes a processor and a memory, wherein computer instructions are stored in the memory, and the processor is configured to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps of the method as described in any one of claims 1 to 9.

Images