Table question and answer method and device based on large language model, medium and program product

Abstract

The invention discloses a table question and answer method and device based on a large language model, a medium and a program product, and the method comprises the steps: firstly, receiving a table file uploaded by a user through a file uploading interface, vectorizing the content of the table file, and storing the vectorized content into a vector library; after a question request sent by a user is received, determining a target column corresponding to the question request in the table file by using the large language model; determining a target row corresponding to the question request according to vector similarity between character strings obtained by splicing the question request and the index column of each row of the table; constructing a question sub-table corresponding to the question request by using the target row and the target column; and finally, based on a large language model, according to the question request and the question sub-table, outputting a reply result. According to the method and the device, the large language model can accurately answer the large-volume table content under the condition that the maximum processing length is limited.

Description

Technical Field

[0001] The present invention relates to the field of large model technology, and in particular to a table question-answering method, device, medium and program product based on a large language model. Background Art

[0002] A large language model refers to a deep learning neural network model for natural language processing with a parameter count of over one billion, represented by Chat-GPT. It is widely used due to its significantly better instruction followability, language fluency, and functional diversity than traditional deep learning neural network models.

[0003] Tabular question answering is a key area of ​​large language models. However, large models perform significantly worse on tabular questions than plain text. The main reason is that text is one-dimensional time series information, while tables are two-dimensional (information needs to be read vertically by columns), have order invariance (exchanging two columns or two rows does not affect the understanding of the table content), and this information is clearly lacking in the training data.

[0004] The problems that exist in the actual application of large models in table question answering are as follows: the large language model will set a maximum length limit that the model input and output can handle during the pre-training stage. When the table is too large (such as hundreds or thousands of rows, dozens of columns), the large language model cannot read the entire table content every time it answers a question, resulting in incorrect information and inconsistent answers with the table; even if the length limit allows, the large language model will also have problems such as slow response and hallucination replies caused by interference from irrelevant information due to the long input length. In the prior art, some people have proposed that in order to reduce interference information, the large language model should first observe the table or the first few rows of the table and select the required rows or columns, but this method is still limited by the length of the observed table or part of the table, and cannot be handled once the limit is exceeded.

[0005] When faced with complex problems, existing methods usually adopt the method of problem decomposition, breaking down complex problems into small problems one by one and sending them to the large model for answer step by step. However, this type of method usually requires passing the previous questions and answers back to the input part of the large language model when processing subsequent questions, resulting in the large language model still facing severe length limitations when processing complex problems.

[0006] In the service segment of large language models, users usually ask common questions in fields other than tables at the same front-end interactive entrance. Large language models are required to automatically identify and switch the basis for answers between table information and domain knowledge, and be able to provide a fallback response to questions that cannot be answered, while reducing the occurrence of hallucinations.

[0007] There is currently no technical solution in the existing methods that can support users to upload tables and domain knowledge at the same time in a single interactive entrance, and to ask questions about the table content and domain knowledge at the same front-end entrance, which means that it lacks convenience and ease of use for users.

[0008] Therefore, it is urgent to invent a table question and answer method based on a large language model to solve the problems that existing table question and answer methods are unable to accurately answer large-volume table contents under limited maximum processing length, are unable to break down and answer complex questions, and are unable to automatically identify table-type information and domain knowledge and switch answering methods. Summary of the invention

[0009] In view of this, embodiments of the present invention provide a table question answering method, device, medium and program product based on a large language model, which at least partially solve the problems existing in the prior art.

[0010] Other features and advantages of the present invention will become apparent from the following detailed description, or may be learned in part by practice of the present invention.

[0011] In order to achieve the above purpose, the embodiment of the present invention provides the following technical solutions:

[0012] According to a first aspect of an embodiment of the present invention, a table question answering method based on a large language model is provided, the method comprising:

[0013] Utilize the file upload interface to receive the table file uploaded by the user, vectorize the content of the table file and store it in the vector library;

[0014] After receiving a question request sent by a user, determine a target column corresponding to the question request in a table file using a large language model;

[0015] Determine a target row corresponding to the question request according to the vector similarity between the string obtained by concatenating the question request and the index columns of each row of the table;

[0016] Using the target row and the target column, construct a question subtable corresponding to the question request;

[0017] Based on the large language model, a reply result is output according to the question request and the question sub-table.

[0018] Furthermore, the table file is a file containing a table or an Excel file containing multiple form pages.

[0019] Further, using the file upload interface to receive the table file uploaded by the user includes:

[0020] Use the file upload interface to receive the table file uploaded by the user;

[0021] Frequently used questions and answers in the same field as the table file are listed separately as question-answer pairs in a question-answer pair table.

[0022] Further, after receiving the question request sent by the user, the target column corresponding to the question request is determined in the table file using the large language model, including:

[0023] After receiving the question request sent by the user, the target column corresponding to the question request in the table is determined according to the table header of the table file using the large language model;

[0024] If the target column is located in the question-answer pair table, the question is answered using a common question-answer process based on domain knowledge.

[0025] Further, based on the large language model, according to the question request and the question sub-table, a reply result is output, including:

[0026] Determine whether the question request is a complex question that cannot be answered through a single round of interaction or a single sub-table;

[0027] If the question request is not a complex question that cannot be answered through a single round of interaction or a single sub-table, then based on the large language model, output a response result according to the question request and the question sub-table;

[0028] If the question request is a complex question that cannot be answered through a single round of interaction or a single sub-table, the question request is decomposed using a large language model to obtain at least two sub-questions;

[0029] For each of the sub-questions, generate a corresponding sub-question answer;

[0030] All sub-question responses are concatenated to get the final response.

[0031] Furthermore, if the question request is a complex question that cannot be answered through a single round of interaction or a single sub-table, the question request is decomposed using a large language model to obtain at least two sub-questions, and a corresponding sub-question answer is generated for each of the sub-questions, including:

[0032] If the question request is a complex question that cannot be answered through a single round of interaction or a single sub-table, the question request will be treated as a sub-question;

[0033] Generate sub-question answers and follow-up question marks corresponding to the sub-questions according to the question sub-table;

[0034] Execute a follow-up questioning process based on the follow-up questioning mark to generate a corresponding follow-up question for the sub-question;

[0035] The follow-up sub-question is sent as a new question request to the large language model to obtain a sub-question answer corresponding to the follow-up sub-question.

[0036] Furthermore, the content of the table file is vectorized and stored in a vector library, including:

[0037] Parse and extract the received table files, vectorize the contents of each table and store them in the vector library;

[0038] For each table, determine the index column of the table based on the table header.

[0039] According to a second aspect of an embodiment of the present invention, there is provided a table question answering device based on a large language model, the device comprising: a processor and a memory;

[0040] The memory is used to store one or more program instructions;

[0041] The processor is used to run one or more program instructions to execute the steps of a table question answering method based on a large language model as described in any of the above items.

[0042] According to a third aspect of an embodiment of the present invention, a computer-readable storage medium is provided, characterized in that a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the steps of a table question answering method based on a large language model as described in any one of the above items are implemented.

[0043] According to a fourth aspect of an embodiment of the present invention, a computer program product is provided, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, which, when executed by a computer, enable the computer to implement the steps of a table question answering method based on a large language model as described in any one of the above items.

[0044] The embodiments of the present invention have the following advantages:

[0045] The embodiment of the present invention significantly reduces the scope of the table information used by the large language model when answering by extracting question sub-tables, effectively alleviating the computational pressure and hallucination answering problems caused by length restrictions, so that the large language model can accurately answer large-volume table contents under the condition of limited maximum processing length. In addition, by breaking down complex questions into several sub-questions and answering them in sequence, after completing one answer, a special mark is used to trigger the follow-up question process, and finally the answers to all sub-questions are spliced ​​and returned to the user, so that the large language model can complete the problem decomposition and accurately answer the complex questions under the length restriction. At the same time, by storing the domain-specific knowledge outside the table separately in a table, automatic switching between the answering method based on the table content and the answering method based on the domain knowledge (frequently asked questions) is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the implementation methods of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the implementation methods or the description of the prior art. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other implementation drawings can be derived from the provided drawings without creative work.

[0047] Figure 1 A flowchart of a table question answering method based on a large language model provided by an embodiment of the present invention;

[0048] Figure 2 A schematic block diagram of the principles of a table question answering method based on a large language model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0050] In addition, the terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus that includes a series of steps or elements is not necessarily limited to those steps or elements explicitly listed, but may include other steps or elements not explicitly listed or inherent to such process, method, product, or apparatus.

[0051] Figure 1The figure shows a flow chart of a table question answering method based on a large language model according to an embodiment of the present invention.

[0052] like Figure 1 As shown, the table question answering method based on a large language model according to an embodiment of the present invention may include step S100, step S200, step S300, step S400 and step S500.

[0053] refer to Figure 2 In step S100, the table file uploaded by the user is received using the file upload interface, and the content of the table file is vectorized and stored in the vector library.

[0054] Specifically, the above steps include:

[0055] The file upload interface is used to receive a table file uploaded by a user, wherein the table file is a file containing a table or an Excel file containing multiple form pages.

[0056] The received table files are parsed and extracted, and the contents of each extracted table are vectorized and stored in vector libraries respectively, isolated from each other.

[0057] For each table, extract the table header, and let the user specify the vector index column of the table on the front-end upload page. For example, select columns A, B, and D of a table. When determining the target row in the subsequent sub-table construction step for the question request raised by the user, use the string concatenated from the contents of columns A, B, and D of each row of the table to calculate the vector similarity with the user's question, and take the row with the highest similarity as the target row.

[0058] If the user has the need to ask questions and answers about the knowledge in this field (outside the table content), the common questions and answers in the same field as the table file can be listed separately as question and answer pairs in a question and answer pair table. For example, the table headers of the above question and answer pair table are "Questions" and "Answers", thereby supporting answers to frequently asked questions about tables and fields in a unified paradigm.

[0059] The embodiment of the present invention lists the commonly used questions and answers in the same field as the table file as question-answer pairs separately as a question-answer pair table, so that the large language model can automatically recognize the field knowledge provided by the user and give accurate answers at the same time.

[0060] Next, in step S200 to step S400, after receiving the question request sent by the user, the large language model is first used to determine the target column corresponding to the question request according to the header of the table file, and if the target column is located in the above-mentioned question and answer table, the common question and answer process of domain knowledge is used to answer the question, and then the target row corresponding to the question request is determined based on the vector similarity between the question request and the index column of each row of the table, and finally the target row and target column are used to construct the question sub-table corresponding to the question request.

[0061] Specifically, the above steps include:

[0062] The user sends a question request on the front end. After the large language model receives the question request sent by the user, it first uses the language understanding ability of the large language model to determine the target column corresponding to the question request in the table file, and then calculates the vector similarity between the question request and the string obtained by concatenating the index columns of each row of the table. Based on the similarity calculation result, the target row corresponding to the question request is determined, and finally the above target row and the above target column are used to construct the question subtable corresponding to the question request.

[0063] For example, the table file looks like this:

[0064] “[Name, gender, age, height, weight, hometown, self-introduction, past experience]

[0065] [Xiao Zhang, female, 28, 165cm, 55cm, Beijing, 200 words omitted, 500 words omitted]

[0066] ……(Abbreviated 500 lines)

[0067] [Xiao Liu, male, 21, 185cm, 75cm, Shanghai, 300 words or less, 600 words or less]

[0068] [Xiao Zhao, female, 29, 175cm, 52cm, Guangdong, 100 words omitted, 400 words omitted]”

[0069] The first line is the table header, and the remaining 503 lines are specific content, including two columns of "Self-introduction" and "Past resume" with long content. For such large tables, the big model cannot view all the content at once, and it is necessary to extract sub-tables related to user questions so that the big model can complete accurate answers within the length limit while removing irrelevant information to prevent interference.

[0070] After receiving the user's question request regarding this table, "What are the height and weight of Xiao Zhang and Xiao Liu respectively?", the question request and the table header are first sent to the large language model. Based on the language understanding ability of the large language model and the semantic information of each column in the table header, the target columns "name", "height", and "weight" corresponding to the question request are determined.

[0071] Then calculate the vector similarity between the question request and the index columns of each row of the table, and determine the target rows corresponding to the question request based on the similarity calculation results, that is, the two rows where "Xiao Zhang" and "Xiao Liu" are located.

[0072] Using the above target rows and columns, the question subtable corresponding to the question request is constructed as follows: "[name, height, weight]

[0073] [Xiao Zhang, 165, 55]

[0074] [Xiao Liu, 185, 75]”.

[0075] The reason why the embodiment of the present invention adopts a large language model to select the target column instead of also adopting vector similarity to select the target column is that: in some scenarios, the questions expressed by users will not directly mention the column name or similar expressions of the column name. The large language model with strong natural language understanding ability is needed to analyze the target columns with potential semantic relevance. For example, the question "What projects has Xiao Zhao done?" does not repeat the language and text with the "Past Experience" in the table header, and the target column cannot be accurately determined by vector indexing, but the semantic understanding ability of the large language model can be used to determine the content of the column.

[0076] In step S500, based on the large language model, a reply result is output according to the question request and the question sub-table.

[0077] Specifically, the above steps include:

[0078] The question request and the corresponding question sub-table are sent to the large language model to determine whether the question request is a complex question that cannot be answered through a single round of interaction or a single sub-table. If the question request is not a complex question that cannot be answered through a single round of interaction or a single sub-table, then based on the large language model, the answer result is output according to the question request and the question sub-table. If the question request is a complex question that cannot be answered through a single round of interaction or a single sub-table, the large language model is used to decompose the question request, first taking the question request as a sub-question, and generating the sub-question answer and follow-up question mark corresponding to the sub-question according to the question sub-table. The follow-up question process is executed based on the follow-up question mark to generate the corresponding follow-up question sub-question for the question. Finally, the follow-up question sub-question is sent to the large language model as a new question request, and the steps of the above-mentioned table question-answering method based on the large language model are executed to obtain the sub-question answer corresponding to the follow-up question sub-question. The same process is repeated in sequence, and finally all the sub-question answers are used to splice the final answer result.

[0079] For example, in a software module purchase scenario, if you need to purchase module A as a dependent module, then when the user asks the question "How much is module A?", the price of module C in the table needs to be returned in the answer. Therefore, the large language model first answers the price of module A, and then generates a follow-up sub-question "

Follow-up question

[0080] The embodiment of the present invention uses vector index to select the row most relevant to the user's question, and uses a large language model to select the column most relevant to the user's question, and then extracts a question sub-table, which is sent to the large language model again for answering in combination with the question; for multiple candidate tables, when using the large language model to select columns according to user questions, the headers of multiple candidate tables are observed at the same time, and the selection of specific tables is completed while selecting relevant columns. For domain-specific knowledge outside the table, it can be stored separately in a table. Therefore, the present invention supports automatic switching between answering based on table content and answering based on domain knowledge (FAQs).

[0081] The embodiment of the present invention breaks down complex questions into several sub-questions and answers them in sequence. After completing one answer, a special mark is used to trigger a follow-up question process, and finally the answers to all sub-questions are concatenated and returned to the user, so that the large language model can complete the question decomposition and accurate answering when facing complex questions under length restrictions.

[0082] In addition, an embodiment of the present invention also provides a device, which includes: a processor and a memory; the memory is used to store one or more program instructions; the processor is used to run one or more program instructions to execute the steps of a table question answering method based on a large language model as described above.

[0083] In addition, an embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the table question answering method based on a large language model are implemented as described above.

[0084] In addition, an embodiment of the present invention further provides a computer program product, which includes computer program instructions, and when the computer program instructions are executed by a processor, the steps of the table question answering method based on a large language model as described above are implemented.

[0085] In order to overcome the problem that the entire table content cannot be sent to the large language model when the table size is too large, and the hallucination problem caused by too much interference information when the large language model observes the entire table, the embodiments of the present invention provide a table question and answer method, device, medium and program product based on a large language model. The language understanding ability of the large language model is first used to select the target column to be used according to the user's question request, and then the target row to be used is obtained according to the content similarity between the user's question request and the index column established for each row in the form. The target column and the target row are used to construct a question sub-table, and then the large language model is used to answer the user's question request according to the content of the question sub-table.

[0086] The embodiment of the present invention significantly reduces the scope of the table information used by the large language model when answering by extracting question sub-tables, effectively alleviating the computational pressure and hallucination answering problems caused by length restrictions, so that the large language model can accurately answer large-volume table contents under the condition of limited maximum processing length. In addition, by breaking down complex questions into several sub-questions and answering them in sequence, after completing one answer, a special mark is used to trigger the follow-up question process, and finally the answers to all sub-questions are spliced ​​and returned to the user, so that the large language model can complete the problem decomposition and accurately answer the complex questions under the length restriction. At the same time, by storing the domain-specific knowledge outside the table separately in a table, automatic switching between the answering method based on the table content and the answering method based on the domain knowledge (frequently asked questions) is achieved.

[0087] In the embodiment of the present invention, the processor may be an integrated circuit chip with signal processing capability. The processor may be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, discrete gates or transistor logic devices, or discrete hardware components. The methods, steps, and logic block diagrams disclosed in the embodiments of the present invention may be implemented or executed. The general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc. The steps of the method disclosed in the embodiment of the present invention may be directly embodied as being executed by a hardware decoding processor, or may be executed by a combination of hardware and software modules in a decoding processor. The software module may be located in a mature storage medium in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The processor reads the information in the storage medium and completes the steps of the above method in combination with its hardware. The storage medium may be a memory, for example, a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory can be a random access memory (RAM), which is used as an external cache. By way of example but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct RAMbus RAM (DRRAM).The storage medium described in the embodiment of the present invention is intended to include but is not limited to these and any other suitable types of memory. Those skilled in the art should be aware that in one or more of the above examples, the functions described in the present invention can be implemented by a combination of hardware and software. When the software is applied, the corresponding function can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein the communication medium includes any medium that is convenient for transmitting a computer program from one place to another. The storage medium can be any available medium that a general or special-purpose computer can access. Although the present invention has been described in detail above with general descriptions and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements can be made to it on the basis of the present invention. Therefore, these modifications or improvements made on the basis of not departing from the spirit of the present invention all belong to the scope of protection claimed in the present invention.

Claims

1. A table question answering method based on a large language model, characterized in that: The method comprises: Utilize the file upload interface to receive the table file uploaded by the user, vectorize the content of the table file and store it in the vector library; After receiving a question request sent by a user, determine a target column corresponding to the question request in a table file using a large language model; Determine a target row corresponding to the question request according to the vector similarity between the string obtained by concatenating the question request and the index columns of each row of the table; Using the target row and the target column, construct a question subtable corresponding to the question request; Based on the large language model, a reply result is output according to the question request and the question sub-table.

2. A table question answering method based on a large language model as claimed in claim 1, characterized in that: The table file is a file containing a table or an Excel file containing multiple form pages.

3. A table question answering method based on a large language model as claimed in claim 1, characterized in that: Use the file upload interface to receive table files uploaded by users, including: Use the file upload interface to receive the table file uploaded by the user; Frequently used questions and answers in the same field as the table file are listed separately as question-answer pairs in a question-answer pair table.

4. A table question answering method based on a large language model as claimed in claim 3, characterized in that: After receiving the question request sent by the user, the target column corresponding to the question request is determined in the table file using the large language model, including: After receiving the question request sent by the user, the target column corresponding to the question request in the table is determined according to the table header of the table file using the large language model; If the target column is located in the question-answer pair table, the question is answered using a common question-answer process based on domain knowledge.

5. The table question answering method based on a large language model as claimed in claim 1, characterized in that: Based on the large language model, according to the question request and the question sub-table, a reply result is output, including: Determine whether the question request is a complex question that cannot be answered through a single round of interaction or a single sub-table; If the question request is not a complex question that cannot be answered through a single round of interaction or a single sub-table, then based on the large language model, output a response result according to the question request and the question sub-table; If the question request is a complex question that cannot be answered through a single round of interaction or a single sub-table, the question request is decomposed using a large language model to obtain at least two sub-questions; For each of the sub-questions, generate a corresponding sub-question answer; All sub-question responses are concatenated to get the final response.

6. A table question answering method based on a large language model as claimed in claim 5, characterized in that: If the question request is a complex question that cannot be answered through a single round of interaction or a single sub-table, the question request is decomposed using a large language model to obtain at least two sub-questions, and a corresponding sub-question answer is generated for each of the sub-questions, including: If the question request is a complex question that cannot be answered through a single round of interaction or a single sub-table, the question request will be treated as a sub-question; Generate sub-question answers and follow-up question marks corresponding to the sub-questions according to the question sub-table; Execute a follow-up questioning process based on the follow-up questioning mark to generate a corresponding follow-up question for the sub-question; The follow-up sub-question is sent as a new question request to the large language model to obtain a sub-question answer corresponding to the follow-up sub-question.

7. A table question answering method based on a large language model as claimed in claim 1, characterized in that: The content of the table file is vectorized and stored in a vector library, including: Parse and extract the received table files, vectorize the contents of each table and store them in the vector library; For each table, determine the index column of the table based on the table header.

8. A table question-answering device based on a large language model, characterized in that: The device comprises: a processor and a memory; The memory is used to store one or more program instructions; The processor is used to run one or more program instructions to execute the steps of a table question answering method based on a large language model as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the steps of the table question answering method based on a large language model as described in any one of claims 1 to 7 are implemented.

10. A computer program product, characterized in that The computer program product includes computer program instructions, which, when executed by a processor, implement the steps of a table question answering method based on a large language model as described in any one of claims 1 to 7.

Images