A hybrid table text question answering method and system based on multi-task
Through the multi-task mixed table text question and answer method, combined with fact linking, retrieval and reasoning modules, the problem of the accuracy and efficiency of answer generation in the existing system in the processing of mixed table text data is solved, achieving more efficient and accurate answer generation .
Patent Information
- Application Number
- CN202311754726.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-19
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2043-12-19
AI Technical Summary
When processing mixed table text data, existing question and answer systems are susceptible to exposure deviations and error propagation, resulting in reduced accuracy in answer generation and low computational efficiency.
Using a multi-task-based mixed table text question-and-answer method, through fact linking, fact retrieval, span extraction and numerical inference modules, splicing questions, tables and text data as input sequences, using pre-trained models for semantic extraction and sorting, combined with classifier judgment Whether to perform numerical inference or span extraction, generate the answer.
It effectively avoids error transmission, improves the accuracy and efficiency of answer generation, enhances the robustness and generalization capabilities of the model, and is suitable for a variety of text and table question-and-answer scenarios.
Smart Images

Figure CN117910573B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of intelligent question answering technology, and in particular to a multi-task based mixed table text question answering method and system. Background Art
[0002] The advancement of information technology has brought about a huge change in the acquisition of knowledge, realizing the transition from traditional document retrieval to intelligent question-answering systems. Intelligent question-answering systems enable computers to better understand natural language questions asked by users and provide precise answers to their questions. Therefore, question-answering systems are widely regarded as the main means of obtaining information in the future. Industry and academia have conducted extensive research in this field, and question-answering systems are currently used in various fields. According to the structure of the data source, question-answering systems can be divided into two categories: single-data source question-answering systems and mixed-data source question-answering systems. In previous research on question-answering systems, people mainly focused on homogeneous data, such as unstructured text or structured data. In contrast, the field of mixed-table text question-answering faces greater challenges because it involves reasoning from heterogeneous information and usually requires numerical reasoning to provide answers to questions in addition to span extraction answers. Currently, numerical reasoning techniques often rely on autoregressive models to decode program sequences. However, these methods are susceptible to exposure bias and error propagation, which significantly reduce the accuracy of program generation as the decoding process unfolds. Summary of the invention
[0003] In order to solve the above problems existing in the prior art, the present invention provides a multi-task based hybrid table text question answering method and system.
[0004] To achieve the above object, the present invention provides the following solutions:
[0005] A multi-task based hybrid table text question answering method, comprising:
[0006] Obtain questions, tables, and text data;
[0007] Converting the content in the table into text information according to a predetermined template;
[0008] Concatenating the question, the text information and the text data to obtain an input sequence;
[0009] Input the input sequence into the retriever to retrieve supporting facts, obtain a fact sequence, and select the first n supporting facts in the fact sequence;
[0010] Re-arrange the selected supporting facts according to their corresponding positions in the input sequence to obtain search results;
[0011] Inputting the search result into a classifier to obtain a classification result, and judging whether to perform numerical reasoning on the search result based on the classification result to obtain a judgment result;
[0012] When the judgment result is yes, reasoning is performed on the search result to obtain an answer;
[0013] When the judgment result is no, span extraction is performed on the search result to obtain an answer.
[0014] Optionally, the input sequence is input into a retriever to retrieve supporting facts, obtain a fact sequence, and select the first n supporting facts in the fact sequence, specifically including:
[0015] Using the IO tagging method to tag the text data and text information in the input sequence to obtain a tag sequence; wherein the word vectors in the text data and the text information that are positively correlated with the answer are marked as I, and the word vectors in the text data and the text information that are negatively correlated with the answer are marked as O;
[0016] Input the tag sequence into the retriever to obtain the fact sequence and the probability of each supporting fact being marked as 1;
[0017] Select n supporting facts based on the probability of each supporting fact being marked as 1.
[0018] Optionally, inputting the search results into a classifier to obtain classification results specifically includes:
[0019] Inputting the search result into the classifier to obtain the probability of containing a special vector [CLS];
[0020] Based on the probability, it is determined whether to perform numerical reasoning on the search result to obtain a determination result.
[0021] Optionally, the process of inferring the search results to obtain an answer includes:
[0022] The special vector [CLS] and the soft mask representation are input into the classifier to obtain the operator prediction result;
[0023] Common numerical values are added as special tags, and the special tags are connected with questions and sentences containing supporting facts as input to the RoBERTa encoder to obtain operand prediction results;
[0024] predicting the order and the numerical units of the operands using a dual classifier;
[0025] The answer is derived based on the operator predictions, operand predictions, the order of the operands, and the numerical units.
[0026] Optionally, the probability of including a special vector [CLS] is:
[0027] P next =softmax(FFN[CLS]);
[0028] Where P next is the probability of containing the special vector [CLS], FFN is a 2-layer feedforward network operation with GELU as the activation function, and [CLS] is a special vector.
[0029] Optionally, the answer probability obtained by span extraction of the retrieval results is the product of the probabilities of the starting position and the ending position in the process of supporting fact retrieval.
[0030] Furthermore, the present invention also provides a multi-task based hybrid table text question answering system, the system is used to implement the multi-task based hybrid table text question answering method provided above; the system comprises:
[0031] A fact linking module, used to obtain questions, tables and text data, to convert the content in the table into text information according to a predetermined template, and to concatenate the questions, the text information and the text data to obtain an input sequence;
[0032] A fact retrieval module, used for inputting the input sequence into a retriever to retrieve supporting facts, obtaining a fact sequence, and selecting the first n supporting facts in the fact sequence, and re-sorting the selected supporting facts according to the corresponding position in the input sequence to obtain a retrieval result, and used for inputting the retrieval result into a classifier to obtain a classification result, and judging whether to perform numerical reasoning on the retrieval result based on the classification result to obtain a judgment result;
[0033] A span extraction module, used for performing span extraction on the search result to obtain an answer when the judgment result is no;
[0034] The digital reasoning module is used to reason the search results to obtain an answer when the judgment result is yes.
[0035] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:
[0036] The present invention splices questions, text information and text data as input sequences, and adopts a parallel generation model, which can improve the efficiency of generating answers to mixed reasoning questions with heterogeneous data; the present invention inputs the input sequence into a retriever to retrieve all supporting facts, so as to select the first n supporting facts, and re-sorts them according to the corresponding position of each selected supporting fact in the input sequence to obtain retrieval results, inputs the retrieval results into a classifier to obtain classification results, and judges whether to perform numerical reasoning on the retrieval results based on the classification results to obtain judgment results, which can effectively avoid error propagation and further improve the accuracy of generating answers; finally, based on the judgment results, it is selected to perform reasoning or span extraction on the retrieval results to obtain the final answer, so that the reasoning effect is better and the model has stronger robustness, thereby effectively solving the problems of exposure bias, error propagation and low computational efficiency brought by traditional decoding programs. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0038] Figure 1 A flowchart of the multi-task based hybrid table text question answering method provided by the present invention;
[0039] Figure 2 This is an implementation block diagram of the multi-task based hybrid table text question answering method provided by the present invention. DETAILED DESCRIPTION
[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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 are within the scope of protection of the present invention.
[0041] The purpose of the present invention is to provide a multi-task based hybrid table text question answering method and system, which can effectively solve the problems of exposure bias, error propagation and low computational efficiency brought by traditional decoding procedures.
[0042] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0043] The multi-task based hybrid table text question answering method provided by the present invention mainly includes the steps of fact linking, fact retrieval, span extraction, and digital reasoning. The specific implementation process is as follows:
[0044] (1) Fact linking: For a given question Q, table T, and text P, first describe the table as text information according to a predetermined template. Then input the described text information, the given question, and the relevant paragraphs into the encoder to obtain the corresponding representation H (i.e., the concatenated input sequence) to provide input for subsequent retrieval. In the process of describing the table, the cell row headers and column headers as well as the content of each cell are converted into text information.
[0045] (2) Fact retrieval: Use the pre-trained model to perform semantic extraction and obtain supporting facts. Select the first n retrieved supporting facts and reorder them in the order they appear in the input sequence to obtain the retrieval results. Input the retrieval results into the numerical reasoning module. At the same time, use the special vector [CLS] to divide the problem into two types: numerical reasoning and span extraction.
[0046] (3) Span extraction: The purpose of this operation is to select the predicted span candidate, that is, a span of the retrieved sentence. The answer probability is defined as the product of the probabilities of the start position and the end position in the fact retrieval.
[0047] (4) Numerical reasoning: Based on the search results retrieved by the retriever, the final answer needs to be determined through numerical reasoning. In order to improve the generalization ability of the model, a multi-task model with a residual learning mechanism can be used to solve and answer questions. First, the operands of numerical reasoning are determined based on the residual coding output label. Secondly, the probability of the operator type is calculated by averaging the residual coding output and the special vector [CLS]. Next, based on the numerical label of the residual coding output, a dual classifier is added to determine whether the order of the operands is consistent with the original order. Finally, a multi-classifier is designed to predict the numerical unit.
[0048] Using weighted negative log-likelihood loss and training question-answering models, the predicted values of each task are input, and the optimal parameters are determined based on the error between the predicted value and the true answer, ultimately guiding the model to generate the optimal answer.
[0049] For example, given a question Q, the goal of the hybrid table-text question answering system is to find the correct answer from the known table T and text P heterogeneous data. In most cases, the model only needs to extract candidate answers from the text data. However, for mathematical questions, mathematical reasoning is required through the table data to obtain the correct answer.
[0050] Define the hybrid table question answering method task as G = {G0, G1, ..., G n}, where G irepresents a subtask, i = 1, 2, ..., n. The task consists of four parts: fact linking, fact retrieval, span extraction, and numerical reasoning. The probability of obtaining the correct answer A is equal to the probability of all possible task sequences G that may produce the answer A. i The sum of the probabilities is:
[0051] P(A|T,P,Q)=∑ i P(G i |T,P,Q).
[0052] Where P(A|T,P,Q) is the probability of the correct answer A, P(G i |T,P,Q) is a task sequence G that may produce answer A i probability.
[0053] like Figure 1 and Figure 2 As shown, in this embodiment, the multi-task based hybrid table text question answering method includes the following steps:
[0054] Step 1: Fact link.
[0055] Question Q, text P and table T, where Q = {q1, q2, ..., q i}, P = {p1, p2, ..., p m}, T={t1,t2,…,t m}, where q i is the problem weight, p m is the text component, t m is the table component. To improve the accuracy and efficiency of table retrieval, the table description, cell row headers, column headers and each cell content are converted into text information using a predetermined template. Next, the question, the flattened table and the text are linked together as the input X of the encoder, expressed as:
[0056] X=(x q ,x t ,x p ).
[0057] Among them, x q For the question after the link, x t For the linked table, x p The text after the link.
[0058] Step 2: Fact retrieval.
[0059] For a given question Q, the input text P and table T sequence are labeled using the IO labeling method. Each sub-token is labeled independently, and the word vectors that are positively related to the answer are labeled as I, and the word vectors that are negatively related to the answer are labeled as O. Those word vectors labeled as I are selected as supporting facts for the answer. Formally, for each word vector input, its label is represented as t, and the probability of being labeled as I is:
[0060]
[0061] in, is the probability marked as I. FFN is a 2-layer feed-forward network with GELU as the activation function. t is the representation of sub-vector t. Subsequently, the top n sentences are selected as input to the next stage based on the prediction of the supporting fact classification model. In order to determine whether the next stage involves span extraction or numerical reasoning, a classifier is used for prediction.
[0062] P next =softmax(FFN[CLS]).
[0063] Where P next is the probability of containing the special vector [CLS].
[0064] Step 3: Span extraction.
[0065] The purpose of span extraction is to predict span candidates based on fact retrieval results, that is, to extract a piece of text as the answer from a given text. The answer probability is defined as the product of the probabilities of the starting and ending positions of the retrieved supporting facts, which is:
[0066]
[0067] Where P span is the probability of the answer, is the probability of the starting position of the retrieved supporting fact, is the probability of the end position of the retrieved supporting fact.
[0068] Step 4: Numerical reasoning.
[0069] Numerical reasoning obtains the final answer by reasoning on the retrieved supporting facts, including operand prediction, operator prediction, order prediction, and numerical scale prediction. Among them, operand prediction is mainly used to predict the accurate probability of operands in numerical reasoning. Operator prediction is used to predict the type of operands in numerical reasoning, such as addition, subtraction, multiplication, and division. Sequence prediction is used to predict the order of operands. Since some numerical operations have strict order requirements for operands, this operation is mainly used to determine whether their order is the same as or opposite to that of the input. For numerical reasoning, the unit of the operand often becomes the key to obtaining the correct answer. For this, numerical scale prediction is used to determine the scale of the value.
[0070] (1) Operator prediction.
[0071] In order to overcome the error propagation caused by the Roberta model prediction error and the overall performance degradation, the present invention adopts a method similar to residual learning to perform element-wise addition of the feature vector obtained by dynamic masking and the original vector probability to obtain the output, and the output is expressed as:
[0072]
[0073] In the formula, is the output, h s is the original vector, h t is the representation of subvector t.
[0074] The operator generator defines several common operators: addition, subtraction, multiplication, division, exponential operation, comparison operation, etc. A multi-class classifier is used as an operator predictor to predict the operation type of numerical reasoning. The classifier adopts a single-layer FFN structure, and the input is a special vector [CLS] containing classification information (such as text span, table cell, addition, subtraction, multiplication, division, count, average, and change rate) and a soft mask representation. The operator type probability is calculated by averaging the two. Among them, the operator type probability is:
[0075]
[0076] In the formula, P operator is the operator type probability.
[0077] (2) Operand selection.
[0078] Correctly selecting operands is crucial for inferring the correct answer and is beneficial to improving the overall performance of answer generation. First, for operations such as averaging and digital order conversion that may occur during numerical reasoning, common values (such as 0 to 9, etc.) are added to special tags, and the special tags are connected to questions and sentences containing supporting facts as the input of the RoBERTa encoder. Next, the residual learning mechanism is used to extract the operands required for numerical reasoning from the input. As with operator selection, a two-layer feedforward network (FFN) is used for operand selection to identify all operands of the reasoning program on the output of the RoBERTa encoder. Among them, the probability of the operand is P operand , P operand The larger the value, the closer the soft mask result is to the original embedding. On the contrary, operand The smaller the value, the closer the result is to the mask embedding. Therefore, the soft mask mechanism can effectively prove the priority of candidate answers. Compared with hard masking using classification results, soft masking is differentiable and can be trained in an end-to-end manner while alleviating the error propagation problem.
[0079]
[0080] (3) Sequential prediction.
[0081] When the operator is subtraction, division, exponentiation or comparison, the order of the two operands is important. In this method, a dual classifier is used to predict the order of the operands, whether they are in the same order as the input or in the opposite order. The probability is 0 if the order of the operands is the same as in the input sequence, otherwise it is 1. The order prediction probability is denoted as P order :
[0082] P order =softmax(FFN(avg(h t1 ,h t2 ))).
[0083] Among them, avg() means taking the average value. t1 ,h t2 Respectively represent the two operands required for the operation.
[0084] (4) Numerical unit prediction.
[0085] The final correct answer is related to the predicted operator, order, and operands, and should also include the correct numerical unit. Here, a multi-class classifier is used to predict the numerical unit (such as unitless, thousands, millions, billions, and percentages). The special vector [CLS], table vector, and paragraph vector are input in sequence, and the probability of predicting the numerical unit is P. scale :
[0086]
[0087] In the formula, is a table vector, is the paragraph vector.
[0088] Step 4: Training.
[0089] In order to optimize the multi-task question-answering model for mixed table-text data, a multi-task loss function is designed, including five sub-tasks: label prediction, operand prediction, operator prediction, order prediction, and numerical unit prediction. It is used to calculate the error between the predicted value and the true answer, and finally output the parameter series of the best answer, which is:
[0090]
[0091] In the formula, is the multi-task loss function value, NLL is the negative log-likelihood loss function, P t The probability of supporting facts extracted for each task, G t Support facts extracted for each task, G operator for, is the probability of the operand extracted in the lth reasoning, The supporting facts of the operands extracted for the lth reasoning, is the sequential prediction probability of the lth inference, is the supporting fact of the sequential prediction extracted for the lth reasoning, is the predicted numerical unit probability of the lth inference, is the supporting fact of the predicted numerical unit extracted by the l-th reasoning, α, β, γ, δ, ε all represent weights, N is the maximum number of reasoning times, l = 1, 2, ..., N.
[0092] Based on the above description, compared with the related art, the present invention has the following beneficial effects:
[0093] 1) The present invention adopts a parallel generation model to realize the answer generation of heterogeneous data mixed reasoning problems, further improving the efficiency of answer generation.
[0094] 2) The present invention adopts a multi-task mode for numerical reasoning, which effectively avoids error transmission and further improves the accuracy of reasoning. At the same time, the noise of different tasks will offset some of the noise to a certain extent, making the reasoning effect better and the model more robust.
[0095] 3) The present invention has strong generalization ability and is suitable for a variety of text, table or data question and answer scenarios, and only needs to add or delete tasks.
[0096] An embodiment is provided below to illustrate the implementation process of the multi-task based hybrid table text question answering method provided above in the present invention.
[0097] Question: How much does the commercial cloud revenue account for the total revenue in 2019?
[0098] Corresponding text and table (see Table 1):
[0099] Revenue from external customers, classified by significant product and service offerings, was as follows:
[0100] Table 1
[0101]
[0102]
[0103] Our commercial cloud revenue, which includes Office 365Commercial, Azure, the commercial portion of LinkedIn, Dynamics 365, and other commercial cloud properties, was $38.1 billion, $26.6 billion and $16.2 billion in fiscal years 2019, 2018, and 2017, respectively. These amounts are primarily included in Office products and cloud services, Server products and cloud services, and LinkedIn in the table above.
[0104] First, convert the table rows and columns into sentences according to the template:
[0105] (...abbreviate...)Table 1show Server products and cloud services of revenue from external customers is 32622(in millions)in 2019; Table 1show Server products and cloud services of revenue from external customers is 26129(in millions)in 2018; 2017;…(…abbreviate…).
[0106] Second, concatenate the questions, text, and table sequences into:
[0107] [CLS]How much…revenue account…in 2019? [SEP]…in millions…5134…[SEP]…$38.1billion…[SEP].
[0108] Again, supporting facts are extracted from the table and related paragraphs by using IO tagging for sequence tagging. First find the fact in the table, otherwise find the fact in its related paragraph.
[0109] If a fact appears multiple times, only the first one found is kept.
[0110]
[0111] Next, perform numerical reasoning: the predicted length is 1. The operands are selected as 38.1 billion and 125843 million respectively. The identification operator is division. If the order of the operands is the same as the order in the input sequence, it is 0, otherwise it is 1, and the predicted order is 0.
[0112] Next, adjust the hyperparameter training model and use the multi-task loss function to train the model until the loss function value on the training set no longer decreases.
[0113] Finally, based on text span extraction and multi-task numerical reasoning, the answer to the question is obtained. For the given question, the model predicts that the answer is "The commercial cloud revenue account for the total revenue in 2019 is 30.28%".
[0114] Furthermore, the present invention also provides a multi-task based hybrid table text question answering system, which is used to implement the multi-task based hybrid table text question answering method provided above. The system includes:
[0115] The fact linking module is used to obtain questions, tables and text data, to convert the contents in the tables into text information according to a predetermined template, and to concatenate questions, text information and text data to obtain an input sequence.
[0116] The fact retrieval module is used to input the input sequence into the retriever to retrieve supporting facts, obtain a fact sequence, and select the first n supporting facts in the fact sequence, and re-sort them according to the corresponding position of each selected supporting fact in the input sequence to obtain retrieval results, and input the retrieval results into the classifier to obtain classification results, and judge whether to perform numerical reasoning on the retrieval results based on the classification results to obtain judgment results.
[0117] The span extraction module is used to extract the span of the search results to obtain the answer when the judgment result is no.
[0118] The digital reasoning module is used to reason about the search results and obtain the answer when the judgment result is yes.
[0119] In addition, when the method provided above is implemented in the form of a computer program through a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory, a random access memory, a magnetic disk or an optical disk.
[0120] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0121] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of the present invention. At the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.
Claims
1. A multi-task hybrid table text question answering method, characterized in that: include: Obtain questions, tables, and text data; Converting the content in the table into text information according to a predetermined template; Concatenating the question, the text information and the text data to obtain an input sequence; Input the input sequence into the retriever to retrieve supporting facts, obtain a fact sequence, and select the first n supporting facts in the fact sequence; Re-arrange the selected supporting facts according to their corresponding positions in the input sequence to obtain search results; Inputting the search result into a classifier to obtain a classification result, and judging whether to perform numerical reasoning on the search result based on the classification result to obtain a judgment result; When the judgment result is yes, reasoning is performed on the search result to obtain an answer; When the judgment result is no, performing span extraction on the search result to obtain an answer; The input sequence is input into the retriever to retrieve supporting facts, obtain a fact sequence, and select the first n supporting facts in the fact sequence, specifically including: Using the IO tagging method to tag the text data and text information in the input sequence to obtain a tag sequence; wherein the word vectors in the text data and the text information that are positively correlated with the answer are marked as I, and the word vectors in the text data and the text information that are negatively correlated with the answer are marked as O; The tag sequence is input into the retriever to obtain the fact sequence and the probability of each supporting fact being marked as I; for each word vector input, its tag is represented as t, and the probability of being marked as I is: In the formula, is the probability marked as I, FFN is a 2-layer feedforward network with GELU as the activation function, h t is the representation of the subvector t; Select n supporting facts based on the probability of each supporting fact being marked as 1.
2. The multi-task hybrid table text question answering method according to claim 1 is characterized in that: The search results are input into a classifier to obtain classification results, specifically including: Inputting the search result into the classifier to obtain the probability of containing a special vector [CLS]; Based on the probability, it is determined whether to perform numerical reasoning on the search result to obtain a determination result.
3. The multi-task hybrid table text question answering method according to claim 2 is characterized in that: The process of inferring the search results to obtain answers includes: The special vector [CLS] and the soft mask representation are input into the classifier to obtain the operator prediction result; Common numerical values are added as special tags, and the special tags are connected with questions and sentences containing supporting facts as input to the RoBERTa encoder to obtain operand prediction results; predicting the order and the numerical units of the operands using a dual classifier; The answer is derived based on the operator predictions, operand predictions, the order of the operands, and the numerical units.
4. The multi-task hybrid table text question answering method according to claim 2, characterized in that: The probability of including a special vector [CLS] is: P next =softmax(FFN[CLS]); Where P next is the probability of containing the special vector [CLS], where [CLS] is a special vector.
5. The multi-task hybrid table text question answering method according to claim 1, characterized in that: The answer probability obtained by span extraction of the retrieval results is the product of the probabilities of the starting position and the ending position in the process of supporting fact retrieval.
6. A multi-task hybrid table text question answering system, characterized in that: The system is used to implement the multi-task based hybrid table text question answering method according to any one of claims 1 to 5; the system comprises: A fact linking module, used to obtain questions, tables and text data, to convert the content in the table into text information according to a predetermined template, and to concatenate the questions, the text information and the text data to obtain an input sequence; A fact retrieval module, used for inputting the input sequence into a retriever to retrieve supporting facts, obtaining a fact sequence, and selecting the first n supporting facts in the fact sequence, and re-sorting the selected supporting facts according to the corresponding position in the input sequence to obtain a retrieval result, and used for inputting the retrieval result into a classifier to obtain a classification result, and judging whether to perform numerical reasoning on the retrieval result based on the classification result to obtain a judgment result; A span extraction module, used for performing span extraction on the search result to obtain an answer when the judgment result is no; The digital reasoning module is used to reason the search results to obtain an answer when the judgment result is yes.
Citation Information
Patent Citations
Multi-hop question answering method based on multi-hop reasoning joint optimization
CN114780707A
Multi-hop reading understanding method and system based on double-view comparative learning and graph pruning
CN116403231A
Adaptive evaluation of text search queries with blackbox scoring functions
US20040243632A1
