Data processing method, device and equipment

Abstract

The embodiment of the invention discloses a data processing method, device and equipment, the method is applied to an open domain question answering system, the open domain question answering system comprises a retriever and a reader, and the method comprises the steps that target question data input by a user is received; based on the content of the target problem data, performing extension processing on a user intention corresponding to the target problem data to obtain one or more different clarified query data; inputting each piece of clear query data into a retriever to obtain retrieval data corresponding to each piece of clear query data; determining correlation information between the retrieval data corresponding to each clear query data and the target problem data, and rewriting the retrieval data corresponding to each clear query data based on the determined correlation information to obtain rewritten retrieval data; and based on the rewritten retrieval data, determining answer data corresponding to the target question data through the reader, and outputting the answer data.

Description

Technical Field

[0001] This document relates to the field of computer technology, and in particular to a data processing method, device and equipment. Background Art

[0002] Open-domain question answering aims to design question answering models to answer open questions without limiting the scope of the topic. In recent years, with the rapid development of large-scale language models and people paying more and more attention to their privacy data, the Retrieve-then-Read system using large-scale language models as readers has achieved good results in open-domain question answering. Such systems use a retriever to retrieve relevant context from the knowledge base, and then use a reader to read these contexts to generate answer data for the question data.

[0003] With the development of large-scale language models, many large-scale language models have achieved good results in open domain question answering. A further approach is to use a specified language model to build a retrieval-reading system to further improve performance. Specifically, the retrieval-reading system first uses a retriever to retrieve relevant documents from an external knowledge source, and then uses a reader to predict answers based on these documents. However, the current retrieval-reading system has a task mismatch between the retriever and the reader. To this end, it is necessary to provide a technical solution that can reduce or eliminate the task mismatch between the retriever and the reader. Summary of the invention

[0004] The purpose of the embodiments of this specification is to provide a technical solution that can reduce or eliminate the situation where the tasks between the retriever and the reader do not match.

[0005] In order to implement the above technical solution, the embodiments of this specification are implemented as follows: A data processing method provided in an embodiment of the present specification is applied to an open domain question-answering system, wherein the open domain question-answering system includes a retriever and a reader, and the method includes: receiving target question data input by a user. Based on the content of the target question data, performing expansion processing on the user intention corresponding to the target question data to obtain one or more different clarification query data. Inputting each of the clarification query data into the retriever respectively to obtain retrieval data corresponding to each of the clarification query data, wherein the retriever is used to retrieve retrieval data matching the input query data from a preset database. Determining the correlation information between the retrieval data corresponding to each of the clarification query data and the target question data, and rewriting the retrieval data corresponding to each of the clarification query data based on the determined correlation information to obtain the rewritten retrieval data. Based on the rewritten retrieval data, determining the answer data corresponding to the target question data through the reader, and outputting the answer data, wherein the reader is used to predict the answer data of the question data input by the user based on the input retrieval data.

[0006] A data processing device provided in an embodiment of the present specification is provided with an open domain question-answering system, wherein the open domain question-answering system includes a retriever and a reader, and the device includes: a question receiving module, receiving target question data input by a user. An expansion module, based on the content of the target question data, performs expansion processing on the user intention corresponding to the target question data to obtain one or more different clarification query data. A retrieval module, inputting each of the clarification query data into the retriever respectively, obtains the retrieval data corresponding to each of the clarification query data, and the retriever is used to retrieve the retrieval data matching the input query data from a preset database. A rewriting module, determining the correlation information between the retrieval data corresponding to each of the clarification query data and the target question data, and rewriting the retrieval data corresponding to each of the clarification query data based on the determined correlation information to obtain the rewritten retrieval data. An answer output module, based on the rewritten retrieval data, determines the answer data corresponding to the target question data through the reader, and outputs the answer data, and the reader is used to predict the answer data of the question data input by the user according to the input retrieval data.

[0007] The embodiment of the present specification provides a data processing device, the data processing device is provided with an open domain question answering system, the open domain question answering system includes a retriever and a reader, the data processing device includes: a processor; and a memory arranged to store computer executable instructions, the executable instructions when being executed, enable the processor to: receive target question data input by a user. Based on the content of the target question data, perform expansion processing on the user intention corresponding to the target question data to obtain one or more different clarification query data. Input each of the clarification query data into the retriever respectively to obtain the retrieval data corresponding to each of the clarification query data, the retriever is used to retrieve the retrieval data matching the input query data from a preset database. Determine the correlation information between the retrieval data corresponding to each of the clarification query data and the target question data, and rewrite the retrieval data corresponding to each of the clarification query data based on the determined correlation information to obtain the rewritten retrieval data. Based on the rewritten retrieval data, determine the answer data corresponding to the target question data through the reader, and output the answer data, the reader is used to predict the answer data of the question data input by the user according to the input retrieval data.

[0008] The embodiment of the present specification also provides a storage medium, which is applied to an open domain question answering system, wherein the open domain question answering system includes a retriever and a reader, wherein the storage medium is used to store computer executable instructions, and the executable instructions implement the following process when executed by the processor: receiving target question data input by a user. Based on the content of the target question data, the user intention corresponding to the target question data is extended to obtain one or more different clarification query data. Each of the clarification query data is input into the retriever respectively to obtain the retrieval data corresponding to each of the clarification query data, and the retriever is used to retrieve the retrieval data matching the input query data from a preset database. Determine the correlation information between the retrieval data corresponding to each of the clarification query data and the target question data, and rewrite the retrieval data corresponding to each of the clarification query data based on the determined correlation information to obtain the rewritten retrieval data. Based on the rewritten retrieval data, the answer data corresponding to the target question data is determined by the reader, and the answer data is output, and the reader is used to predict the answer data of the question data input by the user according to the input retrieval data.

[0009] The embodiment of the present specification also provides a computer program product, which is applied to an open domain question answering system. The open domain question answering system includes a retriever and a reader. The computer program product includes a computer program. When the computer program is executed by a processor, the following process is implemented: receiving target question data input by a user. Based on the content of the target question data, the user intention corresponding to the target question data is extended to obtain one or more different clarification query data. Each of the clarification query data is input into the retriever respectively to obtain the retrieval data corresponding to each of the clarification query data. The retriever is used to retrieve the retrieval data matching the input query data from a preset database. Determine the correlation information between the retrieval data corresponding to each of the clarification query data and the target question data, and rewrite the retrieval data corresponding to each of the clarification query data based on the determined correlation information to obtain the rewritten retrieval data. Based on the rewritten retrieval data, the answer data corresponding to the target question data is determined by the reader, and the answer data is output. The reader is used to predict the answer data of the question data input by the user according to the input retrieval data. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the drawings required for use in the embodiments or the prior art description will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative labor. Figure 1 This is a schematic diagram of a question-answering process of an open domain question-answering system in this specification; Figure 2 This is an embodiment of a data processing method of this specification; Figure 3 This is a schematic diagram of an interface of an open domain question answering system in this specification; Figure 4 A schematic diagram of a question-answering process of another open domain question-answering system in this specification; Figure 5 This is a schematic diagram of a question-answering process of an open domain question-answering system in this specification; Figure 6 Another data processing method embodiment of this specification; Figure 7 This is another data processing method embodiment of the present specification; Figure 8 A schematic diagram of a question-answering process of another open domain question-answering system in this specification; Fig. 9 This is another data processing method embodiment of the present specification; Fig.10 This is another data processing method embodiment of the present specification; Fig.11 This is another data processing method embodiment of the present specification; Fig.12 A schematic diagram of a question-answering process of another open domain question-answering system in this specification; Fig.13 This is another data processing method embodiment of the present specification; Fig.14 This is a schematic diagram of a question-answering processing process of another open domain question-answering system in this specification; Fig.15 This is an embodiment of a data processing device of the present specification; Fig.16 This is an embodiment of a data processing device in this specification. DETAILED DESCRIPTION

[0011] The embodiments of this specification provide a data processing method, device and equipment.

[0012] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.

[0013] The embodiments of this specification provide an open domain question answering mechanism, which aims to design a question answering model to answer open questions that do not limit the scope of the subject. In recent years, with the rapid development of large-scale language models, retrieval-reading systems that use large-scale language models as readers have achieved good results in open domain question answering. Such systems use a retriever to retrieve relevant contexts from a knowledge base, and then use a reader to read these contexts to generate answer data for the question data.

[0014] With the development of large-scale language models, many large-scale language models have achieved good results in open-domain question answering. A further approach is to use a specified language model to build a retrieval-reading system to further improve performance. Specifically, the retrieval-reading system first uses a retriever to retrieve relevant documents from an external knowledge source, and then uses a reader to predict answers based on these documents. However, existing retrieval-reading systems have problems such as task mismatch and limited interaction between the retrieval system and the reading system.

[0015] The reading system takes natural language questions with context and intent as input data, while the retrieval system focuses on keyword retrieval. Figure 1 As shown in the figure, the question data raised by the user is "What genre is The Trip?". It has few common keywords with the relevant document "The Trip (2010 TV series)" in the retrieval database, and the important keyword "TV series" is omitted. Therefore, it is difficult to match the question data with the relevant documents. This mismatch between question data and query results often makes it difficult to retrieve documents related to the question data, and the documents actually retrieved are irrelevant to the question data, and may even mislead the reading system to answer the question correctly. In addition, the intention ambiguity of open domain questions and the mismatch between questions and terms in the knowledge base further aggravate the occurrence of task mismatch. For example, Figure 1 Although the relevant document "The Trip (2010 TV series)" in the search library contains relevant information about the question data, it is difficult to be used directly to answer the question. To this end, this specification embodiment proposes a chain rewriting method (Chain-of-Rewrite), which ensures that the rewritten content meets the user's intention through semantic analysis, while keeping the semantics before and after the rewriting consistent, effectively eliminating the task mismatch between the open domain question and the relevant documents. For specific processing, please refer to the specific content in the following embodiment.

[0016] like Figure 2 As shown, an embodiment of this specification provides a data processing method, and the execution subject of the method can be a terminal device or a server, etc., wherein the terminal device can be a mobile terminal device such as a mobile phone, a tablet computer, or a computer device such as a laptop or a desktop computer, or it can also be an IoT device (specifically such as a smart watch, a car-mounted device, etc.), etc., wherein the server can be an independent server, or it can be a server cluster composed of multiple servers, etc., and the server can be a background server in the financial field or the online shopping field, etc., or it can be a background server of a certain application, etc. In this embodiment, the execution subject is taken as an example for detailed description. For the case where the execution subject is a terminal device, please refer to the following server case processing, which will not be repeated here. In addition, an open domain question-and-answer system can be set in the execution subject, and the open domain question-and-answer system can include a retriever and a reader. The retriever can be used to retrieve retrieval data that matches the input query data from a preset database, and the reader can be used to predict the answer data of the question data input by the user based on the input retrieval data. The method can specifically include the following steps: In step S202, target question data input by a user is received.

[0017] Among them, the target question data can be data of any question, for example, the target question data can be "What genre is The Trip?", or the target question data can be "What type of mobile phone is it?", etc., which can be set according to actual conditions, and the embodiments of this specification do not limit this.

[0018] In implementation, when a user needs to know the answer data corresponding to a question data, the user can obtain the question and answer page through the terminal device, such as Figure 3 As shown, the question-and-answer page may include an input box for question data, an output box for answer data, a confirmation button, a cancel button, etc. The user may enter the question data (i.e., target question data) to be queried in the input box for question data in the question-and-answer page. After the input is completed, the user may click the confirmation button. At this time, the user's terminal device may obtain the target question data in the input box for question data, and may send the target question data to the server. The server may receive the target question data entered by the user.

[0019] In step S204, based on the content of the target question data, the user intention corresponding to the target question data is expanded to obtain one or more different clarification query data.

[0020] In practice, generally, for an open domain question answering system consisting of a retriever and a reader, first, the retriever can be used to retrieve relevant documents from a preset database (such as a database constructed by an external knowledge source, etc.), and then the reader can be used to predict the corresponding answer based on the retrieved documents. In order to reduce the occurrence of the above mismatch, this embodiment explores a chain-of-rewrite mechanism to align the question data raised by the user and the data of the retrieved relevant documents. The chain-of-rewrite mechanism can use a large-scale language model as the basic model, which rewrites the input data according to the given instructions. Figure 4 As shown, given an open domain question, that is, the target question data, the target question data can be converted into search queries using chain rewriting. These queries can further express user intentions and cover more relevant terms, which can effectively improve the recall rate of relevant documents. After that, the search query data is input into the retriever to retrieve a batch of relevant documents. In order to delete irrelevant content in the document and highlight relevant information, the chain rewriting mechanism can be used to convert the retrieved documents into relevant context information that can directly answer the target question data according to the target question data, so as to facilitate the question-answering open domain question-answering system to use relevant documents to correctly answer the questions corresponding to the target question data.

[0021] Specifically, Figure 5As shown in the upper part, the chain rewriting mechanism can be provided with semantic analysis rules. The semantic analysis in the chain rewriting mechanism can be used to identify and focus on the subject and key information contained in the target question data, so as to better align the target question data with the relevant documents. To this end, query expansion can be performed, that is, the user intent corresponding to the target question data can be expanded based on the content of the target question data to obtain one or more different clarification query data. For example, the given target question data "What genre is The Trip?" can be expanded into a series of clarification query data that can further express the user intent, such as "What genre is the TV series The Trip?", "What genre is the film The Trip?", "What genre is the book The Trip?", etc. For another example, the given target question data "What type of mobile phone is?" can be expanded into a series of clarification query data that can further express the user intent, such as "What type of movie is the movie "Mobile Phone"?", "What type of operating system is this mobile phone?", etc. By further expanding the user intent of the target question data, multiple different clarification query data can be obtained, so that more relevant terms (or keywords) can be covered, thereby effectively improving the recall rate of the retrieval.

[0022] In step S206, each clarification query data is input into the retriever to obtain the retrieval data corresponding to each clarification query data. The retriever is used to retrieve the retrieval data matching the input query data from a preset database.

[0023] Among them, the retrieval data may include multiple types, for example, the retrieval data may be text data, image data, audio data or video data, etc. It should be noted that the documents or related documents mentioned above may include multiple types, such as text documents, image documents, audio documents or video documents, etc., and the above documents may also belong to one of the retrieval data. The database may be a database constructed by a specified data source (such as a data source outside the current institution or a data source consisting of large-scale data pre-constructed by the current institution, etc.). The retriever can be constructed in a variety of different ways. For example, the retriever can be constructed by specifying an algorithm or model. Specifically, the retriever can be constructed by a Boolean retrieval algorithm, a truncation retrieval algorithm, an original text retrieval algorithm, etc., or the retriever can be constructed by a neural network model, a large model, etc., which can be set according to actual conditions, and the embodiments of this specification do not limit this.

[0024] In implementation, each clarification query data may be input into a retriever respectively, and the retriever retrieves the search data matching each clarification query data from a preset database to obtain the search data corresponding to each clarification query data.

[0025] In step S208, correlation information between the retrieval data corresponding to each clarification query data and the target question data is determined, and the retrieval data corresponding to each clarification query data is rewritten based on the determined correlation information to obtain rewritten retrieval data.

[0026] In implementation, the retrieval data corresponding to each clarified query data can be expanded through semantic analysis in the chain rewriting mechanism, specifically, Figure 5 As shown in the lower part, through the semantic analysis in the chain rewriting mechanism, the correlation between the retrieval data and the target question data can be analyzed according to the target question data and the retrieval data obtained by retrieval, so as to determine the correlation information between the retrieval data and the target question data corresponding to each clarification query data. The correlation between the retrieval data and the target question data can be determined by converting the data into vectors and then calculating the distance between the vectors. Alternatively, the correlation between the two data can be determined by other methods, which can be set according to actual conditions. Then, the retrieval data corresponding to the corresponding clarification query data can be rewritten based on the determined correlation information. Specifically, a correlation threshold can be set in advance, and based on the correlation threshold, the retrieval data whose correlation value in the correlation information is greater than the correlation threshold can be searched from the retrieval data. The found retrieval data can be retained, and the remaining retrieval data can be deleted, so as to achieve the purpose of rewriting the retrieval data corresponding to each clarification query data. Finally, the rewritten retrieval data can be obtained, or other relevant information can be further calculated based on the obtained correlation information (for example, the part of the retrieval data related to the target question data, the data that can be used as the answer to the question corresponding to the target question data, etc.), and the other relevant information obtained by calculation can be combined with the correlation information to rewrite the retrieval data corresponding to each clarification query data, etc., which can be set specifically according to actual conditions, and the embodiments of this specification do not limit this.

[0027] In step S210, based on the rewritten retrieval data, the reader determines the answer data corresponding to the target question data and outputs the answer data. The reader is used to predict the answer data of the question data input by the user based on the input retrieval data.

[0028] Among them, the reader can be constructed in a variety of different ways. For example, the reader can be constructed by a specified algorithm or model. Specifically, the reader can be constructed by a classification algorithm, a clustering algorithm, etc., or the reader can be constructed by a neural network model, a large model, etc. The specific setting can be based on actual conditions, and the embodiments of this specification do not limit this.

[0029] In implementation, the rewritten search data can be input into the reader, and the rewritten search data can be analyzed and processed by the specified algorithm or model in the reader to obtain the answer data corresponding to the target question data. Specifically, the reader is constructed by a neural network model, and the corresponding training samples can be obtained. The training samples are used to train the neural network model to obtain the trained neural network model, and then the reader can be obtained. At this time, the reader can predict the answer data of the question data input by the user based on the input search data. The rewritten search data can be input into the above reader, and the reader can predict the answer data of the target question data based on the rewritten search data, and can output the predicted answer data.

[0030] It should be noted that if the reader determines that the rewritten retrieval data is insufficient to answer the question corresponding to the target question data, the reader may use content such as "cannot be answered", "Unknown", etc. as answer data, and then the reader may output the above answer data.

[0031] The embodiment of the present specification provides a data processing method, which is applied to an open domain question answering system. The open domain question answering system includes a retriever and a reader. By receiving target question data input by a user, the user intent corresponding to the target question data is expanded based on the content of the target question data to obtain one or more different clarification query data. Then, each clarification query data is input into the retriever to obtain the retrieval data corresponding to each clarification query data. The retriever is used to retrieve the retrieval data matching the input query data from a preset database. After that, the correlation information between the retrieval data corresponding to each clarification query data and the target question data can be determined, and the retrieval data corresponding to each clarification query data can be rewritten based on the determined correlation information to obtain the rewritten retrieval data. Finally, the answer data corresponding to the target question data can be determined by the reader based on the rewritten retrieval data, and the answer data can be output. The reader is used to predict the answer data of the question data input by the user based on the input retrieval data. In this way, the original target question data is rewritten into clarification query data that can further express the user intent and cover more keywords by using an effective query rewriting method, which can not only ensure that the expanded content meets the user intent and maintains the semantic consistency, but also effectively improve the recall rate of the relevant retrieval data. In addition, by utilizing an effective retrieval data rewriting method, the retrieval data corresponding to each clarification query data is rewritten through the correlation information between the retrieval data corresponding to each clarification query data and the target question data, and converted into data related to the target question data, so that the question and answer system can use the relevant data to correctly answer the questions corresponding to the target question data, thereby effectively reducing or avoiding the occurrence of task mismatches in open domain question and answer scenarios.

[0032] In practical applications, the specific processing methods of the above steps S202 to S210 can be various. This embodiment provides a specific implementation method, that is, after a round of data retrieval and retrieval data rewriting, the rewritten retrieval data is input into the reader to generate corresponding answer data. Specifically, the target question data input by the user is input and received; based on the content of the target question data, the user intention corresponding to the target question data is extended to obtain one or more different clarification query data; each clarification query data is input into the retriever respectively to obtain the retrieval data corresponding to each clarification query data, and the retriever is used to retrieve the retrieval data matching the input query data from the preset database; the correlation information between the retrieval data corresponding to each clarification query data and the target question data is determined, and the retrieval data corresponding to each clarification query data is rewritten based on the determined correlation information to obtain the rewritten retrieval data; the rewritten retrieval data is input into the reader to obtain the answer data corresponding to the target question data, and the answer data is output, and the reader is used to predict the answer data of the question data input by the user according to the input retrieval data. The above specific processing process can be referred to the above related content, which will not be repeated here.

[0033] Based on the above implementation, the relevant contents may include the following improvements: In practical applications, the specific processing methods of the above step S204 can be varied. An optional processing method is provided below, such as Figure 6 As shown, the processing may specifically include the following steps S2042 and S2044.

[0034] In step S2042, based on the content of the target question data, the intention extension prompt information corresponding to the target question data is generated.

[0035] Among them, the intent-extended prompt information can be prompt information for the user's intention corresponding to the target question data, and the prompt information can be the Prompt in the large language model. For example, the target question data is "What type of mobile phone is it?", and the corresponding intent-extended prompt information can be "Please provide some information related to "mobile phone"", etc., which can be set according to actual conditions.

[0036] In implementation, the semantic analysis in the chain rewriting mechanism can be used to identify and focus on the topics and key information contained in the target question data, and prompt information that can be used for a large language model can be generated based on the identified topics and key information, that is, the intent extension prompt information corresponding to the target question data.

[0037] In step S2044, the intention expansion prompt information is input into the pre-trained first language model to expand the user intention corresponding to the target question data to obtain one or more different clarification query data.

[0038] Among them, the first largest language model can be a model composed of billions of parameters, based on a deep learning algorithm, and pre-trained using a large-scale corpus. For example, the first largest language model can be a model built by LLM, a large model based on BERT, or a large model based on the Transfomer architecture, etc. Specifically, the first largest language model can include GPT-3, ERNIEBot, etc., which can be set according to actual conditions.

[0039] In implementation, the model architecture of the first language model can be constructed according to actual conditions (for example, the model architecture of the first language model can be constructed based on the Transformer module, etc.), and then a large number of training samples can be obtained, and the first language model can be trained using the training samples to obtain the trained first language model. In another embodiment, the first language model that has been pre-trained can also be directly obtained, and then a certain amount of sample data can be obtained, and the pre-trained first language model can be fine-tuned using the sample data, and finally, the fine-tuned first language model can be obtained.

[0040] After obtaining the intention expansion prompt information in the above manner, the intention expansion prompt information can be input into the above-mentioned first language model, and the user intention corresponding to the target question data can be expanded through the first language model to expand the target question data into one or more new question data that meet the user's intention, so as to obtain one or more different clarification query data.

[0041] In practical applications, in addition to semantic analysis, the chain rewriting mechanism can also be set up with semantic augmentation. Semantic augmentation can use the analysis information obtained from semantic analysis to guide the second language model to generate faithful and consistent extensions based on the original target question data, thereby bridging the task mismatch problem between the retriever and the reader. First, semantic augmentation can be used to expand the clarifying query data, such as Figure 7 As shown, please refer to the processing of the following steps S212 and S214 for details.

[0042] In step S212, based on each clarification query data, query expansion prompt information corresponding to each clarification query data is generated.

[0043] The query expansion prompt information may be prompt information for further clarifying the clarification query data, wherein the prompt information may be the Prompt in the large language model. Specifically, for example, if the clarification query data is "What type of movie is the movie "Cell Phone"?", the corresponding query expansion prompt information may be "Please provide some keywords related to the movie "Cell Phone"", etc. In implementation, the semantic enhancement in the chain rewriting mechanism can be used to obtain the topic and key information contained in the target question data based on semantic analysis, and can be combined with the key information contained in each clarification query data to generate prompt information that can be used in a large language model, that is, the query expansion prompt information corresponding to each clarification query data.

[0044] In step S214, the query expansion prompt information is input into the pre-trained second largest language model to perform expansion processing on each clarification query data to obtain one or more different supplementary keywords corresponding to each clarification query data.

[0045] Among them, the second largest language model can be a model composed of hundreds of millions of parameters, based on a deep learning algorithm, and pre-trained using a large-scale corpus. For example, the second largest language model can be a model built by LLM, a large model built based on BERT, or a large model based on the Transformer architecture, etc. Specifically, the second largest language model can include GPT-3, ERNIEBot, etc., which can be set according to actual conditions. The second largest language model can be the same as the first largest language model mentioned above, or it can be different from the first largest language model, which can be set according to actual conditions.

[0046] In implementation, the model architecture of the second largest language model can be constructed according to actual conditions (for example, the model architecture of the second largest language model can be constructed based on the Transformer module, etc.), and then a large number of training samples can be obtained, and the second largest language model can be trained using the training samples to obtain the trained second largest language model. In another embodiment, the pre-trained second largest language model can also be directly obtained, and then a certain amount of sample data can be obtained, and the pre-trained second largest language model can be fine-tuned using the sample data, and finally, the fine-tuned second largest language model can be obtained.

[0047] The query expansion prompt information can be input into the second language model, and the query expansion prompt information corresponding to each clarification query data can be expanded by the second language model to add one or more supplementary information that meets the user's intention and / or can serve as keywords for the clarification query data to each clarification query data, thereby obtaining one or more different supplementary keywords corresponding to each clarification query data. Figure 8As shown, the clarification query data is “What genre is the TV series The Trip?”, and the supplementary keywords output after the clarification query data is input into the second largest language model may include “director”, “television program”, “episodes”, etc. The clarification query data is “What genre is the film The Trip?”, and the supplementary keywords output after the clarification query data is input into the second largest language model may include “director”, “storyline”, “main characters”, etc. The clarification query data is “What genre is the book The Trip?”, and the supplementary keywords output after the clarification query data is input into the second largest language model may include “author”, “storyline”, “theme”, “summary”, etc. The clarification query data is “What type of movie is the movie “Cell Phone”?”, and the supplementary keywords output after the clarification query data is input into the second largest language model may include “director”, “introduction”, “screenwriter”, “starring”, “movie type”, etc., which can be set specifically according to actual conditions.

[0048] Based on the processing of the above steps S212 and S214, the specific processing method of the above step S206 can be varied. An optional processing method is provided below, which may specifically include the following: input each clarification query data and one or more different supplementary keywords corresponding to each clarification query data into the retriever respectively to obtain the retrieval data corresponding to each clarification query data.

[0049] By clarifying each query data and its corresponding supplementary keywords, the retriever can retrieve more and more accurate data as retrieval data, thereby further bridging the task mismatch problem between the retriever and the reader.

[0050] In practical applications, there are various specific processing methods for determining the correlation information between the search data corresponding to each clarification query data and the target question data in the above step S208. An optional processing method is provided below, such as Fig. 9 As shown, the processing may specifically include the following steps S2082 and S2084.

[0051] In step S2082, relevance prompt information is generated based on the search data and target question data corresponding to each clarification query data.

[0052] Among them, the relevance prompt information can be prompt information of the relevance between the retrieval data and the target question data, and the prompt information can be the Prompt in the large language model, such as, "Please provide the relevance between the retrieval data 1 "xxxx" and the target question data "xxxx"".

[0053] In implementation, the semantic analysis in the chain rewriting mechanism can be used to identify the topics and key information contained in the retrieval data and the target question data, respectively, and then prompt information that can be used in a large language model, that is, relevance prompt information, can be generated.

[0054] In step S2084, based on the relevance prompt information, the relevance information between the retrieval data corresponding to each clarification query data and the target question data, as well as the portion of data in the retrieval data corresponding to each clarification query data whose relevance to the target question data is higher than a preset threshold, are determined through a pre-trained third language model; and based on the above relevance information and the portion of data in the retrieval data corresponding to each clarification query data whose relevance to the target question data is higher than the preset threshold, the candidate answer data for the target question data corresponding to each retrieval data is determined.

[0055] Among them, the third largest language model can be a model composed of hundreds of millions of parameters, based on a deep learning algorithm, and pre-trained using a large-scale corpus. For example, the third largest language model can be a model built by LLM, a large model built based on BERT, or a large model based on the Transformer architecture, etc. Specifically, the third largest language model can include GPT-3, ERNIEBot, etc., which can be set according to actual conditions. The third largest language model can be the same as the first and second largest language models mentioned above, or it can be different from the first and second largest language models, which can be set according to actual conditions. The preset threshold can be set according to actual conditions, such as 80% or 70%, etc., which can be set according to actual conditions.

[0056] In implementation, the model architecture of the third language model can be constructed according to actual conditions (for example, the model architecture of the third language model can be constructed based on the Transformer module, etc.), and then a large number of training samples can be obtained, and the third language model can be trained using the training samples to obtain the trained third language model. In another embodiment, the third language model that has been pre-trained can also be directly obtained, and then a certain amount of sample data can be obtained, and the pre-trained third language model can be fine-tuned using the sample data, and finally, the fine-tuned third language model can be obtained.

[0057] The relevance prompt information can be input into the third language model, and the relevance between the retrieval data corresponding to each clarification query data and the target question data can be calculated by the third language model (wherein, the relevance can be achieved in a variety of ways, for example, the similarity between the retrieval data corresponding to each clarification query data and the target question data can be calculated, and the calculated similarity value can be used as the relevance, or the Pearson correlation coefficient algorithm or the Spearman correlation coefficient algorithm can be used to calculate, and the calculated result can be used as the relevance), and then the relevance information between the retrieval data corresponding to each clarification query data and the target question data can be determined (contained in Figure 8 In addition, Figure 8 As shown, it is also possible to determine the part of the data (included in Figure 8 ), and based on the above-mentioned correlation information and the partial data whose correlation with the target question data in the retrieval data corresponding to each clarification query data is higher than the preset threshold, the candidate answer data of the target question data corresponding to each retrieval data that can be obtained can be determined. It should be noted that if the above-mentioned correlation information and the partial data whose correlation with the target question data in the retrieval data corresponding to each clarification query data is higher than the preset threshold are insufficient to answer the question corresponding to the target question data, then "cannot be answered" or "Unknown" can be used as candidate answer data.

[0058] Based on the processing of the above-mentioned steps S2082 and S2084, the retrieval data corresponding to each clarification query data is rewritten based on the determined relevance information in the above-mentioned step S208, and the specific processing method for obtaining the rewritten retrieval data can be varied. An optional processing method is provided below, which may specifically include the following contents: based on the determined relevance information and the candidate answer data of the target question data corresponding to each retrieval data, and the partial data in the retrieval data corresponding to each clarification query data whose correlation with the target question data is higher than a preset threshold, the retrieval data corresponding to each clarification query data is rewritten to obtain the rewritten retrieval data.

[0059] In implementation, Figure 8As shown, the search data retrieved by the retriever can be reconstructed through semantic enhancement in the chain rewriting mechanism, and the content irrelevant to the target question data in the search data can be deleted, so that the obtained relevant content can be rewritten into a statement that can be directly used to answer the target question data. Specifically, based on the part of the search data corresponding to each clarification query data whose correlation with the target question data is higher than a preset threshold, combined with the determined correlation information (contained in Figure 8 and the candidate answer data of the target question data corresponding to each retrieval data, retain the partial data whose correlation with the target question data in the retrieval data corresponding to each clarification query data is higher than the preset threshold, and delete the remaining partial data (including the candidate answer data in Figure 8 The final data is used as the rewritten retrieval data.

[0060] In practical applications, the above-mentioned retrieval data corresponding to each clarification query data is rewritten based on the determined correlation information and the candidate answer data of the target question data corresponding to each retrieval data, as well as the partial data in the retrieval data corresponding to each clarification query data whose correlation with the target question data is higher than a preset threshold. The specific processing method for obtaining the rewritten retrieval data can be varied. An optional processing method is provided below. For details, please refer to the processing of the following steps A2 and A4.

[0061] In step A2, if the candidate answer data of the target question data corresponding to the first retrieval data included in the retrieval data corresponding to the obtained clarification query data is not unanswerable, then based on the determined correlation information and the candidate answer data of the target question data corresponding to the first retrieval data, and the partial data in the first retrieval data whose correlation with the target question data is higher than the preset threshold, the partial data in the first retrieval data whose correlation with the target question data is higher than the preset threshold is retained, and the remaining part is deleted, so as to rewrite the first retrieval data and obtain the rewritten retrieval data.

[0062] In step A4, if the candidate answer data of the target question data corresponding to the second search data included in the search data corresponding to the obtained clarification query data is unanswerable, the second search data is used as the rewritten search data.

[0063] In practical applications, the specific processing methods of the above steps S202 to S210 can be varied. This embodiment provides another specific implementation method, that is, after two rounds of data retrieval and retrieval data rewriting, the rewritten retrieval data is input into the reader to generate corresponding answer data. Specifically, in practical applications, in many cases, the retrieved retrieval data may be relevant, but may not be sufficient to point to the correct answer. These retrieval data may mislead the reader to generate wrong answers, thereby reducing the effectiveness and robustness of the open domain question-answering system. In order to collect more relevant data to improve the possibly incorrect answers, the feedback provided in the semantic analysis in the chain rewriting mechanism can be used to correct the query data. Based on this, Fig.10 As shown, the processing of the above step S210 may include the processing of the following steps S21002 to S21010.

[0064] In step S21002, based on the rewritten search data, candidate answer data of the target question data corresponding to the search data before the rewriting process of the rewritten search data is obtained.

[0065] In implementation, based on the rewritten retrieval data, candidate answer data of the target question data corresponding to the retrieval data before the rewriting process of the rewritten retrieval data can be obtained from the candidate answer data determined in step S2084.

[0066] In step S21004, the target question data is corrected based on the obtained candidate answer data to obtain corrected query data.

[0067] In implementation, the target question data can be modified based on the obtained candidate answer data in a variety of different ways. For example, the target question data can be modified based on the obtained candidate answer data, and the modified data can be used as the modified query data. For example, if the obtained candidate answer data is "Situation comedy", the target question data "What genre is The Trip?" can be modified to "What genre is the TV series The Trip?", etc.; or, the obtained candidate answer data can be fused with the target question data to obtain fused data, and the fused data can be used as the modified query data, etc. The specific setting can be based on actual conditions.

[0068] In step S21006, the modified query data is input into the retriever to obtain the target search data corresponding to the modified query data.

[0069] In step S21008, target correlation information between the target retrieval data and the target question data is determined, and the target retrieval data is rewritten based on the determined target correlation information to obtain rewritten target retrieval data.

[0070] The specific processing process of the above step S21008 can refer to the relevant content of the above step S208, which will not be repeated here.

[0071] In step S21010, the rewritten target search data is input into the reader to obtain the answer data corresponding to the target question data.

[0072] Based on the above implementation, the relevant contents may include the following improvements: In practical applications, the specific processing methods of the above step S21002 can be varied. An optional processing method is provided below, such as Fig.11 As shown, the processing may specifically include the following steps S210022 to S210026.

[0073] In step S210022, based on the rewritten retrieval data, the probability that the answer data of the target question data predicted using the rewritten retrieval data is unanswerable is determined.

[0074] In implementation, the probability that the answer data of the target question data predicted by the rewritten retrieval data is unanswerable can be pre-evaluated. This can be achieved in a variety of different ways. For example, the probability distribution (such as normal distribution, etc.) of the answer data obtained by using the rewritten retrieval data to predict the target question data can be analyzed and determined. Based on the determined probability distribution, the probability that the answer data of the target question data predicted by the rewritten retrieval data is unanswerable can be determined. Alternatively, Fig.12 As shown, in the above step S2084, based on the relevance prompt information, through the pre-trained third language model, in addition to determining the relevance information between the retrieval data corresponding to each clarification query data and the target question data, and the part of the data in the retrieval data corresponding to each clarification query data whose relevance to the target question data is higher than the preset threshold, and determining the candidate answer data of the target question data corresponding to each retrieval data based on the above relevance information and the part of the data in the retrieval data corresponding to each clarification query data whose relevance to the target question data is higher than the preset threshold, it is also possible to determine the probability that the answer data of the target question data predicted by the rewritten retrieval data is unanswerable, etc., which can be specifically set according to actual conditions. Based on the above processing, based on the rewritten retrieval data, the probability that the answer data of the target question data predicted by the rewritten retrieval data is unanswerable can be obtained from the above determined probability of being unanswerable.

[0075] In step S210024, based on the determined probability, correlation information between the target question data and the rewritten retrieval data is determined.

[0076] In implementation, in order to effectively distinguish the data related to the target question data and the interference data in the rewritten retrieval data, the feedback provided in the chain rewriting mechanism can be used to estimate the relevance between the target question data and the rewritten retrieval data at a deeper semantic level. Fig.12 For example, Fig.12 As shown, the probability of answer data such as "Unknown" or "Unable to answer" output when predicting candidate answers using semantic analysis in the chain rewriting mechanism can be used as an indicator for estimating the relevance between the target question data and the rewritten retrieval data pair. The higher the probability that the third language model outputs answer data such as "Unknown" or "Unable to answer", the more difficult it is for the rewritten retrieval data to be used to answer the user's target question data, and the lower the relevance to the target question data. The lower the probability that the third language model outputs answer data such as "Unknown" or "Unable to answer", the more likely it is that the rewritten retrieval data can be used to answer the user's target question data, and the higher the relevance to the target question data.

[0077] The rewritten search data with a higher probability value corresponding to the answer data such as "Unknown" or "Unable to answer" output by the third language model can be regarded as having a lower correlation with the target question data, and then the correlation information between the rewritten search data and the target question data can be determined. The correlation information between the rewritten search data and the target question data can be determined in the above manner. Fig.12 As shown, the probability value of the third largest language model outputting "Unknown" corresponding to the rewritten retrieval data D1 is 0.133, the probability value of the third largest language model outputting "Unknown" corresponding to the rewritten retrieval data D2 is 0.861, and the probability value of the third largest language model outputting "Unknown" corresponding to the rewritten retrieval data D3 is 0.002. Therefore, the rewritten retrieval data D3 has the highest correlation with the target question data, and the rewritten retrieval data D2 has the lowest correlation with the target question data.

[0078] In step S210026, based on the correlation information between the target question data and the rewritten retrieval data, one or more rewritten retrieval data are selected from the rewritten retrieval data, and the candidate answer data of the target question data corresponding to the retrieval data before the rewriting process is obtained for each selected rewritten retrieval data.

[0079] In implementation, a corresponding correlation threshold may be pre-set, and based on the correlation information between the target question data and the rewritten search data, one or more rewritten search data with a correlation greater than the correlation threshold may be selected from the rewritten search data, or the rewritten search data containing the specified keywords may be selected, etc. The specific setting may be based on actual conditions, and this specification embodiment does not limit this. Then, the candidate answer data of the target question data corresponding to the search data before the rewriting process of each selected rewritten search data may be obtained from the candidate answer data determined in step S2084.

[0080] In practical applications, the specific processing methods for selecting one or more rewritten search data from the rewritten search data based on the correlation information between the target question data and the rewritten search data in the above step S210026 can be various. An optional processing method is provided below, such as Fig.13 As shown, the processing may specifically include the following steps S2100262 and S2100264.

[0081] In step S2100262, based on the correlation information between the target question data and the rewritten search data, the rewritten search data is sorted to obtain sorted data.

[0082] In step S2100264, rewritten retrieval data that satisfies a preset arrangement order is selected from the sorted data.

[0083] In implementation, multiple rewritten retrieval data may be sorted from large to small based on the relevance between each rewritten retrieval data and the target question data to obtain sorted data, and then Top-k rewritten retrieval data may be selected from the sorted data.

[0084] In practical applications, the specific processing methods of the above step S21004 can be varied. An optional processing method is provided below, which may specifically include the following: concatenating the obtained candidate answer data with the target question data to correct the target question data and obtain corrected query data.

[0085] In implementation, Fig.14 As shown, the target question data can be corrected based on the target question data and the obtained candidate answer data, and can be combined with the subject and key information contained in the target question data obtained by semantic analysis, as well as the key information contained in the retrieval data. That is, the target question data can be concatenated with the obtained candidate answer data to obtain a representation that can be directly used to answer the target question data, i.e., the query data is corrected, and the answer data corresponding to the target question data can be subsequently determined by the reader.

[0086] For the above processing, the candidate answer (CandidateAnswers) data extracted from the Top-k rewritten retrieval data can be used as automatic feedback to enhance the original target question data. The above candidate answer data can be spliced ​​with the original target question data to obtain the refined query (Refined Queries) data. These revised query data can provide more retrieval data related to the candidate answer data, helping the reader to improve the wrong answers extracted from the misleading retrieval data, thereby obtaining more accurate and reliable answer data. In addition, the feedback from different components can be used to self-correct the retrieval and reading processes, so that there is more interaction between the retrieval step and the reading step.

[0087] In addition, there are many specific processing methods for determining the target correlation information between the target retrieval data and the target question data in the above step S21008. An optional processing method is provided below, which may specifically include the following: generating target correlation prompt information based on the target retrieval data and the target question data; based on the target correlation prompt information, determining the target correlation information between the target retrieval data and the target question data, and the portion of data in the target retrieval data whose correlation with the target question data is higher than a preset threshold through a third language model, and determining the target candidate answer data of the target question data corresponding to the target retrieval data based on the target correlation information and the portion of data in the target retrieval data whose correlation with the target question data is higher than the preset threshold.

[0088] The specific processing process of the above content can refer to the relevant content of the aforementioned steps S2082 and S2084, which will not be repeated here.

[0089] Based on the above processing, in the above step S21008, the target retrieval data is rewritten based on the determined target relevance information, and the specific processing method for obtaining the rewritten target retrieval data can be varied. An optional processing method is provided below, which may specifically include the following contents: based on the determined target relevance information and the target candidate answer data of the target question data corresponding to the target retrieval data, and some data in the target retrieval data whose correlation with the target question data is higher than a preset threshold, the target retrieval data is rewritten to obtain the rewritten target retrieval data.

[0090] The specific processing process of the above content can be found in the aforementioned related content, which will not be repeated here.

[0091] In addition, the target retrieval data is rewritten based on the determined target relevance information and the target candidate answer data of the target question data corresponding to the target retrieval data, as well as some data in the target retrieval data whose correlation with the target question data is higher than a preset threshold. The specific processing method for obtaining the rewritten target retrieval data can be varied. An optional processing method is provided below, which may specifically include the following: if the target candidate answer data of the target question data corresponding to the first target retrieval data included in the target retrieval data is not unanswerable, then based on the determined target relevance information and the target candidate answer data of the target question data corresponding to the first target retrieval data, as well as some data in the first target retrieval data whose correlation with the target question data is higher than a preset threshold, some data in the first target retrieval data whose correlation with the target question data is higher than the preset threshold is retained, and the remaining part is deleted to rewrite the first target retrieval data to obtain the rewritten target retrieval data; if the target candidate answer data of the target question data corresponding to the second target retrieval data included in the target retrieval data is unanswerable, then the second target retrieval data is used as the rewritten target retrieval data.

[0092] The specific processing process of the above content can refer to the above related content, which will not be repeated here. In practical applications, the reader can be constructed through a large-scale language model.

[0093] In order to verify the superiority of the above processing method, the embodiments of this specification provide some comparative verifications, as follows: Table 1

[0094] As shown in Table 1 (where the F1 index is an evaluation index that takes into account both precision and recall, and the EM index is an index used to evaluate the accuracy of a question-answering system), the open-domain question-answering system with the chain rewriting mechanism outperforms the open-domain question-answering system with the Vicuna baseline method on each of the above open-domain question-answering datasets, and the open-domain question-answering system with the chain rewriting mechanism improves its performance by more than 10 points on the NQ dataset, TriviaQA dataset, and PopQA dataset. The open-domain question-answering system with the chain rewriting mechanism also has similar improvements compared to the retrieval enhancement model Vicuna-13b+BM25.

[0095] In addition, by effectively handling the mismatch between question data and retrieval data and performing self-correction during retrieval and re-ranking, the Vicuna-based approach can achieve comparable performance to approaches based on larger-scale language models (InstructGPT, Text-Davinci-003). This demonstrates the great potential of medium-scale language models for open-domain question answering that can be effectively enhanced by providing high-quality contextual information about question data.

[0096] Finally, it can be observed that using GPT-3.5-Turbo as a reader can further improve the performance of the open-domain question answering system with a chain rewriting mechanism on all datasets, achieving high performance on various datasets, confirming the effectiveness of the open-domain question answering system with a chain rewriting mechanism.

[0097] Table 2

[0098] Table 2 shows the recall performance of retrieved data. It can be seen that on each of the above open-domain question answering datasets, the open-domain question answering system with the chain rewriting mechanism outperforms the query expansion method (Direct Expansion) that directly generates relevant terms based on the original question data. In particular, the open-domain question answering system with the chain rewriting mechanism improves Recall@5 and Recall@20 on PopQA for long-tail open-domain questions by +11.8 and +13.6 respectively (relatively improved by 24.6% and 26.2%). This shows that by adding clear topic information and details to the open-domain question data, the open-domain question answering system with the chain rewriting mechanism can provide reliable guidance and generate detailed and consistent information related to the question, thereby achieving high-quality query expansion and improving the recall of relevant retrieved data.

[0099] Table 3

[0100] As shown in Table 3, adding feedback to the BM25 ranking or BM25+UPR ranking to rerank can further improve the performance on the above datasets, especially the performance improvement on the Top-1 accuracy and Top-5 accuracy. This shows that the open-domain question answering system with a chain rewriting mechanism can effectively self-correct the reranking process.

[0101] The embodiment of the present specification provides a data processing method, which is applied to an open domain question answering system. The open domain question answering system includes a retriever and a reader. By receiving target question data input by a user, the user intent corresponding to the target question data is expanded based on the content of the target question data to obtain one or more different clarification query data. Then, each clarification query data is input into the retriever to obtain the retrieval data corresponding to each clarification query data. The retriever is used to retrieve the retrieval data matching the input query data from a preset database. After that, the correlation information between the retrieval data corresponding to each clarification query data and the target question data can be determined, and the retrieval data corresponding to each clarification query data can be rewritten based on the determined correlation information to obtain the rewritten retrieval data. Finally, the answer data corresponding to the target question data can be determined by the reader based on the rewritten retrieval data, and the answer data can be output. The reader is used to predict the answer data of the question data input by the user based on the input retrieval data. In this way, the original target question data is rewritten into clarification query data that can further express the user intent and cover more keywords by using an effective query rewriting method, which can not only ensure that the expanded content meets the user intent and maintains the semantic consistency, but also effectively improve the recall rate of the relevant retrieval data. In addition, by utilizing an effective retrieval data rewriting method, the retrieval data corresponding to each clarification query data is rewritten through the correlation information between the retrieval data corresponding to each clarification query data and the target question data, and converted into data related to the target question data, so that the question and answer system can use the relevant data to correctly answer the questions corresponding to the target question data, thereby effectively reducing or avoiding the occurrence of task mismatches in open domain question and answer scenarios.

[0102] In addition, semantic analysis and semantic enhancement are set in the chain rewriting mechanism to ensure that the rewritten content meets the user's intention while maintaining semantic consistency. In addition, the candidate answer data extracted from the rewritten retrieval data of Top-k can be used as automatic feedback to enhance the original target question data, thereby obtaining corrected query data. These corrected query data can provide more retrieval data related to the candidate answer data, helping the reader to improve the wrong answers extracted from the misleading retrieval data, thereby obtaining more accurate and reliable answer data. Moreover, the feedback from different components can be used to self-correct the retrieval and reading processes, so that there is more interaction between the retrieval step and the reading step.

[0103] The above is a data processing method provided in the embodiment of this specification. Based on the same idea, the embodiment of this specification also provides a data processing device, which is provided with an open domain question answering system. The open domain question answering system includes a searcher and a reader, such as Fig.15 shown.

[0104] The data processing device comprises: a question receiving module 1501, an expansion module 1502, a retrieval module 1503, a rewriting module 1504 and an answer output module 1505, wherein: The question receiving module 1501 receives the target question data input by the user; An expansion module 1502, based on the content of the target question data, performs expansion processing on the user intention corresponding to the target question data to obtain one or more different clarification query data; The retrieval module 1503 inputs each of the clarification query data into the retriever to obtain the retrieval data corresponding to each of the clarification query data. The retriever is used to retrieve the retrieval data matching the input query data from a preset database. A rewriting module 1504 determines correlation information between the search data corresponding to each of the clarification query data and the target question data, and rewrites the search data corresponding to each of the clarification query data based on the determined correlation information to obtain rewritten search data; The answer output module 1505 determines the answer data corresponding to the target question data through the reader based on the rewritten search data, and outputs the answer data. The reader is used to predict the answer data of the question data input by the user according to the input search data.

[0105] In the embodiment of this specification, the expansion module 1502 includes: A prompt information generating unit, which generates the intention extension prompt information corresponding to the target question data based on the content of the target question data; The expansion unit inputs the intention expansion prompt information into a pre-trained first language model to expand the user intention corresponding to the target question data to obtain one or more different clarification query data.

[0106] In the embodiment of this specification, the device further includes: A prompt generating module, based on each of the clarification query data, generates query expansion prompt information corresponding to each of the clarification query data; A question expansion module, inputting the query expansion prompt information into a pre-trained second language model to perform expansion processing on each of the clarification query data to obtain one or more different supplementary keywords corresponding to each of the clarification query data; The search module 1503 inputs each of the clarification query data and one or more different supplementary keywords corresponding to each of the clarification query data into the search engine to obtain search data corresponding to each of the clarification query data.

[0107] In the embodiment of this specification, the rewriting module 1504 includes: a prompt information generating unit, generating relevant prompt information based on the search data corresponding to each of the clarification query data and the target question data; a relevance determination unit, based on the relevance prompt information, determining, through a pre-trained third language model, relevance information between the retrieval data corresponding to each of the clarification query data and the target question data, and part of the retrieval data corresponding to each of the clarification query data whose relevance to the target question data is higher than a preset threshold, and determining candidate answer data for the target question data corresponding to each of the retrieval data based on the relevance information and part of the retrieval data corresponding to each of the clarification query data whose relevance to the target question data is higher than the preset threshold; The rewriting unit rewrites the retrieval data corresponding to each clarification query data based on the determined correlation information and the candidate answer data of the target question data corresponding to each retrieval data, and the partial data of the retrieval data corresponding to each clarification query data whose correlation with the target question data is higher than a preset threshold, to obtain the rewritten retrieval data.

[0108] In an embodiment of the present specification, if the candidate answer data of the target question data corresponding to the first retrieval data included in the retrieval data corresponding to the obtained clarification query data is not unanswerable, then based on the determined correlation information and the candidate answer data of the target question data corresponding to the first retrieval data, and the partial data in the first retrieval data whose correlation with the target question data is higher than a preset threshold, the rewriting unit retains the partial data in the first retrieval data whose correlation with the target question data is higher than a preset threshold, and deletes the remaining part, so as to rewrite the first retrieval data to obtain rewritten retrieval data; if the candidate answer data of the target question data corresponding to the second retrieval data included in the retrieval data corresponding to the obtained clarification query data is unanswerable, the second retrieval data is used as the rewritten retrieval data.

[0109] In the embodiment of this specification, the answer output module 1505 includes: A candidate answer acquisition unit, based on the rewritten search data, acquires candidate answer data of the target question data corresponding to the search data before the rewriting of the rewritten search data; A correction unit, which corrects the target question data based on the obtained candidate answer data to obtain corrected query data; A retrieval unit, inputting the modified query data into the retriever to obtain target retrieval data corresponding to the modified query data; A data rewriting unit, which determines target correlation information between the target retrieval data and the target question data, and rewrites the target retrieval data based on the determined target correlation information to obtain rewritten target retrieval data; The answer output unit inputs the rewritten target search data into the reader to obtain the answer data corresponding to the target question data.

[0110] In an embodiment of the present specification, the candidate answer acquisition unit determines, based on the rewritten retrieval data, the probability that the answer data of the target question data predicted by using the rewritten retrieval data is unanswerable; based on the determined probability, determines the correlation information between the target question data and the rewritten retrieval data; based on the correlation information between the target question data and the rewritten retrieval data, selects one or more rewritten retrieval data from the rewritten retrieval data, and obtains the candidate answer data of the target question data corresponding to the retrieval data before the rewriting process for each selected rewritten retrieval data.

[0111] In an embodiment of the present specification, the candidate answer acquisition unit sorts the rewritten retrieval data based on the correlation information between the target question data and the rewritten retrieval data to obtain sorted data; and selects the rewritten retrieval data that meets the preset arrangement order from the sorted data.

[0112] In the embodiment of the present specification, the correction unit concatenates the acquired candidate answer data with the target question data to correct the target question data and obtain corrected query data.

[0113] In the embodiments of this specification, the reader is constructed by using a large-scale language model.

[0114] The embodiment of the present specification provides a data processing device, which is provided with an open domain question answering system, wherein the open domain question answering system includes a retriever and a reader, which receives target question data input by a user, and based on the content of the target question data, performs expansion processing on the user intention corresponding to the target question data to obtain one or more different clarification query data, and then inputs each clarification query data into the retriever respectively to obtain the retrieval data corresponding to each clarification query data, and the retriever is used to retrieve the retrieval data matching the input query data from a preset database, and then the correlation between the retrieval data corresponding to each of the clarification query data and the target question data can be determined. Information, and based on the determined correlation information, the retrieval data corresponding to each of the clarification query data is rewritten to obtain the rewritten retrieval data. Finally, based on the rewritten retrieval data, the answer data corresponding to the target question data can be determined by the reader, and the answer data can be output. The reader is used to predict the answer data of the question data input by the user based on the input retrieval data. In this way, the original target question data is rewritten into clarification query data that can further express the user's intention and cover more keywords by using an effective query rewriting method, which can not only ensure that the expanded content meets the user's intention and maintains semantic consistency, but also effectively improve the recall rate of related retrieval data. In addition, by using an effective retrieval data rewriting method, the retrieval data corresponding to each clarification query data is rewritten through the correlation information between the retrieval data corresponding to each clarification query data and the target question data, and converted into data related to the target question data, so that the question answering system can use the related data to correctly answer the question corresponding to the target question data, thereby effectively reducing or avoiding the occurrence of task mismatch in the open domain question answering scenario.

[0115] In addition, the candidate answer data extracted from the Top-k rewritten retrieval data can be used as automatic feedback to enhance the original target question data, thereby obtaining corrected query data. These corrected query data can provide more retrieval data related to the candidate answer data, helping the reader to improve the wrong answers extracted from the misleading retrieval data, thereby obtaining more accurate and reliable answer data. Moreover, the feedback from different components can be used to self-correct the retrieval and reading processes, thereby allowing more interaction between the retrieval step and the reading step.

[0116] The above is a data processing device provided in the embodiment of this specification. Based on the same idea, the embodiment of this specification also provides a data processing device, wherein the data processing device is provided with an open domain question answering system, wherein the open domain question answering system includes a searcher and a reader, such as Fig.16 shown.

[0117] The data processing device may provide a terminal device or a server, etc. for the above embodiments.

[0118] The data processing device may have relatively large differences due to different configurations or performances, and may include one or more processors 1601 and memory 1602, and the memory 1602 may store one or more storage applications or data. Among them, the memory 1602 may be a short-term storage or a persistent storage. The application stored in the memory 1602 may include one or more modules (not shown in the figure), and each module may include a series of computer executable instructions in the data processing device. Furthermore, the processor 1601 may be configured to communicate with the memory 1602 and execute a series of computer executable instructions in the memory 1602 on the data processing device. The data processing device may also include one or more power supplies 1603, one or more wired or wireless network interfaces 1604, one or more input and output interfaces 1605, and one or more keyboards 1606.

[0119] Specifically in this embodiment, the data processing device includes a memory and one or more programs, wherein the one or more programs are stored in the memory, and the one or more programs may include one or more modules, and each module may include a series of computer executable instructions in the data processing device, and the one or more programs are configured to be executed by one or more processors, including computer executable instructions for performing the following: Receive target question data input by the user; Based on the content of the target question data, the user intention corresponding to the target question data is expanded to obtain one or more different clarification query data; Input each of the clarification query data into the retriever to obtain the retrieval data corresponding to each of the clarification query data, wherein the retriever is used to retrieve the retrieval data matching the input query data from a preset database; Determine the correlation information between the search data corresponding to each of the clarification query data and the target question data, and rewrite the search data corresponding to each of the clarification query data based on the determined correlation information to obtain the rewritten search data; Based on the rewritten retrieval data, the answer data corresponding to the target question data is determined by the reader and the answer data is output. The reader is used to predict the answer data of the question data input by the user according to the input retrieval data.

[0120] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the data processing device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0121] The embodiment of the present specification provides a data processing device, which is provided with an open domain question answering system, wherein the open domain question answering system includes a retriever and a reader, which receives target question data input by a user, and based on the content of the target question data, performs expansion processing on the user intention corresponding to the target question data to obtain one or more different clarification query data, and then inputs each clarification query data into the retriever respectively to obtain the retrieval data corresponding to each clarification query data, and the retriever is used to retrieve the retrieval data matching the input query data from a preset database, and then the correlation between the retrieval data corresponding to each of the clarification query data and the target question data can be determined. Information, and based on the determined correlation information, the retrieval data corresponding to each of the clarification query data is rewritten to obtain the rewritten retrieval data. Finally, based on the rewritten retrieval data, the answer data corresponding to the target question data can be determined by the reader, and the answer data can be output. The reader is used to predict the answer data of the question data input by the user based on the input retrieval data. In this way, the original target question data is rewritten into clarification query data that can further express the user's intention and cover more keywords by using an effective query rewriting method, which can not only ensure that the expanded content meets the user's intention and maintains semantic consistency, but also effectively improve the recall rate of related retrieval data. In addition, by using an effective retrieval data rewriting method, the retrieval data corresponding to each clarification query data is rewritten through the correlation information between the retrieval data corresponding to each clarification query data and the target question data, and converted into data related to the target question data, so that the question answering system can use the related data to correctly answer the question corresponding to the target question data, thereby effectively reducing or avoiding the occurrence of task mismatch in the open domain question answering scenario.

[0122] Furthermore, based on the above Figures 2 to 14 In the method shown, one or more embodiments of this specification further provide a storage medium, which is applied to an open domain question answering system, wherein the open domain question answering system includes a retriever and a reader, and the storage medium is used to store computer executable instruction information. In a specific embodiment, the storage medium can be a USB flash drive, an optical disk, a hard disk, etc. When the computer executable instruction information stored in the storage medium is executed by a processor, the following process can be implemented: Receive target question data input by the user; Based on the content of the target question data, the user intention corresponding to the target question data is expanded to obtain one or more different clarification query data; Input each of the clarification query data into the retriever to obtain the retrieval data corresponding to each of the clarification query data, wherein the retriever is used to retrieve the retrieval data matching the input query data from a preset database; Determine the correlation information between the search data corresponding to each of the clarification query data and the target question data, and rewrite the search data corresponding to each of the clarification query data based on the determined correlation information to obtain the rewritten search data; Based on the rewritten retrieval data, the answer data corresponding to the target question data is determined by the reader and the answer data is output. The reader is used to predict the answer data of the question data input by the user according to the input retrieval data.

[0123] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the above-mentioned storage medium embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0124] The embodiment of the present specification provides a storage medium, which is applied to an open domain question answering system. The open domain question answering system includes a retriever and a reader. By receiving target question data input by a user, the user intent corresponding to the target question data is expanded based on the content of the target question data to obtain one or more different clarification query data. Then, each clarification query data is input into the retriever to obtain the retrieval data corresponding to each clarification query data. The retriever is used to retrieve the retrieval data matching the input query data from a preset database. After that, the correlation information between the retrieval data corresponding to each clarification query data and the target question data can be determined, and the retrieval data corresponding to each clarification query data can be rewritten based on the determined correlation information to obtain the rewritten retrieval data. Finally, the answer data corresponding to the target question data can be determined by the reader based on the rewritten retrieval data, and the answer data can be output. The reader is used to predict the answer data of the question data input by the user based on the input retrieval data. In this way, the original target question data is rewritten into clarification query data that can further express the user intent and cover more keywords by using an effective query rewriting method, which can not only ensure that the expanded content meets the user intent and maintains the semantic consistency, but also effectively improve the recall rate of the relevant retrieval data. In addition, by utilizing an effective retrieval data rewriting method, the retrieval data corresponding to each clarification query data is rewritten through the correlation information between the retrieval data corresponding to each clarification query data and the target question data, and converted into data related to the target question data, so that the question and answer system can use the relevant data to correctly answer the questions corresponding to the target question data, thereby effectively reducing or avoiding the occurrence of task mismatches in open domain question and answer scenarios.

[0125] Furthermore, based on the above Figures 2 to 14 In one or more embodiments of the present specification, a computer program product is provided, which is applied to an open domain question answering system, wherein the open domain question answering system includes a searcher and a reader, and the computer program product includes a computer program. When the computer program in the computer program product is executed by a processor, the following process can be implemented: Receive target question data input by the user; Based on the content of the target question data, the user intention corresponding to the target question data is expanded to obtain one or more different clarification query data; Input each of the clarification query data into the retriever to obtain the retrieval data corresponding to each of the clarification query data, wherein the retriever is used to retrieve the retrieval data matching the input query data from a preset database; Determine the correlation information between the search data corresponding to each of the clarification query data and the target question data, and rewrite the search data corresponding to each of the clarification query data based on the determined correlation information to obtain the rewritten search data; Based on the rewritten retrieval data, the answer data corresponding to the target question data is determined by the reader and the answer data is output. The reader is used to predict the answer data of the question data input by the user according to the input retrieval data.

[0126] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the above-mentioned computer program product embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0127] The embodiment of the present specification provides a computer program product, which is applied to an open domain question answering system. The open domain question answering system includes a retriever and a reader. By receiving target question data input by a user, the user intent corresponding to the target question data is expanded based on the content of the target question data to obtain one or more different clarification query data. Then, each clarification query data is input into the retriever to obtain the retrieval data corresponding to each clarification query data. The retriever is used to retrieve the retrieval data matching the input query data from a preset database. After that, the correlation information between the retrieval data corresponding to each clarification query data and the target question data can be determined, and the retrieval data corresponding to each clarification query data can be rewritten based on the determined correlation information to obtain the rewritten retrieval data. Finally, the answer data corresponding to the target question data can be determined by the reader based on the rewritten retrieval data, and the answer data can be output. The reader is used to predict the answer data of the question data input by the user based on the input retrieval data. In this way, the original target question data is rewritten into clarification query data that can further express the user intent and cover more keywords by using an effective query rewriting method, which can not only ensure that the expanded content meets the user intent and maintains the semantic consistency, but also effectively improve the recall rate of the relevant retrieval data. In addition, by utilizing an effective retrieval data rewriting method, the retrieval data corresponding to each clarification query data is rewritten through the correlation information between the retrieval data corresponding to each clarification query data and the target question data, and converted into data related to the target question data, so that the question and answer system can use the relevant data to correctly answer the questions corresponding to the target question data, thereby effectively reducing or avoiding the occurrence of task mismatches in open domain question and answer scenarios.

[0128] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0129] In the 1990s, it was very clear whether the improvement of a technology was hardware improvement (for example, improvement of the circuit structure of diodes, transistors, switches, etc.) or software improvement (improvement of the method flow). However, with the development of technology, many improvements of the method flow today can be regarded as direct improvements of the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved method flow into the hardware circuit. Therefore, it cannot be said that the improvement of a method flow cannot be implemented with a hardware entity module. For example, a programmable logic device (PLD) (such as a field programmable gate array (FPGA)) is such an integrated circuit whose logical function is determined by the user's programming of the device. Designers can "integrate" a digital system on a PLD by programming themselves, without having to ask chip manufacturers to design and make dedicated integrated circuit chips. Moreover, nowadays, instead of manually making integrated circuit chips, this kind of programming is mostly implemented by "logic compiler" software, which is similar to the software compiler used when developing and writing programs, and the original code before compilation must also be written in a specific programming language, which is called hardware description language (HDL). There is not only one kind of HDL, but many kinds, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, RHDL (Ruby Hardware Description Language), etc. The most commonly used ones are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should also know that it is only necessary to program the method flow slightly in the above-mentioned hardware description languages ​​and program it into the integrated circuit, and then it is easy to obtain the hardware circuit that implements the logic method flow.

[0130] The controller may be implemented in any suitable manner, for example, the controller may take the form of a microprocessor or processor and a computer-readable medium storing a computer-readable program code (such as software or firmware) executable by the (micro)processor, a logic gate, a switch, an application-specific integrated circuit (ASIC), a programmable logic controller, and an embedded microcontroller. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicone Labs C8051F320. The memory controller may also be implemented as part of the control logic of the memory. It is also known to those skilled in the art that, in addition to implementing the controller in a purely computer-readable program code manner, the controller may be implemented in the form of a logic gate, a switch, an application-specific integrated circuit, a programmable logic controller, and an embedded microcontroller by logically programming the method steps. Therefore, such a controller may be considered as a hardware component, and the devices for implementing various functions included therein may also be considered as structures within the hardware component. Or even, the devices for implementing various functions may be considered as both software modules for implementing the method and structures within the hardware component.

[0131] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0132] For the convenience of description, the above devices are described in terms of functions and are divided into various units. Of course, when implementing one or more embodiments of this specification, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0133] Those skilled in the art will appreciate that the embodiments of this specification may be provided as methods, systems, or computer program products. Therefore, one or more embodiments of this specification may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, one or more embodiments of this specification may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0134] The embodiments of this specification are described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of this specification. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable fraud case serial and parallel device to produce a machine, so that the instructions executed by the processor of the computer or other programmable fraud case serial and parallel device generate instructions for implementing the processes in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0135] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable fraud case serial and parallel device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0136] These computer program instructions may also be loaded onto a computer or other programmable device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0137] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0138] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0139] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined in this article, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0140] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0141] Those skilled in the art will appreciate that the embodiments of this specification may be provided as methods, systems or computer program products. Therefore, one or more embodiments of this specification may be in the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Furthermore, one or more embodiments of this specification may be in the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0142] One or more embodiments of the present specification may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. One or more embodiments of the present specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules may be located in local and remote computer storage media, including storage devices.

[0143] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0144] The above description is only an embodiment of this specification and is not intended to limit this document. For those skilled in the art, this specification may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of this specification should be included in the scope of the claims of this specification.

Claims

1. A data processing method, applied to an open domain question answering system, wherein the open domain question answering system includes a retriever and a reader, the method comprising: Receive target question data input by the user; Based on the content of the target question data, the user intention corresponding to the target question data is expanded to obtain one or more different clarification query data; Input each of the clarification query data into the retriever to obtain the retrieval data corresponding to each of the clarification query data, wherein the retriever is used to retrieve the retrieval data matching the input query data from a preset database; Determine the correlation information between the search data corresponding to each of the clarification query data and the target question data, and rewrite the search data corresponding to each of the clarification query data based on the determined correlation information to obtain the rewritten search data; Based on the rewritten retrieval data, answer data corresponding to the target question data is determined by the reader, and the answer data is output. The reader is used to predict answer data of question data input by the user based on the input retrieval data.

2. The method according to claim 1, wherein the step of performing expansion processing on the user intent corresponding to the target question data based on the content of the target question data to obtain one or more different clarification query data comprises: Based on the content of the target question data, generating intention extension prompt information corresponding to the target question data; The intention expansion prompt information is input into a pre-trained first language model to expand the user intention corresponding to the target question data to obtain one or more different clarification query data.

3. The method according to claim 2, further comprising: Based on each of the clarification query data, generating query expansion prompt information corresponding to each of the clarification query data; Inputting the query expansion prompt information into a pre-trained second language model to perform expansion processing on each of the clarification query data to obtain one or more different supplementary keywords corresponding to each of the clarification query data; The step of inputting each of the clarification query data into the retriever to obtain the search data corresponding to each of the clarification query data comprises: Each of the clarification query data and one or more different supplementary keywords corresponding to each of the clarification query data are respectively input into the retriever to obtain the search data corresponding to each of the clarification query data.

4. The method according to claim 1, wherein determining the correlation information between the search data corresponding to each of the clarification query data and the target question data comprises: generating relevance prompt information based on the search data corresponding to each of the clarification query data and the target question data; Based on the relevance prompt information, determine, through a pre-trained third language model, relevance information between the retrieval data corresponding to each of the clarification query data and the target question data, as well as part of the retrieval data corresponding to each of the clarification query data whose relevance to the target question data is higher than a preset threshold, and determine candidate answer data for the target question data corresponding to each of the retrieval data based on the relevance information and part of the retrieval data corresponding to each of the clarification query data whose relevance to the target question data is higher than the preset threshold; The rewriting of the search data corresponding to each of the clarification query data based on the determined correlation information to obtain the rewritten search data includes: Based on the determined correlation information and the candidate answer data of the target question data corresponding to each retrieval data, as well as the partial data in the retrieval data corresponding to each clarification query data whose correlation with the target question data is higher than a preset threshold, the retrieval data corresponding to each clarification query data is rewritten to obtain rewritten retrieval data.

5. The method according to claim 4, rewriting the retrieval data corresponding to each clarification query data based on the determined correlation information and the candidate answer data of the target question data corresponding to each retrieval data, and the partial data of the retrieval data corresponding to each clarification query data whose correlation with the target question data is higher than a preset threshold, to obtain the rewritten retrieval data, comprises: If the candidate answer data of the target question data corresponding to the first search data contained in the search data corresponding to the obtained clarification query data is not unanswerable, based on the determined correlation information and the candidate answer data of the target question data corresponding to the first search data, and part of the data in the first search data whose correlation with the target question data is higher than a preset threshold, retain part of the data in the first search data whose correlation with the target question data is higher than a preset threshold, and delete the remaining part, so as to rewrite the first search data and obtain rewritten search data; If the candidate answer data of the target question data corresponding to the second search data included in the search data corresponding to the obtained clarification query data is unanswerable, the second search data is used as the rewritten search data.

6. The method according to claim 4, wherein the determining, by the reader, answer data corresponding to the target question data based on the rewritten search data comprises: Based on the rewritten search data, obtaining candidate answer data of the target question data corresponding to the search data before the rewriting of the rewritten search data; Correcting the target question data based on the acquired candidate answer data to obtain corrected query data; Inputting the modified query data into the retriever to obtain target search data corresponding to the modified query data; Determining target correlation information between the target search data and the target question data, and rewriting the target search data based on the determined target correlation information to obtain rewritten target search data; The rewritten target search data is input into the reader to obtain answer data corresponding to the target question data.

7. The method according to claim 6, wherein the step of obtaining, based on the rewritten search data, candidate answer data of the target question data corresponding to the search data before the rewriting of the rewritten search data comprises: determining, based on the rewritten search data, a probability that answer data of the target question data predicted using the rewritten search data is unanswerable; Based on the determined probability, determining correlation information between the target question data and the rewritten search data; Based on the correlation information between the target question data and the rewritten retrieval data, one or more rewritten retrieval data are selected from the rewritten retrieval data, and candidate answer data of the target question data corresponding to the retrieval data before the rewriting process is obtained for each selected rewritten retrieval data.

8. The method according to claim 7, wherein the selecting one or more rewritten search data from the rewritten search data based on the correlation information between the target question data and the rewritten search data comprises: sorting the rewritten search data based on the correlation information between the target question data and the rewritten search data to obtain sorted data; The rewritten retrieval data satisfying the preset arrangement order is selected from the sorted data.

9. The method according to claim 6, wherein the step of modifying the target question data based on the obtained candidate answer data to obtain modified query data comprises: The obtained candidate answer data is concatenated with the target question data to correct the target question data and obtain corrected query data.

10. The method according to any one of claims 1-9, wherein the reader is constructed by a large-scale language model.

11. A data processing device, provided with an open domain question answering system, wherein the open domain question answering system includes a searcher and a reader, the device comprising: A question receiving module receives target question data input by the user; An expansion module, based on the content of the target question data, performs expansion processing on the user intention corresponding to the target question data to obtain one or more different clarification query data; A retrieval module, inputting each of the clarification query data into the retriever to obtain retrieval data corresponding to each of the clarification query data, wherein the retriever is used to retrieve retrieval data matching the input query data from a preset database; a rewriting module, which determines correlation information between the search data corresponding to each of the clarification query data and the target question data, and rewrites the search data corresponding to each of the clarification query data based on the determined correlation information to obtain rewritten search data; The answer output module determines the answer data corresponding to the target question data through the reader based on the rewritten search data, and outputs the answer data. The reader is used to predict the answer data of the question data input by the user according to the input search data.

12. A data processing device, wherein the data processing device is provided with an open domain question answering system, wherein the open domain question answering system includes a retriever and a reader, and the data processing device comprises: processor; as well as a memory arranged to store computer executable instructions which, when executed, cause the processor to: Receive target question data input by the user; Based on the content of the target question data, the user intention corresponding to the target question data is expanded to obtain one or more different clarification query data; Input each of the clarification query data into the retriever to obtain the retrieval data corresponding to each of the clarification query data, wherein the retriever is used to retrieve the retrieval data matching the input query data from a preset database; Determine the correlation information between the search data corresponding to each of the clarification query data and the target question data, and rewrite the search data corresponding to each of the clarification query data based on the determined correlation information to obtain the rewritten search data; Based on the rewritten retrieval data, answer data corresponding to the target question data is determined by the reader, and the answer data is output. The reader is used to predict answer data of question data input by the user based on the input retrieval data.

Images