Question and answer model training method and device, question and answer method and device and electronic equipment

Abstract

The invention provides a question and answer model training method and device, a question and answer method and device and electronic equipment, and relates to the technical field of artificial intelligence, in particular to the technical fields of deep learning, natural language processing, large models, intelligent search and the like. According to the specific implementation scheme, a wrong question-answer sample is obtained; the wrong question and answer sample comprises a question prompt text and a wrong answer text corresponding to the question prompt text; obtaining a to-be-trained first question and answer model; according to the question prompt text and the error reply text, carrying out training processing on the first question and answer model to obtain a trained second question and answer model; wherein the second question and answer model is obtained by training wrong question and answer samples, so that the trained second question and answer model can avoid generating potential wrong answers, and the question and answer accuracy of the trained second question and answer model is improved.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, and in particular to technical fields such as deep learning, natural language processing, large models, and intelligent search, and in particular to a training method, a question-answering method, a device, and an electronic device for a question-answering model. Background Art

[0002] The current question-answering models are mainly trained using correct question-answering samples, which makes it difficult to avoid generating potential erroneous responses, resulting in low question-answering accuracy of the question-answering models. Summary of the invention

[0003] The present invention provides a question-answering model training method, a question-answering method, a device, and an electronic device.

[0004] According to one aspect of the present disclosure, a method for training a question-answering model is provided, the method comprising: obtaining an incorrect question-answering sample; the incorrect question-answering sample comprises a question prompt text and an incorrect reply text corresponding to the question prompt text; obtaining a first question-answering model to be trained; and training the first question-answering model according to the question prompt text and the incorrect reply text to obtain a trained second question-answering model.

[0005] According to another aspect of the present disclosure, a question-answering method is provided, the method comprising: obtaining a current question prompt text; obtaining a question-answering model; the question-answering model is determined based on the training method of the question-answering model as described above; inputting the current question prompt text into the question-answering model to obtain a reply text output by the question-answering model; and determining the reply text as the reply text corresponding to the current question prompt text.

[0006] According to another aspect of the present disclosure, a training device for a question-answering model is provided, the device comprising: a first acquisition module, used to acquire incorrect question-answering samples; the incorrect question-answering samples include a question prompt text and an incorrect reply text corresponding to the question prompt text; a second acquisition module, used to acquire a first question-answering model to be trained; and a training processing module, used to perform training processing on the first question-answering model according to the question prompt text and the incorrect reply text to obtain a trained second question-answering model.

[0007] According to another aspect of the present disclosure, a question-and-answer device is provided, the device comprising: a first acquisition module, used to acquire a current question prompt text; a second acquisition module, used to acquire a question-and-answer model; the question-and-answer model is determined based on the training method of the question-and-answer model as described above; a third acquisition module, used to input the current question prompt text into the question-and-answer model, and acquire a reply text output by the question-and-answer model; a determination module, used to determine the reply text as the reply text corresponding to the current question prompt text.

[0008] According to another aspect of the present disclosure, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the training method of the question-answering model proposed above in the present disclosure; or, execute the question-answering method proposed above in the present disclosure.

[0009] According to another aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable a computer to execute the training method of the question-answering model proposed above in the present disclosure; or, to execute the question-answering method proposed above in the present disclosure.

[0010] According to another aspect of the present disclosure, a computer program product is provided, including a computer program, which, when executed by a processor, implements the steps of the training method of the question-answering model proposed above in the present disclosure; or, implements the steps of the question-answering method proposed above in the present disclosure.

[0011] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure.

[0013] Figure 1 is a schematic diagram according to a first embodiment of the present disclosure;

[0014] Figure 2 is a schematic diagram according to a second embodiment of the present disclosure;

[0015] Figure 3 is a schematic diagram according to a third embodiment of the present disclosure;

[0016] Figure 4 is a schematic diagram according to a fourth embodiment of the present disclosure;

[0017] Figure 5 It is a schematic diagram of the determination of incorrect question and answer samples;

[0018] Figure 6 is a schematic diagram according to a fifth embodiment of the present disclosure;

[0019] Figure 7 is a schematic diagram according to a sixth embodiment of the present disclosure;

[0020] Figure 8It is a block diagram of an electronic device used to implement the training method of the question-answering model or the question-answering method of the embodiment of the present disclosure. DETAILED DESCRIPTION

[0021] The following is a description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0022] The current question-answering models are mainly trained using correct question-answering samples, which cannot avoid generating potential erroneous responses, resulting in low question-answering accuracy of the question-answering models.

[0023] In response to the above problems, the present disclosure proposes a training method, a question-answering method, a device and an electronic device for a question-answering model.

[0024] Figure 1 It is a schematic diagram according to the first embodiment of the present disclosure. It should be noted that the training method of the question-answering model of the embodiment of the present disclosure can be applied to the training device of the question-answering model, and the device can be configured in an electronic device so that the electronic device can perform the training function of the question-answering model.

[0025] Among them, the electronic device can be any device with computing capabilities, such as a personal computer (PC), a mobile terminal, a server, etc. The mobile terminal can be, for example, a vehicle-mounted device, a mobile phone, a tablet computer, a personal digital assistant, a wearable device, a smart speaker, a server, a server cluster, and other hardware devices with various operating systems, touch screens and / or display screens.

[0026] The training device of the question-answering model may also be software in an electronic device, such as training software of the question-answering model, etc. In the following embodiments, the execution subject is an electronic device as an example for description.

[0027] like Figure 1 As shown, the training method of the question-answering model may include the following steps:

[0028] Step 101, obtaining an incorrect question and answer sample; the incorrect question and answer sample includes a question prompt text and an incorrect answer text corresponding to the question prompt text.

[0029] In the embodiments of the present disclosure, the error reply text may be text that is irrelevant to the question prompt text; or, the error reply text may be text whose relevance to the question prompt text is less than or equal to a relevance threshold.

[0030] In one example, the incorrect question and answer sample can be determined by the error detection model. The question and answer sample can be input into the error detection model to obtain the detection result output by the error detection model; when the detection result indicates that the question and answer sample is wrong, the question and answer sample is determined as an incorrect question and answer sample. In another example, the incorrect question and answer sample can be selected from various question and answer sample annotation libraries. The question and answer sample annotation library includes various question and answer samples and corresponding annotation information; the annotation information indicates whether the question and answer sample is an incorrect question and answer sample.

[0031] In the embodiment of the present disclosure, the question prompt text may include question text and prompt text. The prompt text is used to guide the model to generate a specific type of answer or perform a specific task. The question text indicates the question that needs to be answered. The setting of the prompt text enables the question-answering model to be applicable to answering questions under various tasks; or enables the question-answering model to provide different types of answers to questions.

[0032] In an embodiment of the present disclosure, the incorrect question and answer sample may also include at least one of the following for determining the incorrect reply text: text search results, reasons for the incorrect question and answer; wherein the text search results are determined based on the question prompt text.

[0033] The question-answering model can communicate with a search engine, etc., obtain text search results in the search engine according to the question prompt text, and generate a reply text by combining the text search results and the question prompt text. The incorrect question-answering sample can have at least one of the following conditions: incorrect text search results; incorrect reply text.

[0034] Among them, the reasons for incorrect questions and answers may include at least one of the following: the matching degree between the reply text and the prompt text is less than or equal to the text matching threshold; the relevance between the reply text and the prompt text is less than or equal to the relevance threshold; the degree of logical consistency between each word in the reply text is less than or equal to the degree threshold; the matching degree between the reply text and the target grammar and target style is less than the grammatical style matching threshold respectively.

[0035] Among them, the setting of the reasons for incorrect questions and answers enables the question and answer model to learn the incorrect reasons for incorrect question and answer samples during the training process, and then overcome the incorrect question and answer samples caused by the incorrect reasons for question and answer, and further improve the accuracy of the trained question and answer model.

[0036] Among them, the setting of text search results enables the question-answering model to prompt text for questions in new fields, obtain relevant knowledge through text search processing, and then perform reply processing to improve the accuracy of the reply. In addition, during the training process, the question-answering model can fully learn the correlation between text search results and reply text, further improving the accuracy of the trained question-answering model.

[0037] Step 102, obtaining a first question-answering model to be trained.

[0038] In the disclosed embodiment, the first question-answering model to be trained may be a model obtained by training an initial question-answering model using a plurality of correct question-answering samples, wherein the initial question-answering model may be a pre-trained large language model.

[0039] Step 103: training the first question-answering model according to the question prompt text and the wrong answer text to obtain a trained second question-answering model.

[0040] The training method of the question-answering model of the embodiment of the present disclosure is as follows: obtaining incorrect question-answering samples; the incorrect question-answering samples include question prompt text and incorrect answer text corresponding to the question prompt text; obtaining a first question-answering model to be trained; training the first question-answering model according to the question prompt text and the incorrect answer text to obtain a trained second question-answering model; wherein the second question-answering model is trained using the incorrect question-answering samples, so that the trained second question-answering model can avoid generating potential incorrect answers, thereby improving the question-answering accuracy of the trained second question-answering model.

[0041] In order to further improve the accuracy of the questions and answers of the trained second question and answer model, the first question and answer model can be trained by combining the wrong question and answer samples generated during the use of the first question and answer model. Figure 2 As shown, Figure 2 is a schematic diagram according to a second embodiment of the present disclosure, Figure 2 The illustrated embodiment may include the following steps:

[0042] Step 201, obtaining a first historical question and answer sample of a first question and answer model, and feedback information for the first historical question and answer sample.

[0043] In the disclosed embodiment, for the first question-answering model, the electronic device may display a front-end page. In the front-end page, the electronic device may detect the input operation of the object and obtain the input question prompt text; input the question prompt text into the first question-answering model and obtain the answer text output by the first question-answering model; when detecting the object's related operation on the feedback control, obtain the object's feedback information. The feedback information may indicate at least one of the following: a correct answer and an incorrect answer.

[0044] Step 202: When the feedback information indicates that the answer is wrong, the first historical question and answer sample is determined as an incorrect question and answer sample.

[0045] In which, when the feedback information indicates that the answer is correct, the electronic device can determine the first historical question and answer sample as a correct question and answer sample.

[0046] In an embodiment of the present disclosure, in order to further expand the number of erroneous question and answer samples, so as to fully train the first question and answer model and further improve the accuracy of the trained second question and answer model, the electronic device may also perform the following process: obtaining a second historical question and answer sample of the first question and answer model without feedback information; determining a first matching degree between the second historical question and answer sample and each erroneous question and answer sample; and in the event that there is an erroneous question and answer sample whose first matching degree is greater than or equal to a matching degree threshold, determining the second historical question and answer sample as an erroneous question and answer sample.

[0047] Among them, the process of the electronic device determining the first matching degree between the second historical question and answer sample and each erroneous question and answer sample can, for example, be to determine the first semantic representation vector of the second historical question and answer sample and the second semantic representation vector of each erroneous question and answer sample; determine the similarity between the first semantic representation vector and each second semantic representation vector; and determine the first matching degree based on each similarity.

[0048] For example, the similarity between the first semantic representation vector of the second historical question and answer sample and the second semantic representation vector of the erroneous question and answer sample can be used as the first matching degree between the second historical question and answer sample and the erroneous question and answer sample.

[0049] Among them, determining the first matching degree in combination with the similarity between the first semantic representation vector of the second historical question and answer sample and the second semantic representation vector of the erroneous question and answer sample can improve the accuracy of the determined first matching degree.

[0050] In the embodiments of the present disclosure, it is also necessary to explain that in order to further expand the number of erroneous question and answer samples, the electronic device can obtain multiple question and answer samples from each question and answer sample library; for each question and answer sample, determine the first matching degree between the question and answer sample and each erroneous question and answer sample; in the case that there is an erroneous question and answer sample whose first matching degree is greater than or equal to the matching degree threshold, determine the question and answer sample as an erroneous question and answer sample.

[0051] Step 203, obtaining a first question-answering model to be trained.

[0052] Step 204: training the first question-answering model according to the question prompt text and the incorrect answer text to obtain a trained second question-answering model.

[0053] In an embodiment of the present disclosure, in one example, the process of the electronic device executing step 204 may be, for example, inputting the question prompt text into the first question and answer model to obtain the predicted reply text output by the first question and answer model; determining the fourth degree of match between the incorrect reply text and the predicted reply text; determining the first value of the loss function of the first question and answer model based on the fourth degree of match; and performing parameter adjustment processing on the first question and answer model based on the first value to obtain a second question and answer model.

[0054] Among them, the process of the electronic device determining the fourth matching degree between the erroneous reply text and the predicted reply text can, for example, be to determine a first vector of the erroneous reply text and a second vector of the predicted reply text; determine the similarity between the first vector and the second vector; and determine the similarity as the fourth matching degree between the erroneous reply text and the predicted reply text.

[0055] Among them, the first value of the loss function is determined according to the fourth matching degree between the wrong reply text and the predicted reply text, and then the parameters of the first question-answering model are adjusted, so that when the trained second question-answering model generates a reply for the question prompt text, it can generate a predicted reply text far away from the wrong reply text, thereby improving the accuracy of the predicted reply text.

[0056] In another example, the process of the electronic device executing step 204 may, for example, be to determine the correct answer text corresponding to the question prompt text; input the question prompt text into the first question and answer model to obtain the predicted answer text output by the first question and answer model; determine the fifth matching degree between the predicted answer text and the correct answer text, and the sixth matching degree between the predicted answer text and the incorrect answer text; determine the second value of the loss function of the first question and answer model based on the inverse of the fifth matching degree and the sixth matching degree; and perform parameter adjustment processing on the first question and answer model based on the second value to obtain a second question and answer model.

[0057] Among them, the process of the electronic device determining the correct answer text corresponding to the question prompt text can, for example, be to query the correct question and answer text library according to the question prompt text to obtain the first correct question and answer text therein; wherein the first question prompt text in the first correct question and answer text is matched with the question prompt text; and the answer text in the first correct question and answer text is determined as the correct answer text corresponding to the question prompt text.

[0058] Among them, the second value of the loss function is determined according to the sixth matching degree between the wrong reply text and the predicted reply text, and the fifth matching degree between the correct reply text and the predicted reply text, and then the parameters of the first question-answering model are adjusted, so that when the trained second question-answering model generates a reply for the question prompt text, it can generate a predicted reply text that is far away from the wrong reply text and close to the correct reply text, thereby further improving the accuracy of the predicted reply text.

[0059] The training method of the question-and-answer model of the disclosed embodiment is as follows: obtaining a first historical question-and-answer sample of a first question-and-answer model and feedback information for the first historical question-and-answer sample; determining the first historical question-and-answer sample as an incorrect question-and-answer sample when the feedback information indicates an incorrect answer; obtaining the first question-and-answer model to be trained; training the first question-and-answer model according to a question prompt text and an incorrect answer text to obtain a trained second question-and-answer model; wherein, training the first question-and-answer model in combination with the incorrect question-and-answer samples generated during the use of the first question-and-answer model can improve the question-and-answer accuracy of the trained second question-and-answer model.

[0060] Among them, in order to further improve the accuracy of the question and answer of the trained second question and answer model, each third question and answer sample generated during the use of the first question and answer model can be scored on the question and answer scoring index to determine whether the question and answer sample is an incorrect question and answer sample, and then the first question and answer model is trained in combination with the incorrect question and answer samples to obtain the second question and answer model. Figure 3 As shown, Figure 3 is a schematic diagram according to a third embodiment of the present disclosure, Figure 3 The illustrated embodiment may include the following steps:

[0061] Step 301, obtaining a third historical question and answer sample of the first question and answer model.

[0062] Step 302, determining the question and answer scoring index and the weight of the question and answer scoring index.

[0063] In the disclosed embodiment, the process of the electronic device executing step 302 may, for example, be to determine the domain category of the third historical question and answer sample; and obtain the question and answer scoring index corresponding to the domain category and the weight from the index database.

[0064] The indicator database may include question-and-answer scoring indicators corresponding to various field categories. The question-and-answer scoring indicators corresponding to the field categories may be preset, or determined according to the question-and-answer requirements under the field categories. Field categories such as medicine, machinery, communications, education, law, and art may be set according to actual needs.

[0065] Among them, the setting of question and answer scoring indicators under different field categories can set personalized question and answer scoring indicators for different field categories, thereby personalizing the determination of incorrect question and answer samples under each field category and improving the accuracy of incorrect question and answer samples under each field category.

[0066] In the disclosed embodiment, the third historical question and answer sample may include a historical question prompt text and a historical answer text; the historical question prompt text may include a historical question text and a historical prompt text. Correspondingly, the question and answer scoring index may include at least one of the following: a second matching degree between the historical answer text and the historical prompt text; a correlation degree between the historical answer text and the historical prompt text; a logical consistency degree between each word in the historical answer text; and a third matching degree between the historical answer text and the target grammar and the target style, respectively.

[0067] The second matching degree between the historical reply text and the historical prompt text may refer to the similarity between the historical reply text and the historical prompt text. The correlation between the historical reply text and the historical prompt text may indicate whether the historical reply text is a reply text to the historical prompt text.

[0068] Among them, the setting of multiple question and answer scoring indicators can determine the scoring of the third historical question and answer sample on various indicators, and then determine whether the third historical question and answer sample is an incorrect question and answer sample, thereby further improving the accuracy of the incorrect question and answer sample determined.

[0069] In the embodiments of the present disclosure, in order to achieve flexible setting of question and answer scoring indicators under various field categories and improve the flexibility of question and answer scoring indicators under various field categories, the electronic device can also perform the following processes: displaying the indicator configuration interface; detecting the question and answer scoring indicator setting operation for the field category in the indicator configuration interface; and setting and processing the question and answer scoring indicators under the field category according to the setting operation.

[0070] Among them, in the indicator configuration interface, the object can add, delete, modify, etc. the question and answer scoring indicators under the field category.

[0071] Step 303: determine the first score value of the third historical question and answer sample on each question and answer scoring indicator.

[0072] In the disclosed embodiment, the process of the electronic device executing step 303 may, for example, be to input the third historical question and answer sample and each question and answer scoring indicator into the evaluation model, obtain the indicator content on each question and answer scoring indicator output by the evaluation model; and determine the first scoring value based on the indicator content.

[0073] Among them, the question-and-answer scoring indicators include at least one of the following: the second matching degree between the historical reply text and the historical prompt text; the relevance between the historical reply text and the historical prompt text; the degree of logical consistency between each word in the historical reply text; the third matching degree between the historical reply text and the target grammar and target style respectively.

[0074] For the second matching degree between the historical reply text and the historical prompt text, the indicator content may be a second matching degree value; and the first scoring value may be a scoring value determined according to the second matching degree value. The second matching degree value and the first scoring value may be positively correlated, and the larger the second matching degree value, the larger the first scoring value.

[0075] Among them, for the correlation between the historical reply text and the historical prompt text, the indicator content may be a correlation value. Among them, the first scoring value may be determined based on the correlation value. The correlation value and the first scoring value may be positively correlated. The larger the correlation value, the larger the first scoring value. Among them, for the logical consistency between each word in the historical reply text, the indicator content may be a degree value. The first scoring value may be determined based on the degree value. The degree value and the first scoring value may be positively correlated. The larger the degree value, the larger the first scoring value.

[0076] Wherein, for the third matching degree between the historical reply text and the target grammar and the target style respectively, the indicator content may be two third matching degree values. Wherein, the first scoring value may be determined based on the two third matching degree values.

[0077] Among them, in combination with the evaluation model, the first score value of the third historical question and answer sample on each question and answer scoring indicator is determined; based on the accuracy of the evaluation model, the accuracy of the determined first score value can be improved.

[0078] Step 304: Determine a second scoring value for the third historical question and answer sample based on the first scoring value and the weight.

[0079] In the disclosed embodiment, the electronic device may perform weighted summation processing on the first score values ​​of the third historical question and answer sample on each question and answer score indicator according to the weights of each question and answer score indicator to obtain the second score value of the third historical question and answer sample.

[0080] Step 305: When the second score value is less than the score threshold, the third historical question and answer sample is determined as an erroneous question and answer sample.

[0081] Step 306, obtaining the first question-answering model to be trained.

[0082] Step 307: training the first question-answering model according to the question prompt text and the wrong answer text to obtain a trained second question-answering model.

[0083] It should be noted that the details of step 306 to step 307 can be found in Figure 2 Steps 203 to 204 in the illustrated embodiment will not be described in detail herein.

[0084] The training method of the question and answer model of the embodiment of the present disclosure is as follows: obtaining a third historical question and answer sample of the first question and answer model; determining a question and answer scoring index and a weight of the question and answer scoring index; determining a first scoring value of the third historical question and answer sample on each question and answer scoring index; determining a second scoring value of the third historical question and answer sample based on the first scoring value and the weight; determining the third historical question and answer sample as an incorrect question and answer sample when the second scoring value is less than a scoring threshold; obtaining the first question and answer model to be trained; training the first question and answer model based on a question prompt text and an incorrect answer text to obtain a trained second question and answer model; wherein, each third historical question and answer sample generated during the use of the first question and answer model is scored on the question and answer scoring index, thereby determining whether the third historical question and answer sample is an incorrect question and answer sample, and then training the first question and answer model, which can further improve the accuracy of the trained second question and answer model.

[0085] Figure 4 It is a schematic diagram according to the fourth embodiment of the present disclosure. It should be noted that the question-answering method of the embodiment of the present disclosure can be applied to a question-answering device, which can be configured in an electronic device so that the electronic device can perform a question-answering function.

[0086] Among them, the electronic device can be any device with computing capabilities, such as a personal computer (PC), a mobile terminal, a server, etc. The mobile terminal can be, for example, a vehicle-mounted device, a mobile phone, a tablet computer, a personal digital assistant, a wearable device, a smart speaker, a server, a server cluster, and other hardware devices with various operating systems, touch screens and / or display screens.

[0087] The question-answering device may also be software in an electronic device, such as question-answering software, etc. In the following embodiments, the execution subject is an electronic device as an example for description.

[0088] like Figure 4 As shown, the question-answering method may include the following steps:

[0089] Step 401, obtaining the prompt text of the current question.

[0090] In the disclosed embodiment, the current question prompt text may include question text and prompt text. The prompt text is used to guide the model to generate a specific type of answer or perform a specific task. The question text indicates the question that needs to be answered.

[0091] Step 402, obtaining a question-answering model; the question-answering model is based on Figures 1 to 3 The training method of the question-answering model in any embodiment is determined.

[0092] In the disclosed embodiment, it is assumed that the electronic device that executes the training method of the question-answering model is a first electronic device, and the electronic device that executes the question-answering method is a second electronic device. In one example, the first electronic device and the second electronic device may be the same electronic device. That is, the question-answering model is trained by the same electronic device, and then the question-answering process is performed in combination with the trained question-answering model. The training process of the question-answering model and the question-answering process may be performed separately. For example, the electronic device may first train the question-answering model offline, and then perform online processing, and use the trained question-answering model for question-answering processing.

[0093] In another example, the first electronic device and the second electronic device may be different electronic devices. That is, after the first electronic device determines the trained question-answering model, it may provide the trained question-answering model to the second electronic device; the second electronic device performs question-answering processing based on the trained question-answering model. During the process of the second electronic device performing question-answering processing based on the trained question-answering model, the first electronic device may retrain the question-answering model in combination with the generated erroneous question-answering samples, and synchronize the new question-answering model to the second electronic device.

[0094] Step 403: input the current question prompt text into the question-answering model, and obtain the answer text output by the question-answering model.

[0095] Step 404, determining the reply text as the reply text corresponding to the current question prompt text.

[0096] The question-answering method of the embodiment of the present disclosure obtains the current question prompt text; obtains the question-answering model; the question-answering model is based on Figures 1 to 3 The training method of the question-answering model in any embodiment is determined; the current question prompt text is input into the question-answering model to obtain the reply text output by the question-answering model; the reply text is determined as the reply text corresponding to the current question prompt text; wherein the question-answering model is trained using erroneous question-answering samples, so that the trained question-answering model can avoid generating potential erroneous replies, thereby improving the accuracy of the determined reply text.

[0097] The following examples are used to illustrate this. Figure 5 The figure below is a schematic diagram of determining incorrect question and answer samples. Figure 5In the step 501, the following steps may be specifically included. Step 501, obtain the test set badcase and the business feedback badcase (the collected incorrect question and answer samples). Step 502, determine the badcase embedding (the semantic representation vector of the incorrect question and answer sample) in combination with the semantic vector model. Step 503, obtain the online query (i.e., the question and answer sample generated during the use of the first question and answer model), and determine the query embedding (the semantic representation vector of the question and answer sample) in combination with the semantic vector model. Step 504, perform semantic matching processing on the badcase embedding and query embedding to determine whether the online query is similar to the badcase (i.e., determine whether the question and answer sample is an incorrect question and answer sample). Step 505, perform self-correction in combination with similar badcases (i.e., train the first question and answer model in combination with the incorrect question and answer samples).

[0098] In order to implement the above embodiment, the present disclosure also provides a training device for a question-answering model. Figure 6 As shown, Figure 6 6 is a schematic diagram according to the fifth embodiment of the present disclosure. The training device 60 of the question-answering model may include: a first acquisition module 601 , a second acquisition module 602 and a training processing module 603 .

[0099] Among them, the first acquisition module 601 is used to obtain incorrect question and answer samples; the incorrect question and answer samples include question prompt text and incorrect answer text corresponding to the question prompt text; the second acquisition module 602 is used to obtain the first question and answer model to be trained; the training processing module 603 is used to train the first question and answer model according to the question prompt text and the incorrect answer text to obtain the trained second question and answer model.

[0100] As a possible implementation method of an embodiment of the present disclosure, the first acquisition module 601 is specifically used to obtain a first historical question and answer sample of the first question and answer model, and feedback information for the first historical question and answer sample; when the feedback information indicates that the answer is wrong, the first historical question and answer sample is determined as the wrong question and answer sample.

[0101] As a possible implementation method of the embodiment of the present disclosure, the device also includes: a third acquisition module, a first determination module and a second determination module; the third acquisition module is used to obtain a second historical question and answer sample of the first question and answer model without feedback information; the first determination module is used to determine a first matching degree between the second historical question and answer sample and each of the erroneous question and answer samples; the second determination module is used to determine the second historical question and answer sample as an erroneous question and answer sample when there is an erroneous question and answer sample whose first matching degree is greater than or equal to a matching degree threshold.

[0102] As a possible implementation method of an embodiment of the present disclosure, the first determination module is specifically used to determine the first semantic representation vector of the second historical question and answer sample, and the second semantic representation vector of each of the erroneous question and answer samples; determine the similarity between the first semantic representation vector and each of the second semantic representation vectors; and determine the first matching degree based on each of the similarities.

[0103] As a possible implementation method of an embodiment of the present disclosure, the first acquisition module 601 includes: an acquisition unit, a first determination unit, a second determination unit, a third determination unit and a fourth determination unit; the acquisition unit is used to acquire the third historical question and answer sample of the first question and answer model; the first determination unit is used to determine the question and answer scoring index and the weight of the question and answer scoring index; the second determination unit is used to determine the first scoring value of the third historical question and answer sample on each of the question and answer scoring indicators; the third determination unit is used to determine the second scoring value of the third historical question and answer sample based on the first scoring value and the weight; the fourth determination unit is used to determine the third historical question and answer sample as the incorrect question and answer sample when the second scoring value is less than the scoring threshold.

[0104] As a possible implementation manner of an embodiment of the present disclosure, the first determination unit is specifically used to determine the domain category of the third historical question and answer sample; and obtain the question and answer scoring index and the weight corresponding to the domain category from an index database.

[0105] As a possible implementation method of the embodiment of the present disclosure, the second determination unit is specifically used to input the third historical question and answer sample and each of the question and answer scoring indicators into the evaluation model, obtain the indicator content of each of the question and answer scoring indicators output by the evaluation model; and determine the first scoring value based on the indicator content.

[0106] As a possible implementation method of an embodiment of the present disclosure, the third historical question and answer sample includes a historical question prompt text and a historical answer text; the historical question prompt text includes a historical question text and a historical prompt text; the question and answer scoring index includes at least one of the following: a second matching degree between the historical answer text and the historical prompt text; a correlation between the historical answer text and the historical prompt text; a degree of logical consistency between each word in the historical answer text; and a third matching degree between the historical answer text and the target grammar and target style respectively.

[0107] As a possible implementation method of the embodiment of the present disclosure, the device also includes: a display module, a detection module and a setting processing module; the display module is used to display the indicator configuration interface; the detection module is used to detect the question and answer scoring indicator setting operation for the field category in the indicator configuration interface; the setting processing module is used to set and process the question and answer scoring indicator under the field category according to the setting operation.

[0108] As a possible implementation method of an embodiment of the present disclosure, the incorrect question and answer sample also includes at least one of the following for determining the incorrect reply text: text search results, reasons for incorrect question and answer; wherein, the text search results are determined based on the question prompt text.

[0109] As a possible implementation method of an embodiment of the present disclosure, the training processing module 603 is specifically used to input the question prompt text into the first question and answer model to obtain the predicted answer text output by the first question and answer model; determine the fourth matching degree between the incorrect answer text and the predicted answer text; determine the first numerical value of the loss function of the first question and answer model based on the fourth matching degree; and perform parameter adjustment processing on the first question and answer model based on the first numerical value to obtain the second question and answer model.

[0110] As a possible implementation method of an embodiment of the present disclosure, the training processing module 603 is specifically used to determine the correct answer text corresponding to the question prompt text; input the question prompt text into the first question and answer model to obtain the predicted answer text output by the first question and answer model; determine the fifth matching degree between the predicted answer text and the correct answer text, and the sixth matching degree between the predicted answer text and the incorrect answer text; determine the second value of the loss function of the first question and answer model based on the inverse of the fifth matching degree and the sixth matching degree; and perform parameter adjustment processing on the first question and answer model based on the second value to obtain the second question and answer model.

[0111] The training device of the question-answering model of the embodiment of the present disclosure obtains incorrect question-answering samples; the incorrect question-answering samples include question prompt text and incorrect answer text corresponding to the question prompt text; obtains a first question-answering model to be trained; trains the first question-answering model according to the question prompt text and the incorrect answer text to obtain a trained second question-answering model; wherein the second question-answering model is trained using the incorrect question-answering samples, so that the trained second question-answering model can avoid generating potential incorrect answers, thereby improving the question-answering accuracy of the trained second question-answering model.

[0112] In order to implement the above embodiment, the present disclosure also provides a question-answering device. Figure 7 As shown, Figure 7 The question-answering device 70 may include: a first acquisition module 701 , a second acquisition module 702 , a third acquisition module 703 and a determination module 704 .

[0113] The first acquisition module 701 is used to acquire the prompt text of the current question;

[0114] The second acquisition module 702 is used to acquire a question-answering model; the question-answering model is based on Figures 1 to 3 The training method of the question-answering model described in any embodiment is determined; the third acquisition module 703 is used to input the current question prompt text into the question-answering model to obtain the reply text output by the question-answering model; the determination module 704 is used to determine the reply text as the reply text corresponding to the current question prompt text.

[0115] The question-answering device of the embodiment of the present disclosure obtains the current question prompt text; obtains the question-answering model; the question-answering model is based on Figures 1 to 3 The training method of the question-answering model in any embodiment is determined; the current question prompt text is input into the question-answering model to obtain the reply text output by the question-answering model; the reply text is determined as the reply text corresponding to the current question prompt text; wherein the question-answering model is trained using erroneous question-answering samples, so that the trained question-answering model can avoid generating potential erroneous replies, thereby improving the accuracy of the determined reply text.

[0116] In the technical solution of the present disclosure, the collection, storage, use, processing, transmission, provision and disclosure of user personal information are all carried out with the user's consent, comply with the relevant laws and regulations, and do not violate public order and good morals.

[0117] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium and a computer program product.

[0118] Figure 8A schematic block diagram of an example electronic device 800 that can be used to implement an embodiment of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or required herein.

[0119] like Figure 8 As shown, the device 800 includes a computing unit 801, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 802 or a computer program loaded from a storage unit 808 into a random access memory (RAM) 803. In the RAM 803, various programs and data required for the operation of the device 800 can also be stored. The computing unit 801, the ROM 802, and the RAM 803 are connected to each other via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0120] A number of components in the device 800 are connected to the I / O interface 805, including: an input unit 806, such as a keyboard, a mouse, etc.; an output unit 807, such as various types of displays, speakers, etc.; a storage unit 808, such as a disk, an optical disk, etc.; and a communication unit 809, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 809 allows the device 800 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0121] The computing unit 801 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 801 performs the various methods and processes described above, such as the training method or the question-answering method of the question-answering model. For example, in some embodiments, the training method or the question-answering method of the question-answering model may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit 808. In some embodiments, part or all of the computer program may be loaded and / or installed on the device 800 via the ROM 802 and / or the communication unit 809. When the computer program is loaded into the RAM 803 and executed by the computing unit 801, one or more steps of the training method or the question-answering method of the question-answering model described above may be executed. Alternatively, in other embodiments, the computing unit 801 may be configured to execute a training method or a question-answering method for a question-answering model in any other appropriate manner (e.g., by means of firmware).

[0122] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include: being implemented in one or more computer programs, which may be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general programmable processor, which may receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0123] The program code for implementing the method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0124] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0125] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0126] The systems and techniques described herein may be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), and the Internet.

[0127] A computer system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The relationship of client and server is generated by computer programs running on respective computers and having a client-server relationship with each other. The server may be a cloud server, a server of a distributed system, or a server combined with a blockchain.

[0128] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved, and this document does not limit this.

[0129] The above specific implementations do not constitute a limitation on the protection scope of the present disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A method for training a question-answering model, the method comprising: Obtaining an incorrect question and answer sample; the incorrect question and answer sample includes a question prompt text and an incorrect answer text corresponding to the question prompt text; Obtain a first question-answering model to be trained; The first question-answering model is trained based on the question prompt text and the incorrect answer text to obtain a trained second question-answering model.

2. The method according to claim 1, wherein: The obtaining of the error question and answer sample includes: Obtaining a first historical question and answer sample of the first question and answer model, and feedback information for the first historical question and answer sample; In a case where the feedback information indicates that the answer is wrong, the first historical question and answer sample is determined as the wrong question and answer sample.

3. The method according to claim 1 or 2, wherein: The method further comprises: Obtaining a second historical question-and-answer sample of the first question-and-answer model without any feedback information; Determining a first matching degree between the second historical question and answer sample and each of the erroneous question and answer samples; In the case where there is an erroneous question and answer sample whose first matching degree is greater than or equal to a matching degree threshold, the second historical question and answer sample is determined as an erroneous question and answer sample.

4. The method according to claim 3, wherein: The determining of a first matching degree between the second historical question and answer sample and each of the erroneous question and answer samples includes: Determining a first semantic representation vector of the second historical question and answer sample and a second semantic representation vector of each of the erroneous question and answer samples; Determining the similarity between the first semantic representation vector and each of the second semantic representation vectors; The first matching degree is determined according to each of the similarities.

5. The method according to claim 1, wherein: The obtaining of the error question and answer sample includes: Obtaining a third historical question and answer sample of the first question and answer model; Determining question-answering scoring indicators and weights of the question-answering scoring indicators; Determine a first scoring value of the third historical question and answer sample on each of the question and answer scoring indicators; Determining a second scoring value for the third historical question and answer sample according to the first scoring value and the weight; When the second score value is less than a score threshold, the third historical question and answer sample is determined as the incorrect question and answer sample.

6. The method according to claim 5, wherein: The determining of the question-and-answer scoring index and the weight of the question-and-answer scoring index includes: determining the domain category of the third history question and answer sample; The question-and-answer scoring index and the weight corresponding to the field category are obtained from an index database.

7. The method according to claim 5, wherein: The determining of the first score value of the third historical question and answer sample on each of the question and answer scoring indicators includes: Input the third historical question and answer sample and each of the question and answer scoring indicators into an evaluation model, and obtain the indicator content on each of the question and answer scoring indicators output by the evaluation model; The first scoring value is determined according to the indicator content.

8. The method according to any one of claims 5 to 7, wherein: The third historical question and answer sample includes a historical question prompt text and a historical answer text; the historical question prompt text includes a historical question text and a historical prompt text; The question-and-answer scoring indicators include at least one of the following: a second matching degree between the historical reply text and the historical prompt text; a correlation degree between the historical reply text and the historical prompt text; a degree of logical consistency between each word in the historical reply text; and a third matching degree between the historical reply text and the target grammar and target style respectively.

9. The method according to claim 5 or 6, wherein: The method further comprises: Display indicator configuration interface; Detecting the question and answer scoring indicator setting operation for the field category in the indicator configuration interface; According to the setting operation, the question and answer scoring indicators under the field category are set and processed.

10. The method according to claim 1 or 5, wherein: The incorrect question and answer sample also includes at least one of the following for determining the incorrect reply text: text search results, incorrect question and answer reasons; The text search result is determined based on the question prompt text.

11. The method according to claim 1, wherein: The training process of the first question-answering model according to the question prompt text and the wrong answer text to obtain a trained second question-answering model includes: Inputting the question prompt text into the first question-answering model to obtain a predicted answer text output by the first question-answering model; determining a fourth degree of match between the erroneous reply text and the predicted reply text; Determining a first value of a loss function of the first question-answering model according to the fourth matching degree; The first question-answering model is subjected to parameter adjustment processing according to the first numerical value to obtain the second question-answering model.

12. The method according to claim 1, wherein: The training process of the first question-answering model according to the question prompt text and the wrong answer text to obtain a trained second question-answering model includes: Determine the correct answer text corresponding to the question prompt text; Inputting the question prompt text into the first question-answering model to obtain a predicted answer text output by the first question-answering model; Determining a fifth degree of match between the predicted reply text and the correct reply text, and a sixth degree of match between the predicted reply text and the incorrect reply text; Determining a second value of the loss function of the first question-answering model according to the inverse of the fifth matching degree and the sixth matching degree; The first question-answering model is subjected to parameter adjustment processing according to the second numerical value to obtain the second question-answering model.

13. A question-answering method, the method comprising: Get the prompt text of the current question; Get the question-answering model; The question-answering model is determined based on the training method of the question-answering model according to any one of claims 1 to 12; Input the current question prompt text into the question-answering model, and obtain the answer text output by the question-answering model; The reply text is determined as the reply text corresponding to the current question prompt text.

14. A training device for a question-answering model, the device comprising: The first acquisition module is used to acquire an incorrect question and answer sample; the incorrect question and answer sample includes a question prompt text and an incorrect answer text corresponding to the question prompt text; A second acquisition module, used to acquire a first question-answering model to be trained; The training processing module is used to train the first question-answering model according to the question prompt text and the wrong answer text to obtain a trained second question-answering model.

15. The device according to claim 14, wherein: The first acquisition module is specifically used to: Obtaining a first historical question and answer sample of the first question and answer model, and feedback information for the first historical question and answer sample; In a case where the feedback information indicates that the answer is wrong, the first historical question and answer sample is determined as the wrong question and answer sample.

16. The device according to claim 14 or 15, wherein: The device further includes: a third acquisition module, a first determination module, and a second determination module; The third acquisition module is used to acquire a second historical question and answer sample of the first question and answer model without feedback information; The first determination module is used to determine a first matching degree between the second historical question and answer sample and each of the erroneous question and answer samples; The second determination module is configured to determine the second historical question and answer sample as an erroneous question and answer sample when there is an erroneous question and answer sample whose first matching degree is greater than or equal to a matching degree threshold.

17. The device according to claim 16, wherein: The first determination module is specifically configured to: Determining a first semantic representation vector of the second historical question and answer sample and a second semantic representation vector of each of the erroneous question and answer samples; Determining the similarity between the first semantic representation vector and each of the second semantic representation vectors; The first matching degree is determined according to each of the similarities.

18. The device according to claim 14, wherein: The first acquisition module includes: an acquisition unit, a first determination unit, a second determination unit, a third determination unit and a fourth determination unit; The acquisition unit is used to acquire a third historical question and answer sample of the first question and answer model; The first determining unit is used to determine the question and answer scoring index and the weight of the question and answer scoring index; The second determining unit is used to determine a first scoring value of the third historical question and answer sample on each of the question and answer scoring indicators; The third determining unit is used to determine a second scoring value of the third historical question and answer sample according to the first scoring value and the weight; The fourth determining unit is used to determine the third historical question and answer sample as the incorrect question and answer sample when the second score value is less than a score threshold.

19. The device according to claim 18, wherein The first determining unit is specifically configured to: determining the domain category of the third history question and answer sample; The question-and-answer scoring index and the weight corresponding to the field category are obtained from an index database.

20. The device according to claim 18, wherein The second determining unit is specifically configured to: Input the third historical question and answer sample and each of the question and answer scoring indicators into an evaluation model, and obtain the indicator content on each of the question and answer scoring indicators output by the evaluation model; The first scoring value is determined according to the indicator content.

21. The device according to any one of claims 18 to 20, wherein: The third historical question and answer sample includes a historical question prompt text and a historical answer text; the historical question prompt text includes a historical question text and a historical prompt text; The question-and-answer scoring indicators include at least one of the following: a second matching degree between the historical reply text and the historical prompt text; a correlation degree between the historical reply text and the historical prompt text; a degree of logical consistency between each word in the historical reply text; and a third matching degree between the historical reply text and the target grammar and target style respectively.

22. The device according to claim 18 or 19, wherein: The device also includes: a display module, a detection module and a setting processing module; The display module is used to display the indicator configuration interface; The detection module is used to detect the question and answer scoring indicator setting operation for the field category in the indicator configuration interface; The setting processing module is used to set and process the question and answer scoring indicators under the field category according to the setting operation.

23. The device according to claim 14 or 18, wherein: The incorrect question and answer sample also includes at least one of the following for determining the incorrect reply text: text search results, incorrect question and answer reasons; The text search result is determined based on the question prompt text.

24. The device according to claim 14, wherein: The training processing module is specifically used to: Inputting the question prompt text into the first question-answering model to obtain a predicted answer text output by the first question-answering model; determining a fourth degree of match between the erroneous reply text and the predicted reply text; Determining a first value of a loss function of the first question-answering model according to the fourth matching degree; The first question-answering model is subjected to parameter adjustment processing according to the first numerical value to obtain the second question-answering model.

25. The device according to claim 14, wherein: The training processing module is specifically used to: Determine the correct answer text corresponding to the question prompt text; Inputting the question prompt text into the first question-answering model to obtain a predicted answer text output by the first question-answering model; Determining a fifth degree of match between the predicted reply text and the correct reply text, and a sixth degree of match between the predicted reply text and the incorrect reply text; Determining a second value of the loss function of the first question-answering model according to the inverse of the fifth matching degree and the sixth matching degree; The first question-answering model is subjected to parameter adjustment processing according to the second numerical value to obtain the second question-answering model.

26. A question-answering device, comprising: The first acquisition module is used to obtain the current question prompt text; The second acquisition module is used to acquire the question-answering model; The question-answering model is determined based on the training method of the question-answering model according to any one of claims 1 to 12; A third acquisition module is used to input the current question prompt text into the question-answering model to obtain the answer text output by the question-answering model; A determination module is used to determine the reply text as the reply text corresponding to the current question prompt text.

27. An electronic device comprising: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 12; Alternatively, perform the method of claim 13.

28. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1 to 12; or, to execute the method according to claim 13.

29. A computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements the method according to any one of claims 1 to 12; or implements the method according to claim 13.

Images