Intelligent question and answer method and device and storage medium

Abstract

The invention discloses an intelligent question and answer method and device and a storage medium. The method comprises the steps that after a first question currently proposed by a user is obtained, entity objects, belonging to a preset entity category, in the question are determined; then candidate knowledge of the first question is selected from the multiple pieces of domain knowledge according to the text similarity between the entity object and each piece of pre-stored domain knowledge; and then performing feature recognition on the entity object and the candidate knowledge through a trained question and answer model to obtain a question answer output by the model for the first question. According to the process, the user question and the candidate knowledge related to the user question are input into the model at the same time, so that the extraction amount of the model for question features can be effectively improved; and the question answering precision of the model is improved.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing, and in particular to an intelligent question-answering method, device and storage medium. Background Art

[0002] The intelligent question-answering system uses a large amount of data to train and learn semantic models, so that it can identify questions and give corresponding answers based on the content input by users. The role of the intelligent question-answering system is to provide people with accurate, fast and efficient services. For example, it is widely used in customer service, voice recognition, robot management and other fields.

[0003] Intelligent question-answering systems currently perform poorly in dealing with domain-related questions. The reason is that domain knowledge (i.e., the knowledge reserve of intelligent question-answering systems) is different from the training samples in the form of conventional question-answering. Domain knowledge is mainly structured data in relational databases and cannot be used directly as training samples. It is necessary to convert structured data into natural language. In addition, domain knowledge (such as medical, economic, and engineering) is difficult to obtain, and the number of samples built based on domain knowledge is usually insufficient to support the model to have high answer accuracy and generalization ability. Summary of the invention

[0004] The embodiments of the present application provide an intelligent question-answering method, device, and storage medium for improving the answer accuracy of an intelligent question-answering model.

[0005] In a first aspect, an embodiment of the present application provides an intelligent question-answering method, comprising:

[0006] Acquire a first question currently raised by a user, and determine an entity object in the first question that belongs to a preset entity category; wherein the preset entity category is determined according to a specified field;

[0007] Determine the text similarity between the entity object and each pre-stored domain knowledge, and select candidate knowledge for the first question from the pre-stored multiple domain knowledge according to the text similarity; wherein any domain knowledge is extracted from the structured data of the specified domain;

[0008] The entity object and the candidate knowledge are feature-recognized by a trained question-answering model to obtain an answer to the first question, and the answer is displayed to the user; wherein the question-answering model is obtained by training a basic question-answering model based on the domain knowledge; the basic question-answering model is any trained intelligent question-answering model for answering non-domain questions.

[0009] In some possible embodiments, before determining the text similarity between the entity object and each pre-stored domain knowledge, the method further includes:

[0010] Determining the number of categories of the preset entity categories included in the first question;

[0011] If the number of categories is greater than a first number threshold and less than a second number threshold, then obtaining all second questions raised by the user in the first historical period;

[0012] The first question is updated according to the entity objects of each preset entity category in the first question and the second question.

[0013] In some possible embodiments, updating the first question according to the entity objects of each preset entity category in the first question and the second question includes:

[0014] For any second question, if the preset entity categories included in the second question include all target categories, and the entity objects of the second question under at least one target category are different from those of the first question, the second question is used as the third question; wherein any target category is a preset entity category included in the first question;

[0015] Selecting a fourth question from the third questions, the fourth question being asked at a time adjacent to the first question;

[0016] For any target category in the fourth question, the entity object of the fourth question under the target category is replaced with the entity object of the first question under the target category; and the fourth question after the entity object is replaced is used as the first question after the question is updated.

[0017] In some possible embodiments, the method further includes:

[0018] If the fourth question does not exist within the first historical period, a prompt message indicating that the current question cannot be answered is output to the user.

[0019] In some possible embodiments, before determining the text similarity between the entity object and each pre-stored domain knowledge, the method further includes:

[0020] If the number of entity objects under any preset entity category in the first question is greater than a third quantity threshold, splitting the first question into multiple sub-questions;

[0021] Among them, any sub-question includes all preset entity categories in the first question, and each preset entity category in the sub-question includes an entity object; any two sub-questions have different entity objects under at least one preset entity category.

[0022] In some possible embodiments, determining the text similarity between the entity object and each pre-stored domain knowledge, and selecting candidate knowledge for the first question from the pre-stored domain knowledge according to the text similarity includes:

[0023] For any sub-question, extract all entity objects under the preset entity categories from the sub-question, and determine the text similarity between the entity object and each piece of domain knowledge based on the text similarity algorithm;

[0024] Selecting knowledge to be processed whose text similarity is greater than a first similarity threshold from the plurality of domain knowledge;

[0025] Determine the semantic similarity between the entity object and each piece of knowledge to be processed, and use the knowledge to be processed whose semantic similarity is greater than a second similarity threshold as candidate knowledge for the sub-problem;

[0026] The candidate knowledge of all sub-questions is used as the candidate knowledge of the first question.

[0027] In some possible embodiments, the step of performing feature recognition on the entity object and the candidate knowledge by using a trained question-answering model to obtain an answer to the first question includes:

[0028] Obtaining historical questions raised by the user in the second historical period and answers to each historical question output by the question-answering model;

[0029] Constructing a first text consisting of the first question, the historical question, and answers to each historical question, and constructing a second text consisting of the first question and each candidate knowledge;

[0030] By inputting the first text and the second text into the question-answering model, the question model performs feature extraction on the first text and the second text respectively to obtain first text features of the first text and second text features of the second text, and the question-answering model performs feature recognition on the feature splicing result of the first text features and the second text features to obtain a question answer to the first question.

[0031] In some possible embodiments, the pre-stored multiple pieces of domain knowledge are obtained in the following manner:

[0032] In response to the knowledge reserve indication, obtaining structured data of the specified field; wherein the structured data at least includes form data and a knowledge graph;

[0033] For any form data, if each list of the form data contains a preset column name, the data content under the preset column name is extracted, and the domain knowledge is constructed according to the data content extracted under each preset column name in the form data; wherein the preset list is determined according to the preset entity category;

[0034] For any entity object in any knowledge graph, determine all directly connected entity objects in the knowledge graph that are directly connected to the entity object;

[0035] A target entity object belonging to the preset entity object category is selected from the entity object and the entity objects directly connected to the entity object, and the domain knowledge is constructed according to the entity object and the target entity object.

[0036] In a second aspect, an embodiment of the present application further provides an intelligent question-answering device, the device comprising:

[0037] A question acquisition unit is configured to acquire a first question currently raised by a user and determine an entity object in the first question that belongs to a preset entity category; wherein the preset entity category is determined according to a specified field;

[0038] A candidate knowledge unit is configured to determine the text similarity between the entity object and each pre-stored domain knowledge, and select candidate knowledge for the first question from the pre-stored multiple domain knowledge according to the text similarity; wherein any domain knowledge is extracted from the structured data of the specified domain;

[0039] The answer display unit is configured to perform feature recognition on the entity object and the candidate knowledge through a trained question-answering model, obtain a question answer to the first question, and display the question answer to the user; wherein the question-answering model is obtained by training a basic question-answering model based on the domain knowledge; the basic question-answering model is any trained intelligent question-answering model for answering non-domain questions.

[0040] In a third aspect, an embodiment of the present application further provides an electronic device, including a data transmission unit and a processor:

[0041] The data transmission unit is configured to: obtain a first question currently rejected by the user;

[0042] The processor is configured to: determine an entity object in the first question that belongs to a preset entity category; wherein the preset entity category is determined according to a specified field;

[0043] Determine the text similarity between the entity object and each pre-stored domain knowledge, and select candidate knowledge for the first question from the pre-stored multiple domain knowledge according to the text similarity; wherein any domain knowledge is extracted from the structured data of the specified domain;

[0044] The entity object and the candidate knowledge are feature-recognized by a trained question-answering model to obtain an answer to the first question, and the answer is displayed to the user; wherein the question-answering model is obtained by training a basic question-answering model based on the domain knowledge; the basic question-answering model is any trained intelligent question-answering model for answering non-domain questions.

[0045] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a computer, the computer executes any one of the methods described in the first aspect.

[0046] In the embodiment of the present application, after obtaining the first question currently raised by the user, the entity objects belonging to the preset entity category in the question are determined. Then, candidate knowledge for the first question is selected from multiple domain knowledge based on the text similarity between the entity object and each pre-stored domain knowledge. Next, the trained question-answering model is used to perform feature recognition on the entity object and the candidate knowledge to obtain the answer to the question output by the model for the first question. The above process can effectively improve the model's extraction of question features by simultaneously inputting the user question and the candidate knowledge related to the user question into the model, thereby improving the accuracy of the model's answer to the question.

[0047] Other features and advantages of the present application will be described in the following description, and partly become apparent from the description, or be understood by practicing the present application. The purpose and other advantages of the present application can be realized and obtained by the structures specifically pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings introduced below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0049] Figure 1 A flowchart of question-answering model training provided in an embodiment of the present application;

[0050] Figure 2A schematic diagram of basic question-answering model training provided in an embodiment of the present application;

[0051] Figure 3 A schematic diagram of form data in structured data provided in an embodiment of the present application;

[0052] Figure 4 A schematic diagram of the knowledge graph in the structured data provided in the embodiment of the present application;

[0053] Figure 5 A schematic diagram of preset listings provided for an embodiment of the present application;

[0054] Figure 6 A schematic diagram of extracting domain knowledge from form data provided in an embodiment of the present application;

[0055] Figure 7 A schematic diagram of extracting domain knowledge from a knowledge graph provided in an embodiment of the present application;

[0056] Figure 8 An overall flow chart of an intelligent question-answering method provided in an embodiment of the present application;

[0057] Fig. 9 A schematic diagram of a process for updating the first question provided in an embodiment of the present application;

[0058] Fig.10 A schematic diagram of the third problem provided in the embodiment of the present application;

[0059] Fig.11 A schematic diagram of the fourth problem provided in an embodiment of the present application;

[0060] Fig.12 A schematic diagram of updating the first question according to the fourth question provided in an embodiment of the present application;

[0061] Fig.13 A schematic diagram of prompt information provided in an embodiment of the present application;

[0062] Fig.14 A schematic diagram of a first problem that can be divided into sub-problems provided in an embodiment of the present application;

[0063] Fig.15 A schematic diagram of historical issues provided for embodiments of the present application;

[0064] Fig.16 A schematic diagram of the model internal feature processing flow provided in the embodiment of the present application;

[0065] Fig.17 A schematic diagram of the structure of an intelligent question-answering device provided in an embodiment of the present application;

[0066] Fig.18A structural block diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical scheme and advantages of the present application clearer, the technical scheme in the embodiment of the present application will be clearly and completely described below in conjunction with the drawings in the embodiment of the present application. Obviously, the described embodiment is only a part of the embodiment of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application. In the absence of conflict, the embodiments in the present application and the features in the embodiments can be combined with each other arbitrarily. In addition, although the logical order is shown in the flow chart, in some cases, the steps shown or described can be performed in an order different from that here.

[0068] The terms "first" and "second" in the specification and claims of the present application and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the term "comprising" and any of their variations are intended to cover non-exclusive protection. For example, a process, method, system, product or device comprising a series of steps or units is not limited to the listed steps or units, but optionally also includes steps or units that are not listed, or optionally also includes other steps or units inherent to these processes, methods, products or devices. "Multiple" in the present application can mean at least two, for example, two, three or more, and the embodiments of the present application are not limited.

[0069] The current intelligent question-answering system performs poorly in dealing with domain-related questions. The reason is that domain knowledge is mainly represented by structured data in relational databases, which cannot be used directly as training samples. It is necessary to convert structured data into natural language. In addition, domain knowledge (such as medical, economic, and engineering) is difficult to obtain. The number of samples built based on domain knowledge is usually not enough to support the model with high answer accuracy and generalization ability.

[0070] Specifically, domain knowledge usually exists in the form of structured tables, graphs, etc. Taking economic domain knowledge as an example, it often exists in the form of economic analysis tables or knowledge graphs compiled by relevant units or platforms. How to use this knowledge in question-answering is a very critical issue.

[0071] There are two common practices in the industry. One is to construct a large number of knowledge question-answer pairs based on domain knowledge, and use the question-answer pair training model to perform semantic retrieval based on the questions raised by users to achieve domain question-answering. However, this method relies too much on the constructed question-answer pairs, and the massive generation of question-answer pairs will inevitably lead to the inability to guarantee their quality. The other is to perform intent recognition and entity extraction on the questions, perform data query and assembly to obtain the answers to the questions. This method has a high accuracy rate under the premise of sufficient training samples. However, it is difficult to obtain domain knowledge, and the training samples cannot cover all user intentions and entity types, so it performs poorly in answering domain-related questions.

[0072] To solve the above problems, the invention of the present application is as follows: after obtaining the first question currently raised by the user, the entity object belonging to the preset entity category in the question is determined. Then, candidate knowledge for the first question is selected from multiple domain knowledge according to the text similarity between the entity object and each pre-stored domain knowledge. Next, the trained question-answering model is used to perform feature recognition on the entity object and the candidate knowledge to obtain the question answer output by the model for the first question, and finally the question answer is displayed to the user.

[0073] The basic question-answering model of the present application is trained with conventional training samples, and is a conventional question-answering model for answering non-domain questions, while the question-answering model is obtained by secondary training the basic question-answering model with pre-stored domain knowledge, so the trained question-answering model has a knowledge reserve for domain questions. However, considering that it is difficult to obtain domain knowledge in practical applications, it is usually impossible to obtain enough data for model generalization. Therefore, in the model use stage, the embodiment of the present application uses text similarity to find candidate knowledge related to the user question (i.e., the first question mentioned above) from the pre-stored domain knowledge, and by simultaneously inputting the user question and the candidate knowledge related to the user question into the model, the model can effectively improve the amount of question feature extraction, thereby improving the accuracy of the model's answer to the question.

[0074] In order to facilitate understanding of the technical solution of the present application, the question-answering model training process in the embodiment of the present application is described in advance, as follows: Figure 1 As shown:

[0075] Step 101: training a basic question-answering model based on conventional intelligent question-answering model training samples;

[0076] During implementation, a 1TB (terabyte)-scale conventional Chinese-English bilingual data set was used as a training sample, covering daily question-and-answer conversations, articles, translations, summaries, etc. The model was trained by using a bidirectional attention mechanism on the model input and a unidirectional attention mechanism on the model output, and using an autoregressive text completion method.

[0077] The basic question-answering model training phase of this application can be specifically described as follows Figure 2 As shown in the figure, by randomly masking part of the content in the training sample and combining it with position encoding, the model generates the masked content. The generated content is compared with the masked content to calculate the loss function, and the model parameters are adjusted based on the loss function obtained in each round until the loss function obtained in any round is less than the loss function threshold, that is, the model iteration is completed.

[0078] Step 102: extracting domain knowledge from structured data of domain classes;

[0079] During implementation, structured data of the specified field is obtained in advance. For the convenience of subsequent explanation, the structured data in the economic field is used as an example. It should be stated that the technical solution of this application is not limited to the economic field or a certain field. In practical applications, the model needs to have the ability to answer questions in certain fields, that is, the model can be trained using structured data in the field.

[0080] As mentioned above, knowledge in the economic field mostly exists in the form of economic analysis forms or knowledge graphs compiled by relevant units or platforms. In natural language processing, objective individuals are abstracted into entities, while forms and knowledge graphs are composed of multiple entities. Figure 3 Take the monthly gross domestic product (GDP) statistics form shown as an example, each column name in the form is an entity. Figure 4 Taking the monthly gross domestic product map as an example, each entity object in the map is an entity. Based on this, the embodiment of the present application pre-sets multiple preset entity categories based on common questions asked by users in the economic field, namely time, place, indicator name, indicator unit and indicator value.

[0081] The column names contained in different types of economic forms are different. For example, the column names in a form for monthly GDP statistics may include month, location, GDP, etc., and the column names in a form for annual payment statistics may include year, location, payment details, etc.

[0082] Based on this, this application is as follows Figure 5 As shown, the column names in various types of forms in the economic field are statistically analyzed in advance, and the column names representing time, such as month, quarter, week, year, etc. are counted as time-type preset columns. The column names representing places, such as a province, a city, a region, etc. are counted as place-type preset columns. GDP, payment names (such as cleaning fees, water and electricity fees, etc.), and payment names (such as wages, taxes, etc.) are counted as indicator name-type preset columns. In this way, Figure 5 The three preset column names shown in respectively correspond to the aforementioned preset entity categories (time, place and indicator name).

[0083] During implementation, for any form data, if each list of the form data contains a preset column name, the data content under the preset column name is extracted, and domain knowledge is constructed based on the data content extracted under each preset column name in the form data.

[0084] With the aforementioned Figure 2 Take the first form data shown in as an example, Figure 6 As shown. The list of items in this form data contains three preset column names: time, location, and indicator name. By extracting the data content under each preset column name, a field knowledge corresponding to this form data is obtained: the GDP of City A in January is 90 billion yuan.

[0085] In the same field (such as the aforementioned economic field), the meanings expressed by forms and knowledge graphs are basically the same, both of which are statistical displays of certain economic data. Therefore, this application adopts the aforementioned similar approach to form data, and through the naming statistics of entity objects in a large number of knowledge graphs, preset entity object categories that represent time, place, indicator name, indicator value, and indicator unit are preset.

[0086] During implementation, for any entity object in any knowledge graph, all directly connected entity objects in the knowledge graph that are directly connected to the entity object are pre-determined; and a target entity object belonging to a preset entity object category is selected from the entity object and the directly connected entity objects, and domain knowledge is constructed based on the entity object and the target entity object.

[0087] With the aforementioned Figure 3 Take the entity object "Area A" in the knowledge graph as an example. Figure 7 As shown. The directly connected entity objects directly connected to the entity object "Area A" in the knowledge graph include "December", "Per capita income", "5700", "Yuan", "GDP", "4 billion", and "100 million yuan". Since its directly connected entity objects all belong to the target entity objects of the preset entity object category, the two field data of the per capita income of Area A in December of 5700 yuan and the GDP of Area A in December of 4 billion yuan can be obtained according to the entity object and the corresponding target entity objects.

[0088] Step 103: construct training samples for two-stage training of the basic question-answering model based on domain knowledge, and perform two-stage training on the basic question-answering model to obtain a trained question-answering model;

[0089] Through the above process, the corresponding domain knowledge can be extracted from the structured data in the economic field, and the extracted domain knowledge can be used as training samples for the second-stage training of the basic question-answering model trained in the above step 101.

[0090] Next, the basic question-answering model is trained in two stages by the training method mentioned in the aforementioned step 101, and after the model converges, the trained question-answering model of the embodiment of the present application is obtained. The question-answering model has a certain amount of economic knowledge reserves.

[0091] After introducing the question-answering model of the embodiment of the present application through the aforementioned steps 101 to 103, the overall process of an intelligent question-answering method provided by the embodiment of the present application is explained below. Figure 8 As shown, including:

[0092] Step 801: obtaining a first question currently raised by a user, and determining an entity object in the first question that belongs to a preset entity category; wherein the preset entity category is determined according to a specified field;

[0093] During implementation, it is predetermined whether the first question currently raised by the user contains entity objects representing the aforementioned preset entity categories (ie, time, place, indicator name, indicator unit and indicator value).

[0094] If the first question does not contain any entity object of the preset entity category, such as "Hello", "How is the weather today", etc., it means that the current first question is a regular question of non-domain category. At this time, the first question can be directly input into the question-answering model trained by the above steps 101 to 103 to obtain the answer to the question.

[0095] If the first question contains at least one entity object of a preset entity category, it means that the first question may be a domain-type question. As mentioned earlier, it is difficult to obtain domain knowledge in practical applications, and it is usually impossible to obtain a sufficient amount of data for model generalization. Therefore, the answer to the question that may be obtained by only inputting the first question into the question-answering model may have a large deviation. Therefore, when determining that the first question contains at least one entity object of a preset entity category, the embodiment of the present application will jump to the following step 803 to determine a more accurate answer to the first question through steps 802 to 803.

[0096] Step 802: Determine the text similarity between the entity object and each pre-stored domain knowledge, and select candidate knowledge for the first question from the pre-stored multiple domain knowledge according to the text similarity; wherein any domain knowledge is extracted from the structured data of the specified domain;

[0097] Considering that users usually ask multiple rounds of questions in intelligent question-and-answer scenarios, for example, a user asks in the previous round "What is the GDP of Region A in 2019?", and the first question asked by the user in the current round may be "What about 2020?".

[0098] For questions such as "What about 2020?", although it is easy to know from the context semantics that the user is currently asking what the GDP of Region A is in 2020, it is impossible to parse it into a complete question from the model's perspective. If the conventional method is used to directly perform feature recognition on "What about 2020?" through the model, the correct answer cannot be output.

[0099] Based on this, before executing step 802, it is also necessary to determine whether the first question currently raised by the user is a complete question. Fig. 9 As shown, the following steps are included:

[0100] Step 901: Determine the number of preset entity categories included in the first question.

[0101] That is, determine which of the five preset entity categories, namely time, place, indicator name, indicator unit and indicator value, are included in the first question. For example, the above example "What is the GDP of District A in 2019?" includes three items, namely time, place and indicator name, and the number of its categories is 3; while the above example "What about 2020?" only includes the item of time, and the number of its categories is 1.

[0102] Step 902: Determine whether the number of categories is greater than a first number threshold and less than a second number threshold;

[0103] In the embodiment of the present application, the first quantity threshold is set to 0, and the second quantity threshold is set to 3. That is, if the current question only contains 1 or 2 preset entity categories, it is considered that the current question is not a complete question.

[0104] Step 903: If the number of questions is not greater than the first quantity threshold and is less than the second quantity threshold, determining that the first question is a complete question;

[0105] Step 904: If the number of questions is greater than the first quantity threshold and less than the second quantity threshold, then obtaining all second questions raised by the user in the first historical period;

[0106] Usually, when the question currently asked by the user is not a complete question, ignoring the fact that the user input the question incorrectly, the question currently asked by the user is mostly an extension of the previous rounds of questions. After determining that the first question is an incomplete question, it is necessary to obtain all the second questions asked by the user in the first historical period. For example, obtain all the second questions asked by the user in the last 5 minutes.

[0107] For example, if a user asks "What is the GDP of Region A in 2019?" in the previous round and asks "What about 2020?" in the current round, in actual applications, incomplete questions are mostly related to questions in the last two or three rounds. Therefore, when executing step 904, the last three questions asked by the user in the last 5 minutes can be used as the second question.

[0108] Step 905: Update the first question according to the entity objects of each preset entity category in the first question and the second question.

[0109] When executing step 905, for any second question, if the preset entity categories included in the second question include all target categories, and the entity object of the second question under at least one target category is different from that of the first question, the second question is used as the third question. After the third question is selected from the second questions, a fourth question whose question time is adjacent to the first question is selected from the third questions.

[0110] The above target categories are all the preset entity categories included in this round of questions (i.e. the first question). Fig.10 As shown, the first question of this round and the second questions obtained from the first historical period are sorted in order from early to late in the questioning sequence to obtain:

[0111] Question 2 1 “Hello”;

[0112] Question 2: “What is the per capita income of District A in 2020?”;

[0113] Question 2 3 “What is the GDP of Region A in 2019?”;

[0114] The first question is “What about 2020?”.

[0115] Since the first question contains only one preset entity category "time category", the target category is "time category". At this time, the second question 2 and the second question 3 containing "time category" are found from all the second questions. Next, the third question is selected from the second question 2 and the second question 3, which has at least one entity object under the target category different from that of the first question.

[0116] The entity object under the preset entity category "time category" in the second question 2 is "2020", while the entity object under the preset entity category "time category" in the second question 3 is "2019". Therefore, there is at least one entity object under the target category in the second question 3 that is different from the first question, and the second question 3 is regarded as the third question.

[0117] Another example Fig.11As shown, assume that the first question is "What about District B in 2020?". At this time, the preset entity categories contained in the first question are "time category" and "place category". Therefore, the target categories are "time category" and "place category". At this time, the second question 2 and the second question 3 that contain both "time category" and "place category" are found from all the second questions. Furthermore, the entity object of the "time category" in the second question 2 is "2020", and the entity object of the "place category" is "District A". The entity objects under these two preset entity categories are different from those in the first question, so the second question 2 is the third question. Although the entity object of the "time category" in the second question 3 is the same as that in the first question, its entity object of the "place category" is different from that in the first question, so the second question 3 is also regarded as the third question.

[0118] Then, find the fourth question whose question time is closest to the question time of the first question from the third question. For any target category in the fourth question, replace the entity object of the fourth question under the target category with the entity object of the first question under the target category; and use the fourth question after the entity object is replaced as the first question after the question is updated.

[0119] Still with the above Fig.11 For example, Fig.12 As shown, Fig.11 The third question in is the second question 2 and the second question 3. Since the question time of the second question 3 is closest to the first question, the second question 3 is taken as the fourth question. At this time, the entity objects under the target category in the first question and the fourth question are replaced, and the updated first question is obtained. Fig.12 As shown, the target categories in the first question "What about District B in 2020?" are preset entity categories of "time category" and "place category", and the corresponding entity objects in the fourth question "What is the GDP of District A in 2019?" are "2019" and "District A" respectively. By swapping the entity objects under the target categories in the first question and the fourth question, a new first question "What is the GDP of District B in 2019" is generated.

[0120] In addition, in the aforementioned process of updating the first question, if the fourth question does not exist in the first historical period, for example, if there is no second question or no third question in the first historical period, the fourth question will not exist. Fig.13 As shown, a prompt message indicating that the current question cannot be answered is output to the user.

[0121] Through the above process, it can be determined that the first question before executing step 802 is a complete question of the domain type. Next, when executing step 802, it is pre-determined whether the first question can be split into multiple sub-questions.

[0122] During implementation, it is determined whether the number of entity objects under any preset entity category in the first question is greater than a third quantity threshold (value is 1). When it is greater than the third quantity threshold, it is determined that the first question can be split into multiple sub-questions, otherwise it is determined that the first question cannot be split.

[0123] Specific as Fig.14 As shown, assuming that the first question is "What is the GDP of District A and District B in 2019?" The preset entity categories contained in this question are time category, location category, and indicator name category. Among them, the entity object of the time category is 2019, the entity objects of the location category are District A and District B, and the entity object of the indicator name category is GDP. Since the number of entity objects of the location category is 2, which is greater than the third quantity threshold of 1. Therefore, it is shown that the first question can be split into sub-questions.

[0124] The sub-question splitting rule in the embodiment of the present application is to ensure that any sub-question contains all the preset entity categories in the first question, and there is only one entity object under each preset entity category in the sub-question. And the entity objects of any two sub-questions under at least one preset entity category are different. Taking the aforementioned first question "What is the GDP of District A and District B in 2019?" as an example, the only preset entity category in which the number of entity objects under the same preset entity category is greater than the third threshold is the location category. And the number of entity objects in the location class is 2 (Area A and Area B). Therefore, under the premise of meeting the above rules, the first question can be split into two sub-questions: "What is the GDP of District A in 2019?" and "What is the GDP of District B in 2019?".

[0125] Assume that the first question is "What is the GDP of District A and District B in 2019 and 2020?" Since the number of entity objects under the same preset entity category is greater than the third threshold, the preset entity categories include the time category and the location category. The number of entity objects in the time category is 2 (2019 and 2020), and the number of entity objects in the location category is also 2 (District A and District B). Therefore, by arranging and combining the entity objects of the time category and the location category, four combinations can be obtained: (2019, District A), (2019, District B), (2020, District A) and (2020, District B), that is, four sub-questions are obtained, namely "What is the GDP of District A in 2019?", "What is the GDP of District A in 2020?", "What is the GDP of District B in 2019?" and "What is the GDP of District B in 2020?".

[0126] After the first problem is divided into multiple sub-problems through the above process, for any sub-problem, all entity objects under the preset entity category are extracted from the sub-problem, and the text similarity between the entity object and each domain knowledge is determined based on the text similarity algorithm. Then, the knowledge to be processed whose text similarity is greater than the first similarity threshold is selected from the multiple pre-stored domain knowledge.

[0127] The domain knowledge in the above process is extracted from the structured data of the domain class in the above steps 101 to 103, and is used as the knowledge reserve of the model for the domain. Fig.14 Taking the sub-question "What is the GDP of District A in 2019?" obtained by splitting the first question as an example, the entity objects under each preset entity category are pre-removed from the word question in advance during implementation. That is, 2019, District A, and GDP. Then these entity objects are used as a text, and the text similarity is calculated with each domain knowledge. Then, the knowledge to be processed whose similarity is greater than the first similarity threshold (70%) is found from the multiple pre-stored domain knowledge.

[0128] Next, the semantic similarity between the text composed of these entity objects (i.e., GDP of District A in 2019) and each piece of knowledge to be processed is calculated, and the knowledge to be processed whose semantic similarity is greater than the second similarity threshold (75%) is used as the candidate knowledge for the sub-problem. It should be noted that text similarity and semantic similarity calculations are both commonly used calculation methods in the field of comparative learning, and the embodiments of this application do not limit the similarity algorithm. This application pre-uses text similarity to filter the domain knowledge here to find the knowledge to be processed that is similar to the sub-problem in terms of text content. Then the semantic similarity of the sub-problem and each piece of knowledge to be processed is calculated one by one to screen out candidate knowledge that is closer to the semantic expression of the sub-problem from the perspective of semantic features. Finally, the candidate knowledge of all sub-problems is used as the candidate knowledge of the first problem, thereby finding candidate knowledge with a greater correlation with the first problem from the pre-stored domain knowledge.

[0129] It should also be noted that if the first question cannot be split into sub-questions (i.e., the number of entities under each preset entity category in the first question is 1), the first question can be processed using the aforementioned processing method for any sub-question. Specifically, entity objects under all preset entity categories are extracted from the first question, and after the entity objects are combined into text, the text similarity and semantic similarity are screened, and the screening results are used as candidate knowledge for the first question.

[0130] Step 803: Perform feature recognition on the entity object and the candidate knowledge through the trained question-answering model to obtain the answer to the first question, and display the answer to the user; wherein the question-answering model is obtained by training a basic question-answering model based on the domain knowledge; the basic question-answering model is any trained intelligent question-answering model for answering non-domain questions.

[0131] During implementation, historical questions raised by users in the second historical period and the answers to each historical question output by the question-answering model are obtained. A first text consisting of the first question, historical questions, and the answers to each historical question is constructed, and a second text consisting of the first question and each candidate knowledge is constructed.

[0132] The second historical period in the embodiment of the present application can be set to the last 30 minutes or 60 minutes. Specifically, each historical question and corresponding answer of the user in the second period is obtained. Fig.15 For example, the user's historical questions and corresponding answers in the second period are in the following order from earliest to latest:

[0133] History Question 1 “Hello”;

[0134] Question 1: "Hello, how can I help you?"

[0135] History Question 2 “What is the per capita income of District A in 2020?”;

[0136] Question 2: "5,700 yuan"

[0137] History Question 3 “What is the GDP of Region A in 2019?”;

[0138] Answer to question 3: “4 billion yuan”.

[0139] The text consisting of the first question, each historical question and the corresponding answer to the historical question is taken as the first text. Since there are three historical questions and answers here, three first texts can be constructed. They are "first question + historical question 1 + answer to question 1", "first question + historical question 2 + answer to question 2" and "first question + historical question 3 + answer to question 3".

[0140] Assume that the first question currently raised by the user is "What is the GDP of District A and District B in 2019?", and the candidate knowledge of the first question is "The GDP of District A in 2019 is 4 billion yuan" and "The GDP of District B in 2019 is 5.2 billion yuan". Then the text composed of the first question and each candidate knowledge is used as the second text, thereby obtaining two second texts, namely "What is the GDP of District A and District B in 2019? The GDP of District A in 2019 is 4 billion yuan" and "What is the GDP of District A and District B in 2019? The GDP of District B in 2019 is 5.2 billion yuan".

[0141] Next, the first text and the second text are input into the question-answering model, so that the question model extracts features from the first text and the second text respectively to obtain first text features of the first text and second text features of the second text, and the question-answering model performs feature recognition on the feature splicing result of the first text features and the second text features to obtain an answer to the first question.

[0142] Specific as Fig.16 As shown, the encoder and decoder parts of the model are improved in the embodiment of the present application. Specifically, the encoder part is set to perform a shared encoder operation on the text input within a preset time length, so that the model extracts features from each text input within the preset time length one by one, and sets the decoder to perform feature fusion on the feature extraction results of each text, and performs feature recognition on the fused features. On the one hand, the above structure only performs encoding on one input text at a time, and the calculation time increases linearly with the number of input texts, rather than exponentially, which greatly improves the performance compared to traditional encoding structures (such as encoder-decoder). On the other hand, aggregating information in the decoder can give full play to the powerful capabilities of the pre-trained model itself, and can better aggregate features from multiple texts, thereby improving the authenticity and accuracy of the answer to the question.

[0143] Based on the same inventive concept, the present application embodiment provides the following Fig.17 The intelligent question-answering device shown comprises:

[0144] The question acquisition unit 171 is configured to acquire a first question currently raised by a user and determine an entity object in the first question that belongs to a preset entity category; wherein the preset entity category is determined according to a specified field;

[0145] The candidate knowledge unit 172 is configured to determine the text similarity between the entity object and each pre-stored domain knowledge, and select candidate knowledge for the first question from the pre-stored multiple domain knowledge according to the text similarity; wherein any domain knowledge is extracted from the structured data of the specified domain;

[0146] The answer display unit 173 is configured to perform feature recognition on the entity object and the candidate knowledge through a trained question-answering model, obtain a question answer to the first question, and display the question answer to the user; wherein the question-answering model is obtained by training a basic question-answering model based on the domain knowledge; the basic question-answering model is any trained intelligent question-answering model for answering non-domain questions.

[0147] In some possible embodiments, before determining the text similarity between the entity object and each pre-stored domain knowledge, the candidate knowledge unit is further configured as follows:

[0148] Determining the number of categories of the preset entity categories included in the first question;

[0149] If the number of categories is greater than a first number threshold and less than a second number threshold, then obtaining all second questions raised by the user in the first historical period;

[0150] The first question is updated according to the entity objects of each preset entity category in the first question and the second question.

[0151] In some possible embodiments, the first question is updated according to the entity objects of each preset entity category in the first question and the second question, and the candidate knowledge unit is configured as follows:

[0152] For any second question, if the preset entity categories included in the second question include all target categories, and the entity objects of the second question under at least one target category are different from those of the first question, the second question is used as the third question; wherein any target category is a preset entity category included in the first question;

[0153] Selecting a fourth question from the third questions, the fourth question being asked at a time adjacent to the first question;

[0154] For any target category in the fourth question, the entity object of the fourth question under the target category is replaced with the entity object of the first question under the target category; and the fourth question after the entity object is replaced is used as the first question after the question is updated.

[0155] In some possible embodiments, the candidate knowledge unit is further configured as:

[0156] If the fourth question does not exist within the first historical period, a prompt message indicating that the current question cannot be answered is output to the user.

[0157] In some possible embodiments, before performing the step of determining the text similarity between the entity object and each pre-stored domain knowledge, the candidate knowledge unit of the method is further configured as follows:

[0158] If the number of entity objects under any preset entity category in the first question is greater than a third quantity threshold, splitting the first question into multiple sub-questions;

[0159] Among them, any sub-question includes all preset entity categories in the first question, and each preset entity category in the sub-question includes an entity object; any two sub-questions have different entity objects under at least one preset entity category.

[0160] In some possible embodiments, the step of determining the text similarity between the entity object and each pre-stored domain knowledge and selecting candidate knowledge for the first question from the pre-stored domain knowledge according to the text similarity is performed, and the candidate knowledge unit is configured as follows:

[0161] For any sub-question, extract all entity objects under the preset entity categories from the sub-question, and determine the text similarity between the entity object and each piece of domain knowledge based on the text similarity algorithm;

[0162] Selecting knowledge to be processed whose text similarity is greater than a first similarity threshold from the plurality of domain knowledge;

[0163] Determine the semantic similarity between the entity object and each piece of knowledge to be processed, and use the knowledge to be processed whose semantic similarity is greater than a second similarity threshold as candidate knowledge for the sub-problem;

[0164] The candidate knowledge of all sub-questions is used as the candidate knowledge of the first question.

[0165] In some possible embodiments, the step of performing feature recognition on the entity object and the candidate knowledge by using the trained question-answering model to obtain a question answer to the first question is performed, and the answer display unit is configured as follows:

[0166] Obtaining historical questions raised by the user in the second historical period and answers to each historical question output by the question-answering model;

[0167] Constructing a first text consisting of the first question, the historical question, and answers to each historical question, and constructing a second text consisting of the first question and each candidate knowledge;

[0168] By inputting the first text and the second text into the question-answering model, the question model performs feature extraction on the first text and the second text respectively to obtain first text features of the first text and second text features of the second text, and the question-answering model performs feature recognition on the feature splicing result of the first text features and the second text features to obtain a question answer to the first question.

[0169] In some possible embodiments, the pre-stored multiple pieces of domain knowledge are obtained in the following manner:

[0170] In response to the knowledge reserve indication, obtaining structured data of the specified field; wherein the structured data at least includes form data and a knowledge graph;

[0171] For any form data, if each list of the form data contains a preset column name, the data content under the preset column name is extracted, and the domain knowledge is constructed according to the data content extracted under each preset column name in the form data; wherein the preset list is determined according to the preset entity category;

[0172] For any entity object in any knowledge graph, determine all directly connected entity objects in the knowledge graph that are directly connected to the entity object;

[0173] A target entity object belonging to the preset entity object category is selected from the entity object and the entity objects directly connected to the entity object, and the domain knowledge is constructed according to the entity object and the target entity object.

[0174] Based on the same inventive concept, the present application embodiment also provides an electronic device, specifically Fig.18 As shown, it includes a data transmission unit 181 and a processor 182:

[0175] The data transmission unit 181 is configured to: obtain the first question currently rejected by the user;

[0176] The processor 182 is configured to: determine an entity object in the first question that belongs to a preset entity category; wherein the preset entity category is determined according to a specified field;

[0177] Determine the text similarity between the entity object and each pre-stored domain knowledge, and select candidate knowledge for the first question from the pre-stored multiple domain knowledge according to the text similarity; wherein any domain knowledge is extracted from the structured data of the specified domain;

[0178] The entity object and the candidate knowledge are feature-recognized by a trained question-answering model to obtain an answer to the first question, and the answer is displayed to the user; wherein the question-answering model is obtained by training a basic question-answering model based on the domain knowledge; the basic question-answering model is any trained intelligent question-answering model for answering non-domain questions.

[0179] In some possible embodiments, before determining the text similarity between the entity object and each pre-stored domain knowledge, the processor 182 is further configured to:

[0180] Determining the number of categories of the preset entity categories included in the first question;

[0181] If the number of categories is greater than a first number threshold and less than a second number threshold, then obtaining all second questions raised by the user in the first historical period;

[0182] The first question is updated according to the entity objects of each preset entity category in the first question and the second question.

[0183] In some possible embodiments, the processor 182 is configured to:

[0184] For any second question, if the preset entity categories included in the second question include all target categories, and the entity objects of the second question under at least one target category are different from those of the first question, the second question is used as the third question; wherein any target category is a preset entity category included in the first question;

[0185] Selecting a fourth question from the third questions, the fourth question being asked at a time adjacent to the first question;

[0186] For any target category in the fourth question, the entity object of the fourth question under the target category is replaced with the entity object of the first question under the target category; and the fourth question after the entity object is replaced is used as the first question after the question is updated.

[0187] In some possible embodiments, the processor 182 is further configured to:

[0188] If the fourth question does not exist within the first historical period, a prompt message indicating that the current question cannot be answered is output to the user.

[0189] In some possible embodiments, before determining the text similarity between the entity object and each pre-stored domain knowledge, the processor 182 is further configured to:

[0190] If the number of entity objects under any preset entity category in the first question is greater than a third quantity threshold, splitting the first question into multiple sub-questions;

[0191] Among them, any sub-question includes all preset entity categories in the first question, and each preset entity category in the sub-question includes an entity object; any two sub-questions have different entity objects under at least one preset entity category.

[0192] In some possible embodiments, to perform the step of determining the text similarity between the entity object and each pre-stored domain knowledge, and selecting candidate knowledge for the first question from the pre-stored domain knowledge according to the text similarity, the processor 182 is configured to:

[0193] For any sub-question, extract all entity objects under the preset entity categories from the sub-question, and determine the text similarity between the entity object and each piece of domain knowledge based on the text similarity algorithm;

[0194] Selecting knowledge to be processed whose text similarity is greater than a first similarity threshold from the plurality of domain knowledge;

[0195] Determine the semantic similarity between the entity object and each piece of knowledge to be processed, and use the knowledge to be processed whose semantic similarity is greater than a second similarity threshold as candidate knowledge for the sub-problem;

[0196] The candidate knowledge of all sub-questions is used as the candidate knowledge of the first question.

[0197] In some possible embodiments, the processor 182 is configured to:

[0198] Obtaining historical questions raised by the user in the second historical period and answers to each historical question output by the question-answering model;

[0199] Constructing a first text consisting of the first question, the historical question, and answers to each historical question, and constructing a second text consisting of the first question and each candidate knowledge;

[0200] By inputting the first text and the second text into the question-answering model, the question model performs feature extraction on the first text and the second text respectively to obtain first text features of the first text and second text features of the second text, and the question-answering model performs feature recognition on the feature splicing result of the first text features and the second text features to obtain a question answer to the first question.

[0201] In some possible embodiments, the pre-stored multiple pieces of domain knowledge are obtained in the following manner:

[0202] In response to the knowledge reserve indication, obtaining structured data of the specified field; wherein the structured data at least includes form data and a knowledge graph;

[0203] For any form data, if each list of the form data contains a preset column name, the data content under the preset column name is extracted, and the domain knowledge is constructed according to the data content extracted under each preset column name in the form data; wherein the preset list is determined according to the preset entity category;

[0204] For any entity object in any knowledge graph, determine all directly connected entity objects in the knowledge graph that are directly connected to the entity object;

[0205] A target entity object belonging to the preset entity object category is selected from the entity object and the entity objects directly connected to the entity object, and the domain knowledge is constructed according to the entity object and the target entity object.

[0206] The embodiment of the present application further provides a computer storage medium, in which computer program instructions are stored. When the instructions are executed on a computer, the computer executes the steps of the above-mentioned traffic signal control method.

[0207] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0208] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the 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 202 of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor 202 of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. 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.

[0209] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing 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.

[0210] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0211] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. An intelligent question-answering method, characterized in that: The method comprises: Acquire a first question currently raised by a user, and determine an entity object in the first question that belongs to a preset entity category; wherein the preset entity category is determined according to a specified field; Determine the text similarity between the entity object and each pre-stored domain knowledge, and select candidate knowledge for the first question from the pre-stored multiple domain knowledge according to the text similarity; wherein any domain knowledge is extracted from the structured data of the specified domain; The entity object and the candidate knowledge are feature-recognized by a trained question-answering model to obtain an answer to the first question, and the answer is displayed to the user; wherein the question-answering model is obtained by training a basic question-answering model based on the domain knowledge; the basic question-answering model is any trained intelligent question-answering model for answering non-domain questions.

2. The method according to claim 1, characterized in that: Before determining the text similarity between the entity object and each pre-stored domain knowledge, the method further includes: Determining the number of categories of the preset entity categories included in the first question; If the number of categories is greater than a first number threshold and less than a second number threshold, then obtaining all second questions raised by the user in the first historical period; The first question is updated according to the entity objects of each preset entity category in the first question and the second question.

3. The method according to claim 2, characterized in that The updating of the first question according to the entity objects of each preset entity category in the first question and the second question includes: For any second question, if the preset entity categories included in the second question include all target categories, and the entity objects of the second question under at least one target category are different from those of the first question, the second question is used as the third question; wherein any target category is a preset entity category included in the first question; Selecting a fourth question from the third questions, the fourth question being asked at a time adjacent to the first question; For any target category in the fourth question, the entity object of the fourth question under the target category is replaced with the entity object of the first question under the target category; and the fourth question after the entity object is replaced is used as the first question after the question is updated.

4. The method according to claim 3, characterized in that The method further comprises: If the fourth question does not exist within the first historical period, a prompt message indicating that the current question cannot be answered is output to the user.

5. The method according to claim 1, characterized in that Before determining the text similarity between the entity object and each pre-stored domain knowledge, the method further includes: If the number of entity objects under any preset entity category in the first question is greater than a third quantity threshold, splitting the first question into multiple sub-questions; Among them, any sub-question includes all preset entity categories in the first question, and each preset entity category in the sub-question includes an entity object; any two sub-questions have different entity objects under at least one preset entity category.

6. The method according to claim 5, characterized in that The determining of the text similarity between the entity object and each pre-stored domain knowledge, and selecting candidate knowledge for the first question from the pre-stored domain knowledge according to the text similarity, includes: For any sub-question, extract all entity objects under the preset entity categories from the sub-question, and determine the text similarity between the entity object and each piece of domain knowledge based on the text similarity algorithm; Selecting knowledge to be processed whose text similarity is greater than a first similarity threshold from the plurality of domain knowledge; Determine the semantic similarity between the entity object and each piece of knowledge to be processed, and use the knowledge to be processed whose semantic similarity is greater than a second similarity threshold as candidate knowledge for the sub-problem; The candidate knowledge of all sub-questions is used as the candidate knowledge of the first question.

7. The method according to claim 1, characterized in that The step of performing feature recognition on the entity object and the candidate knowledge by using the trained question-answering model to obtain an answer to the first question includes: Obtaining historical questions raised by the user in the second historical period and answers to each historical question output by the question-answering model; Constructing a first text consisting of the first question, the historical question, and answers to each historical question, and constructing a second text consisting of the first question and each candidate knowledge; By inputting the first text and the second text into the question-answering model, the question model performs feature extraction on the first text and the second text respectively to obtain first text features of the first text and second text features of the second text, and the question-answering model performs feature recognition on the feature splicing result of the first text features and the second text features to obtain a question answer to the first question.

8. The method according to any one of claims 1 to 7, characterized in that: The pre-stored multiple pieces of domain knowledge are obtained in the following manner: In response to the knowledge reserve indication, obtaining structured data of the specified field; wherein the structured data at least includes form data and a knowledge graph; For any form data, if each list of the form data contains a preset column name, the data content under the preset column name is extracted, and the domain knowledge is constructed according to the data content extracted under each preset column name in the form data; wherein the preset list is determined according to the preset entity category; For any entity object in any knowledge graph, determine all directly connected entity objects in the knowledge graph that are directly connected to the entity object; A target entity object belonging to the preset entity object category is selected from the entity object and the entity objects directly connected to the entity object, and the domain knowledge is constructed according to the entity object and the target entity object.

9. An electronic device, characterized in that: Including data transmission unit and processor: The data transmission unit is configured to: obtain a first question currently rejected by the user; The processor is configured to: determine an entity object in the first question that belongs to a preset entity category; wherein the preset entity category is determined according to a specified field; Determine the text similarity between the entity object and each pre-stored domain knowledge, and select candidate knowledge for the first question from the pre-stored multiple domain knowledge according to the text similarity; wherein any domain knowledge is extracted from the structured data of the specified domain; The entity object and the candidate knowledge are feature-recognized by a trained question-answering model to obtain an answer to the first question, and the answer is displayed to the user; wherein the question-answering model is obtained by training a basic question-answering model based on the domain knowledge; the basic question-answering model is any trained intelligent question-answering model for answering non-domain questions.

10. The device according to claim 9, characterized in that Before executing the step of determining the text similarity between the entity object and each pre-stored domain knowledge, the processor is further configured to: Determining the number of categories of the preset entity categories included in the first question; If the number of categories is greater than a first number threshold and less than a second number threshold, then obtaining all second questions raised by the user in the first historical period; The first question is updated according to the entity objects of each preset entity category in the first question and the second question.

Images