An intelligent answering method and device

Abstract

The application relates to the technical field of natural language processing, and is an intelligent answering method and device. The intelligent answering method comprises the following steps: acquiring multi-round question analysis information of a user and answer analysis information corresponding to each round of question analysis information; determining an answer key point set and an answer key point sequence based on all the answer analysis information; fusing the multi-round question analysis information to generate global enhanced context information; and generating guided answering information based on the answer key point set, the answer key point sequence and the global enhanced context information. The application can provide a thinking-guided answering mode for students, cultivate the independent thinking ability of the students, and improve the comprehensiveness of the students in mastering knowledge points.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing, and in particular to an intelligent question answering method and device. BACKGROUND

[0002] With the continuous development of science and technology, intelligent question answering technology brings great convenience to students, parents and teachers. The current question answering method is usually that the student inputs a question to an electronic device, and the electronic device directly matches the answer to the question according to the key words in the question recognized by the semantic recognition function.

[0003] The current question answering method only directly answers the single question itself, cannot provide coherent thinking guidance for students according to the logical relationship of knowledge points in the previous and subsequent questions of the students, and cannot convert the hesitation and fragmented information in the student's expression into the knowledge points of the students in the learning process, so as to further dig out the potential needs of the students in the questioning process.

[0004] Therefore, it is a problem to be solved to provide a question answering method to guide students to think coherently. SUMMARY

[0005] To solve the above technical problems, the present application provides an intelligent question answering method, device, electronic device and storage medium to solve the problem that the question answering method in the related art cannot provide coherent thinking guidance for students.

[0006] In a first aspect, an intelligent question answering method is provided in the present embodiment, comprising:

[0007] obtaining multi-round question analysis information of a user and answer analysis information corresponding to each round of question analysis information;

[0008] determining an answer key point set and a corresponding answer key point sequence based on all the answer analysis information;

[0009] fusing the multi-round question analysis information to generate global enhanced context information;

[0010] generating guided question answering information based on the answer key point set, the answer key point sequence and the global enhanced context information, the guided question answering information including guided question information and answer information corresponding to the guided question information.

[0011] In some embodiments, determining the answer key point set and the corresponding answer key point sequence based on all the answer analysis information comprises:

[0012] The answer analysis information is divided into semantic granularity to obtain first granularity semantic features, second granularity semantic features and third granularity semantic features, and the semantic granularity of the first granularity semantic features, the second granularity semantic features and the third granularity semantic features decreases gradually;

[0013] The first granularity semantic features, the second granularity semantic features and the third granularity semantic features are fused to obtain fusion features;

[0014] The fusion features are decoded to obtain the answer key point set and the answer key point sequence.

[0015] In some embodiments, the first granularity semantic features, the second granularity semantic features and the third granularity semantic features are fused to obtain fusion features, including:

[0016] The second granularity semantic features and the third granularity semantic features are spliced to obtain spliced semantic features;

[0017] The cross-level association weight between the first granularity semantic features and the spliced semantic features is determined;

[0018] The spliced semantic features are semantically enhanced based on the cross-level association weight to obtain enhanced spliced semantic features;

[0019] The enhanced spliced semantic features and the first granularity semantic features are fused to obtain the fusion features.

[0020] In some embodiments, the fusion features are decoded to obtain the answer key point set and the answer key point sequence, including:

[0021] The fusion features are convoluted to obtain convolutional fusion features;

[0022] The fusion features are screened based on the convolutional fusion features to obtain screened features;

[0023] The answer key point set and the answer key point sequence are obtained based on the screened features.

[0024] In some embodiments, the multiple rounds of question analysis information are fused to generate global enhanced context information, including:

[0025] The multiple rounds of question analysis information are contextually fused to obtain global context features;

[0026] The global context features are associated in multiple rounds of context to obtain the global enhanced context information.

[0027] In some embodiments, the context fusion of the multi-turn question analysis information obtains global context features, including:

[0028] The current-turn question analysis information in the multi-turn question analysis information is time-series convoluted with all the question analysis information before the current-turn question analysis information, to obtain current-turn time-series convolution features, the current-turn question analysis information being any question analysis information in the multi-turn question analysis information;

[0029] The current-turn time-series convolution features are context fused with all the time-series convolution features before the current-turn time-series convolution features, to obtain current-turn context features, the global context features including context features of all turns.

[0030] In some embodiments, the multi-turn context association of the global context features obtains global enhanced context information, including:

[0031] The current-turn context features are associated with each context feature before the current-turn context features, to obtain feature association weights between the current-turn context features and each context feature before the current-turn context features;

[0032] All the context features before the current-turn context features are weighted based on each feature association weight, to obtain weighted context features;

[0033] The weighted context features are fused with the current-turn context features, to obtain enhanced context features of the current turn;

[0034] The global enhanced context information includes enhanced context features of all turns.

[0035] In some embodiments, after the multi-turn context association of the global context features obtains the global enhanced context information, the method further includes:

[0036] The global enhanced context information is decoded to generate thinking intent description information.

[0037] In some embodiments, the generation of the guided question answering information based on the answer key points, the answer key point sequence, and the global enhanced context information includes:

[0038] The semantic similarity between the answer key point set and the thinking intent description information is determined;

[0039] The semantic association matching features are generated based on the semantic similarity and the answer key point set;

[0040] perform a decoding operation on the semantic association matching feature, the answer key point set and the answer key point sequence to obtain the guided questioning information;

[0041] perform a decoding operation on the guided questioning information, the answer key point set and the answer key point sequence to obtain answer information corresponding to the guided questioning information.

[0042] In a second aspect, an intelligent question answering device is provided in the present embodiment, comprising:

[0043] An acquisition module is configured to acquire multi-round question analysis information of a user and answer analysis information corresponding to each round of question analysis information;

[0044] A determination module is configured to determine an answer key point set and a corresponding answer key point sequence based on all answer analysis information;

[0045] A fusion module is configured to fuse the multi-round question analysis information to generate global enhanced context information;

[0046] An information generation module is configured to generate guided question answering information based on the answer key point set, the answer key point sequence and the global enhanced context information, wherein the guided question answering information comprises guided questioning information and answer information corresponding to the guided questioning information.

[0047] In a third aspect, an electronic device is provided in the present embodiment, comprising a memory and a processor, characterized in that the memory stores a computer program, and the processor executes the computer program to implement the intelligent question answering method of the first aspect and any one of the embodiments.

[0048] In a fourth aspect, a storage medium is provided in the present embodiment, which stores a computer program, and the program is executed by a processor to implement the intelligent question answering method of the first aspect and any one of the embodiments.

[0049] Compared with the related art, the intelligent answering method provided in the embodiments of the present application determines the answer key point set and the corresponding answer key point sequence from the multi-round answer analysis information of the user, instead of extracting the answer key point through a single question, thereby avoiding the omission of the key information of the question and ensuring the completeness of the answering and the logical sequence of the key points. Further, the multi-round question analysis information is fused, thereby facilitating the acquisition of the global question from the fused information and the acquisition of the potential demand and the knowledge card point of the user from the global question, which provides corresponding support for the accurate answering in the subsequent question and answer process. Finally, the answer key point set, the answer key point sequence and the global enhanced context information are combined, and guided questions are given according to the logical sequence of the key points, thereby providing the student with coherent thinking guidance. And through the global enhanced context information, the knowledge card point and the potential demand of the student can be acquired from the fragmented information in the student's expression, the comprehensiveness of the guided question and answer content is improved, and the student's mastery of the overall knowledge is improved through the answer to all the question key points.

[0050] The details of one or more embodiments of the present application are presented in the following drawings and description to make other features, objects and advantages of the present application more apparent. BRIEF DESCRIPTION OF DRAWINGS

[0051] The drawings described herein are intended to provide further understanding of the present application, form a part of the present application, and the illustrative embodiments of the present application and the description thereof are used to explain the present application, and do not constitute an improper limitation on the present application. In the drawings:

[0052] Figure 1 is a hardware structure block diagram of a terminal of an intelligent answering method provided by the embodiments of the present application;

[0053] Figure 2 is a flowchart of an intelligent answering method provided by the embodiments of the present application;

[0054] Figure 3 is a flowchart of generating fused features provided by the embodiments of the present application;

[0055] Figure 4 is a flowchart of generating global enhanced context information provided by the embodiments of the present application;

[0056] Figure 5 is a structure block diagram of an intelligent answering device provided by the embodiments of the present application. DETAILED DESCRIPTION

[0057] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all the other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0058] The method embodiments provided in the present embodiment can be executed in a terminal, a computer or a similar computing device. For example, the method embodiments are run on a terminal, Figure 1 is a hardware structure block diagram of a terminal of an intelligent answering method provided by the embodiments of the present application. As shown in Figure 1 , the terminal can include one or more (only one is shown in Figure 1 ) processor 102 and memory 104 for storing data, wherein the processor 102 can include but is not limited to a processing device such as a microcontroller unit (MCU) or a field programmable gate array (FPGA). The terminal can also include a transmission device 106 for communication function and an input / output device 108. Those skilled in the art can understand that Figure 1 the structure shown is only schematic, which does not limit the structure of the terminal. For example, the terminal can include more or less components than those shown in Figure 1 , or have a different configuration from that shown in Figure 1 .

[0059] The memory 104 can be used to store computer programs, for example, software programs of application software and modules, such as computer programs corresponding to the intelligent answering method in the present embodiment. The processor 102 executes various functional applications and data processing by running the computer programs stored in the memory 104, that is, implements the above method. The memory 104 can include a high-speed random access memory, and can also include a non-volatile memory, such as one or more magnetic storage devices, flash memories, or other non-volatile solid-state memories. In some embodiments, the memory 104 can further include a memory remotely arranged with respect to the processor 102, which can be connected to the terminal through a network. Embodiments of the above network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0060] The transmission device 106 is configured to receive or send data via a network. The network includes a wireless network provided by a communication provider of the terminal. In an embodiment, the transmission device 106 includes a network interface controller (NIC) that can be connected to other network devices through a base station so as to communicate with the Internet. In an embodiment, the transmission device 106 can be a radio frequency (RF) module configured to communicate with the Internet in a wireless manner.

[0061] With the continuous development of science and technology, intelligent answering technology brings great convenience to students, parents and teachers.

[0062] The current answering mode is usually that the student inputs a question to an electronic device, and the electronic device directly matches an answer to the question according to a keyword in the question recognized by a semantic recognition function.

[0063] The current answering mode only directly answers a single question itself, cannot provide coherent thinking guidance for the student according to a logical relationship between knowledge points in front and back questions of the student, and cannot convert hesitation and fragmented information in the student's expression into a knowledge card point of the student in a learning process, and thus it is difficult to mine potential needs of the student in a questioning process.

[0064] Therefore, in the intelligent answering scenario, how to provide an answering mode to guide the student to think coherently is a problem to be solved.

[0065] In the embodiment, an intelligent answering method is provided, Figure 2 is a flowchart of an intelligent answering method provided by the embodiment, and an execution subject of the method can be an electronic device. Optionally, the electronic device can be a server or a terminal device, but the application is not limited thereto. Specifically, as shown in Figure 2 the flowchart includes the following steps:

[0066] In step S201, multi-round question analysis information of a user and answer analysis information corresponding to each round of question analysis information are obtained.

[0067] For example, in an intelligent answering scenario, a user usually inputs a question to be answered to an electronic device in a manner of voice, text or handwriting, and then the electronic device can give a corresponding answer to the question by using a conventional intelligent algorithm. The user is usually a student, and in the process, the student needs to input multiple questions to be answered due to incomplete mastery of knowledge points, and thus the electronic device outputs multiple corresponding answers.

[0068] Further, the electronic device can collect the multi-round question information input by the student, and collect the answer information output by the electronic device for each question, and analyze the collected question information and answer information, wherein the answer information can include problem solving thought information corresponding to the question. Then the collected question information and the corresponding answer information can be analyzed respectively, so as to obtain the corresponding multi-round question analysis information and the answer analysis information corresponding to each round question analysis information. Specifically, the collected question information and answer information can be text data, and then the collected text is analyzed by word segmentation, format standardization and other analysis methods, so as to obtain the multi-round question analysis information of the user and the answer analysis information corresponding to each round question analysis information.

[0069] Step S202, determining the answer key point set and the corresponding answer key point sequence based on all the answer analysis information.

[0070] Exemplarily, the key points in all the answer analysis information are extracted to obtain the key point set and the corresponding key point sequence.

[0071] Specifically, the core terms in all the answer analysis information are extracted, for example, concepts, data, proper nouns and logical words. Further, the repeated words in all the answer analysis information keywords are deleted, the semantically repeated words are merged, and the containing relationship words are screened. The extracted and de-duplicated keywords are sorted into a static set without duplication and omission according to the "classification logic", and finally the answer key point set is obtained.

[0072] Further, the keywords in the answer key point set are sorted according to the logical derivation or cognitive habit sequence, so as to obtain the answer key point sequence, which ensures the completeness, accuracy and logic of the answer key points.

[0073] Specifically, taking "finding the area of a triangle" as an example, the answer key point set obtained by the above method includes: "right triangle, the longest side is the hypotenuse, the right angle side includes a=4, b=5, the area formula of the triangle is S=½×a×b, and the data is substituted, S=10". Further, according to the logical derivation sequence, the following answer key point sequence can be determined: answer key point sequence 1 is to determine the type of triangle: right triangle, the longest side is the hypotenuse, and the right angle side includes a=4, b=5; answer key point sequence 2 is to recall the area formula of the right triangle: S=½×a×b; answer key point sequence 3 is to substitute the data: a=4, b=5, to get S=½×4×5; answer key point sequence 4 is to conclude that S=10.

[0074] Step S203, fusing the multi-round question analysis information to generate global enhanced context information.

[0075] Exemplarily, the core attributes in each round of question resolution information can be marked, the relevance between the multi-round question resolution information is further determined according to the core attributes, the multi-round question resolution information is fused according to the relevance between the multi-round question resolution information, and the global enhanced context information is obtained.

[0076] Specifically, if the first round question is “What is the Pythagorean theorem?” the core attributes in the corresponding question resolution information are: the Pythagorean theorem defines a²+b²=c², the hypotenuse and the right angle side of a right triangle; the second round question is “Use the Pythagorean theorem to find the right angle side a, (b=3, c=5)”, the core attributes in the corresponding question resolution information are: substitute the formula a²=5²-3², the calculation steps a²=16, a=4; the user supplements in the second round of questioning: “I need to confirm if c is the hypotenuse”, the core attributes in the corresponding question resolution information are: the user has doubts about “hypotenuse determination”, which needs to be strengthened in fusion; the third round question is “If it is an obtuse triangle, can the Pythagorean theorem still be used?” The core attributes in the corresponding question resolution information are: the user extends from “right triangle” to “obtuse triangle”, which needs to supplement the application scope of the theorem.

[0077] The relevance between each core attribute is further determined: the definition, application condition and application scope of the Pythagorean theorem, and the user's potential needs can be mined according to the relevance. If c is not the hypotenuse, it is an obtuse triangle, and the corresponding cosine theorem of the obtuse triangle (c²=a²+b²-2abcosC, C is an obtuse angle) can be supplemented.

[0078] Finally, the multi-round question resolution information is fused according to the relevance, and the global enhanced context information that can be obtained is: the Pythagorean theorem formula (a²+b²=c²), which is applied to a right triangle, the longest side of a right triangle is a hypotenuse, a right triangle length calculation example (a=4, b=3, c=5), the Pythagorean theorem cannot be directly applied to an obtuse triangle, and the cosine theorem of an obtuse triangle (c²=a²+b²-2abcosC, C is an obtuse angle) is needed.

[0079] In step S204, the guided question answering information is generated based on the answer key point set, the answer key point sequence and the global enhanced context information.

[0080] The guided question answering information includes guided question information and answer information corresponding to the guided question information.

[0081] Exemplarily, the answer key point set and the answer key point sequence are combined with the global enhanced context information to guide the user to think about the key points in a question mode. Specifically, the electronic device can output guiding question information: you can first recall the formula of the Pythagorean theorem (a²+b²=c²), where a and b are the legs of a right triangle, and c is the hypotenuse. What problem will occur if the hypotenuse is not determined first? After the user answers the thinking prompt or after a preset time, the corresponding answer is given: the hypotenuse must be determined first because in the Pythagorean theorem, the square of the hypotenuse is equal to the sum of the squares of the two legs (c²=a²+b²). If the legs are substituted for the hypotenuse, the calculation result will be wrong; the basis for determining the hypotenuse is that “in a right triangle, the longest side is the hypotenuse”. Further output the thinking prompt: if c is not the hypotenuse, how to calculate the length of c. After the user answers the thinking prompt or after a preset time, the corresponding answer is given: apply the cosine theorem (c²=a²+b²-2abcosC, C is an obtuse angle) to solve the length of c. In this way, the user is converted from “passive reception” to “active understanding”, and the user's thinking ability is improved by combining the logical information and potential needs in the global context.

[0082] In the above implementation process, the answer key point set and the answer key point sequence are determined from the user's multiple rounds of answer analysis information, rather than extracting the answer key points through a single question, which avoids missing the key information of the question and ensures the completeness of the answer and the logical order of the key points. Further, the multiple rounds of question analysis information are fused, so that the global question can be obtained from the fused information. The global question contains the key information of all questions, so that the user's potential needs and knowledge points can be obtained from the global question, providing corresponding support for accurate answering in the subsequent question and answer process. Finally, according to the combination of the answer key point set, the answer key point sequence, and the global enhanced context information, guiding questions are given according to the logical order of the key points, thereby providing coherent thinking guidance for the user. And through the global enhanced context information, the user's knowledge points and potential needs can be obtained from the fragmented information in the user's expression, and the user is further answered through all the key points of the question, improving the comprehensiveness of the user's knowledge points.

[0083] In some embodiments, determining the answer key point set and the corresponding answer key point sequence based on all the answer analysis information can include the following steps:

[0084] Step 1: performing semantic granularity division on all the answer analysis information to obtain first granularity semantic features, second granularity semantic features, and third granularity semantic features, the semantic granularity of the first granularity semantic features, the second granularity semantic features, and the third granularity semantic features decreasing in stages.

[0085] Exemplarily, all the answer analysis information is divided into three levels of semantic features according to semantic granularity, i.e., a first-granularity semantic feature, a second-granularity semantic feature, and a third-granularity semantic feature, and the semantic granularity of the first-granularity semantic feature, the second-granularity semantic feature, and the third-granularity semantic feature decreases step by step. That is, three different levels of semantic features are obtained, the first-granularity semantic feature is a high-level abstract feature, the second-granularity semantic feature is a medium-level abstract feature, and the third-granularity semantic feature is a low-level abstract feature.

[0086] Specifically, the original text is analyzed as: “To find the area of a right triangle, you need to first determine the formula, then substitute the data of the right angle side to calculate, and finally get the area of 6”. According to the semantic granularity, all the answer analysis information can be divided into a first-granularity semantic feature: calculating the area of a right triangle by substituting the formula.

[0087] The first-granularity semantic feature is decomposed into two second-granularity semantic features: 1. determining the area formula of a right triangle (S = ½ × a × b); 2. substituting the data of the right angle side (a = 3, b = 4) to calculate the area.

[0088] The second-granularity semantic feature is decomposed into three third-granularity semantic features: 1. parameter definition of the area formula (a and b are the right angle sides, and S is the area); 2. known data (a = 3, b = 4); 3. calculation process (½ × 3 × 4 = 6).

[0089] Step 2: Feature fusion is performed on the first-granularity semantic feature, the second-granularity semantic feature, and the third-granularity semantic feature to obtain a fused feature.

[0090] Further, the three different granularity semantic features are converted into vector form, and the three granularity vectors are fused by combining the attention mechanism. Specifically, the BERT model can be used to encode the three different granularity semantic features into a first-granularity word vector, a second-granularity word vector, and a third-granularity word vector. Further, the attention mechanism is used to assign corresponding weights to the three-granularity word vectors. For example, the first-granularity word vector is assigned the highest attention weight (such as 0.4) because it is a global framework; the second-granularity word vector is assigned a medium weight (such as 0.35) because it is a logical core; the third-granularity word vector is assigned a lower weight (such as 0.25) because it is a detail support; in addition, in order to preserve the relevance between levels, the “key correlation information” (such as the correspondence between the formula and the data) can also be enhanced locally. And through weighted summation and residual connection, the three-granularity word vectors are fused into a “fused feature vector”, and the key correlation information is added, i.e., the fused feature is: first-granularity word vector × 0.4 + second-granularity word vector × 0.35 + third-granularity word vector × 0.25 + key correlation information increment.

[0091] Step 3: Decode the fused features to obtain the set of answer key points and the sequence of answer key points.

[0092] Furthermore, the fused features can be transformed into structured key information units, namely, a set and a sequence of answer key points, through decoding. Specifically, a Named Entity Recognition (NER) model can be used to extract core entities from the fused features, such as "right-angled triangle", "3", "4", and "6"; and a key phrase extraction model, such as RAKE or TextRank, can be used to extract core methods or objectives, such as "calculate area" and "S=½×a×b"; and a semantic similarity-based clustering algorithm can be used to remove duplicate key points, forming a set of answer key points. Then, a dependency parsing method can be used to identify the sequential or causal relationships between key points, such as "choosing a formula" being a prerequisite for "substituting data", and "substituting data" being a prerequisite for "obtaining a result"; a topological sorting algorithm can be used to sort the answer key points according to logical relationships, generating a sequence of answer key points.

[0093] The key points to arrive at the final answer are: 1. Calculation objective: Area of ​​a right triangle; 2. Calculation method: Area formula S = ½ × a × b (a and b are the legs of the right triangle); 3. Given conditions: Leg a = 3, leg b = 4; 4. Calculation result: Area = 6.

[0094] Key points of the answer sequence: 1. Clarify the calculation objective (find the area of ​​a right triangle); 2. Determine the calculation method (use the right triangle area formula S=½×a×b, where a and b are the legs); 3. Substitute the known conditions (substitute a=3 and b=4 into the formula); 4. Complete the calculation and obtain the result (½×3×4=6, that is, the area is 6).

[0095] In the above implementation process, all answer parsing information is semantically divided to effectively separate semantic information at different levels, laying the foundation for accurate processing of answer information subsequently. Furthermore, the three different granularities of semantic information are fused to avoid semantic confusion at different levels while ensuring that the semantic relationships between features at each level are not disrupted. Finally, the fused features are decoded, effectively transforming the fused semantic features into structured key answer information, ensuring the logical consistency and completeness of the answer information.

[0096] In some embodiments, the first granularity semantic features, the second granularity semantic features, and the third granularity semantic features are fused to obtain fused features, which can be obtained through, for example... Figure 3 The process shown yields, Figure 3 This is a flowchart of generating fusion features provided in an embodiment of this application, specifically, as shown below. Figure 3 As shown, the process includes the following steps:

[0097] Step S301: Splicing the second granularity semantic feature and the third granularity semantic feature to obtain a spliced semantic feature.

[0098] Exemplarily, semantic granularity division is performed on all answer analysis information, specifically:

[0099]

[0100] wherein, X is all answer analysis information, H is a complete semantic feature, BiLSTM is a bidirectional long short-term memory network, and is used to convert the answer analysis information into the complete semantic feature, are respectively a first granularity semantic feature, a second granularity semantic feature and a third granularity semantic feature, are respectively a dimension selection matrix of the first granularity semantic feature, a dimension selection matrix of the second granularity semantic feature and a dimension selection matrix of the third granularity semantic feature, and are all diagonal matrices, diagonal elements of which are 0 or 1, and satisfy I is an identity matrix.

[0101] Further, the second granularity semantic feature and the third granularity semantic feature are spliced by a splicing function Concat to obtain a spliced semantic feature .

[0102] Step S302: Determining a cross-level association weight between the first granularity semantic feature and the spliced semantic feature.

[0103] The cross-level association weight between the first granularity semantic feature and the spliced semantic feature is determined by an attention mechanism A , that is wherein, Attention is the attention mechanism.

[0104] Step S303: Performing semantic enhancement on the spliced semantic feature based on the cross-level association weight to obtain an enhanced spliced semantic feature.

[0105] Further, the spliced semantic feature is enhanced in semantics according to a value projection matrix and the cross-level association weight to obtain an enhanced spliced semantic feature As an example, the enhanced spliced semantic feature can be determined by the following expression:

[0106]

[0107] Step S304: Performing feature fusion on the enhanced spliced semantic feature and the first granularity semantic feature to obtain a fused feature.

[0108] Further, the enhanced spliced semantic feature is fused with the first granularity semantic feature according to the structured fusion matrix to obtain a fusion feature . Specifically, if the structured fusion matrix is , the fusion feature may be determined by the following expression:

[0109]

[0110] In the above implementation process, the second granularity semantic feature and the third granularity semantic feature are spliced to obtain a spliced semantic feature, and the cross-level association weight between the first granularity semantic feature and the spliced semantic feature is determined through the attention mechanism, so that the cross-level association weight can identify the key information closely associated between the spliced semantic feature and the first granularity semantic feature, and then the spliced semantic feature is semantically enhanced according to the cross-level association weight, so as to obtain an enhanced spliced semantic feature. Finally, the enhanced spliced semantic feature is fused with the first granularity semantic feature to obtain a fusion feature. The feature fusion process is not indiscriminate feature splicing and fusion, but targeted fusion based on semantic association degree, which retains the semantic association of each level and reduces redundant information, avoiding logical break in problem solving caused by isolated hierarchical features.

[0111] In some embodiments, the decoding operation on the fusion feature to obtain the answer key point set and the answer key point sequence can include the following steps:

[0112] Step 1: Perform convolution operation on the fusion feature to obtain a convolution fusion feature.

[0113] Exemplarily, the convolution operation on the fusion feature can be performed by a convolution layer to obtain a convolution fusion feature. Specifically, the convolution fusion feature Z may be determined by the following expression:

[0114]

[0115] wherein, Conv1D is a one-dimensional convolution layer, , the value of which is 3, and Sigmoid is a Sigmoid function. It should be noted that the one-dimensional convolution layer with a convolution kernel size of 3 is taken as an example for description in the embodiments of the present application. In actual application, the convolution layer can also be a two-dimensional convolution layer or a three-dimensional convolution layer, and the convolution kernel size can be 4 or 5. The dimension of the convolution layer and the convolution kernel size can be adaptively adjusted according to actual data, which is not limited herein.

[0116] Step 2: Screen the fusion feature based on the convolution fusion feature to obtain a screened feature.

[0117] Step 3: decoding operation based on the screened features to obtain the answer key point set and the answer key point sequence.

[0118] Further, the fusion features are screened based on the convolution fusion features to obtain screened features, and the decoding operation is performed on the screened features to obtain the answer key point set and the corresponding answer key point sequence.

[0119] Specifically, the answer key point set Y and the corresponding answer key point sequence Seq can be determined by the following expression:

[0120]

[0121] wherein, Y is the answer key point set, Seq is the answer key point sequence, Decoder is the Transformer block decoder operation, is the multi-layer perception.

[0122] In the above implementation process, the convolution operation is performed on the fusion features to obtain the convolution fusion features, and the fusion features are screened based on the convolution fusion features to obtain the screened features. The filtering of redundant semantic information is realized through feature screening. Finally, the answer key point set and the corresponding answer key point sequence are generated through the decoding operation, so as to realize the whole process accurate processing from feature extraction to result generation, and completely solve the three major problems of hierarchical extraction, correlation reservation and ordered generation in the scene.

[0123] In some embodiments, the multi-round question analysis information is fused to generate global enhanced context information, which can include the following steps:

[0124] Step 1: context fusion of multi-round question analysis information to obtain global context features.

[0125] Exemplarily, the core entity of the question analysis information can be extracted by using the NER model, the relationship between entities can be identified by using the relationship extraction (RE) model, and each round of question can be encoded into a semantic vector by using a pre-trained language model, so as to obtain the deep semantic information in the multi-round question information. The pre-trained language model can be a BERT model and a RoBERTa model, etc., which is not limited herein.

[0126] Further, the dynamic weight fusion can be realized by using an attention mechanism. Specifically, a weight is assigned to each round of semantic vector according to the relevance between the question information of each round and the total question, then the local weight of the core entity in the semantic vector is enhanced, and finally the global context features are obtained by weighted summation of all semantic vectors, so as to convert the fragmented single question into structured and logical global context features, and realize the integration of multi-round question information.

[0127] Step 2: Multi-round context association is performed on the global context features to obtain global enhanced context information.

[0128] For example, the association information in the global context features is extracted, and the global enhanced context information is obtained according to the association information. Specifically, the association information can include logical information, demand information and extension information. The logical information can be information representing causal relationship or dependency relationship in the global context features, the demand information can be information representing demand in the global context features, and the extension information can be information representing that the user can have demand in the global context features. Further, the logical information, the demand information and the extension information are associated to the global context features to obtain the global enhanced context information.

[0129] In addition, the global enhanced context information can also be decoded into structured text such as classification list and paragraph by a text generation model such as GPT and T5, so as to ensure that the context features generated by the machine are consistent with the format of the manually sorted ones, and facilitate subsequent use.

[0130] In the above implementation process, the multi-round question analysis information is contextually fused to convert the fragmented single question into structured and logical global context features, and the global context features are contextually associated to make the obtained global enhanced context information associated with the logical context of all question information, and associated with the potential demand and extension knowledge of the user, thereby improving the accuracy and integrity of the context information.

[0131] In some embodiments, the context fusion of the multi-round question analysis information to obtain the global context features can include the following steps:

[0132] Step 1: The current round question analysis information in the multi-round question analysis information is time-series convoluted with all the question analysis information before the current round question analysis information to obtain the current round time-series convolution feature, and the current round question analysis information is any round question analysis information in the multi-round question analysis information.

[0133] Exemplarily, any round of question resolution information in the multi-round question resolution information is determined as the current round of question resolution information, the time sequence convolution network has a long time sequence modeling capability and can fully capture the time sequence semantic dependency relationship between the multi-round questions, and thus the current round of question resolution information and all the round of question resolution information before the current round of question resolution information can be time sequence convoluted through the time sequence convolution network, so as to obtain the current round of time sequence convolution feature.

[0134] Specifically, if the kth round of time sequence convolution feature is then the kth round of time sequence convolution feature can be determined by the following expression:

[0135]

[0136] wherein, TCN is the time sequence convolution network, respectively, are the 1st, 2nd, …, kth round of question resolution information.

[0137] Step 2: context fusion is performed on the current round of time sequence convolution feature and all the round of time sequence convolution features before the current round of time sequence convolution feature, so as to obtain the current round of context feature, and the global context feature includes all the round of context features.

[0138] Further, the current round of time sequence convolution feature and all the round of time sequence convolution features before the current round of time sequence convolution feature are spliced to obtain spliced time sequence convolution features, and context fusion is performed on the spliced convolution features through a context fusion matrix to obtain the current round of context feature.

[0139] Specifically, if the kth round of context feature is then the kth round of context feature can be determined by the following expression:

[0140]

[0141] wherein, is the context fusion matrix, respectively, are the 1st, 2nd, …, kth round of time sequence convolution feature. Therefore, all the round of context features can be determined in this way.

[0142] In the above implementation process, the time sequence convolution is performed on the current question resolution information and all the round of question resolution information before the current question resolution information, the time sequence semantic dependency relationship of the multi-round question is effectively captured, the time sequence convolution feature of each round of question is extracted from the 1st round to the kth round of question resolution information data, and then the context feature covering the complete information of the multi-round dialogue is obtained through the context fusion matrix after splicing the time sequence convolution features of each round, which avoids the defect of ignoring the time sequence association in the traditional single-round feature extraction mode and provides a comprehensive context basis for subsequent accurate understanding of the intention.

[0143] In some embodiments, the global context features are associated in multiple rounds of context association to obtain global enhanced context information. As an example, the global enhanced context information can be obtained by the flow as shown in Figure 4 Figure 4 is a flowchart of generating global enhanced context information provided by the embodiments of the present application. Specifically, as shown in Figure 4

[0144] Step S401: The association degree of the current round of context features and each round of context features before the current round of context features is evaluated to obtain the feature association weight between the current round of context features and each round of context features before the current round of context features.

[0145] Exemplarily, the current round of context features and each round of context features before the current round of context features are spliced to obtain spliced context features, and then the spliced context features are evaluated by a linear function and a nonlinear activation function to obtain the feature association weight between the current round of context features and each round of context features before the current round of context features.

[0146] Specifically, if the kth round of context features is , any round of context features before the kth round of context features is , where i < k, the feature association weight between and can be determined by the following expression:

[0147]

[0148] where Linear is a linear function and Sigmoid is a nonlinear activation function. Therefore, the feature association weight between the kth round of context features and each round of context features before the kth round of context features can be determined in this way.

[0149] Step S402: All rounds of context features before the current round of context features are weighted based on each feature association weight to obtain weighted context features.

[0150] Step S403: The weighted context features are fused with the current round of context features to obtain the enhanced context features of the current round. The global enhanced context information includes all rounds of enhanced context features.

[0151] Further, all rounds of context features before the current round of context features are weighted according to each feature association weight and then fused with the current round of context features to obtain the enhanced context features of the current round.​​​​​

[0152] Specifically, if the enhanced contextual feature in the kth round is Then the enhanced contextual features in the kth round It can be determined in the following way:

[0153]

[0154] This method was used to further identify the enhanced contextual features of all rounds.

[0155] In the above implementation process, a correlation weight calculation mechanism constructed by linear and nonlinear activation functions is used to evaluate the correlation between the contextual features of the k-th round and the previous k-1 rounds, obtaining feature correlation weights that reflect the closeness of the correlation between features in different rounds. Then, the contextual features of the previous k-1 rounds are weighted according to the correlation weights and fused with the contextual features of the k-th round to generate enhanced contextual features. This process is not a simple superposition of contextual features from multiple rounds, but a dynamic enhancement based on semantic correlation. It can highlight historical contextual information closely related to the current round's problem, weaken the interference of irrelevant information, and ensure that the enhanced contextual features contain information from multiple rounds while focusing on the core related semantics, thus avoiding the ambiguity of intent caused by the redundancy of information from multiple rounds.

[0156] In some embodiments, after performing multiple rounds of context association on the global context features to obtain global enhanced context information, the method further includes: decoding the global enhanced context information to generate information describing the thought intention.

[0157] For example, the enhanced context features of each round in the global enhanced context information can be decoded to obtain the corresponding round's thinking intention description information.

[0158] Specifically, the information describing the thinking intentions in round k. It can be determined by the following expression:

[0159]

[0160] in, Decoder This can be achieved through Transformer block decoder operations. Furthermore, this method can be used to obtain the thought intention description information corresponding to the enhanced contextual features in each round.

[0161] In the above implementation process, by decoding the enhanced contextual information, the contextual features focusing on the core related semantics can be directly transformed into a textual description of the user's thinking intention, realizing the transformation from multi-round contextual information to clear intention expression, and improving the effect of multi-round context capture and semantic association.

[0162] In some embodiments, generating the guided question answering information based on the answer key points, the answer key point sequence, and the global enhanced contextual information can include the following steps:

[0163] Step 1: determining the semantic similarity between the answer key point set and the thinking intention description information.

[0164] Exemplarily, the semantic similarity between the answer key point set and the thinking intention description information of each round can be determined by a cosine similarity function. Specifically, the semantic similarity between the answer key point set and the thinking intention description information of the kth round is The semantic similarity between the answer key point set and the thinking intention description information of the kth round can be determined by the following expression:

[0165]

[0166] wherein Sim is the cosine similarity calculation, Embed is the embedding function, and the semantic similarity between the answer key point set and the thinking intention description information of the kth round is The semantic similarity between the answer key point set and the thinking intention description information of the kth round can be expressed in the form of a matrix.

[0167] Step 2: generating the semantic association matching feature based on the semantic similarity and the answer key point set.

[0168] Further, the semantic association matching feature of the corresponding round is generated based on the product of the semantic similarity of each round and the answer key point set. Specifically, the semantic association matching feature of the kth round is The semantic association matching feature of the kth round can be determined by the following expression:

[0169]

[0170] Step 3: decoding the semantic association matching feature, the answer key point set, and the answer key point sequence to obtain the guided question information.

[0171] Step 4: decoding the guided question information, the answer key point set, and the answer key point sequence to obtain the answer information corresponding to the guided question information.

[0172] As an example, the guided question answering information includes the guided question information and the corresponding answer information, and the semantic association matching feature includes the semantic association matching features of all rounds. The guided question information is obtained by decoding the semantic association matching features of all rounds, the answer key point set, and the answer key point sequence; and the answer information corresponding to the guided question information is obtained by decoding the guided question information, the answer key point set, and the answer key point sequence.

[0173] Specifically, the semantic association matching features of all rounds are The guided question information is obtained by inputting the semantic association matching feature, the answer key point set Y, and the answer key point sequence Seq into a decoder of a Transformer block. Question The corresponding answer information can be determined by the following expression:

[0174]

[0175] Further, the guided question information is obtained by inputting the semantic association matching feature, the answer key point set Y, and the answer key point sequence Seq into a decoder of a Transformer block. Question The guided question information is obtained by inputting the semantic association matching feature, the answer key point set Y, and the answer key point sequence Seq into a decoder of a Transformer block. Question The corresponding answer information Answer The corresponding answer information can be determined by the following expression:

[0176]

[0177] The guided question information and the corresponding answer information are obtained by decoding the semantic association matching feature, the answer key point set, and the answer key point sequence into structured text information through the decoder of the Transformer block.

[0178] In the above implementation process, the semantic similarity between the answer key point set and the thinking intention description information is determined according to the cosine similarity, and the semantic association matching feature is generated according to the semantic similarity and the answer key point set, so that the subsequent generated question and answer information effectively integrates the multi-round question information and the corresponding answer information. The semantic association matching feature, the answer key point set, and the answer key point sequence are decoded to generate the guided question information and the corresponding answer information in the form of text, which can effectively guide students to think independently through the text information.

[0179] It should be noted that the steps shown in the above process or the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0180] In the present embodiment, an intelligent question answering device is also provided, which is used to implement the above-mentioned embodiments and preferred embodiments, and the description of which has been described above. The terms "module", "unit", "sub-unit" and the like used below can be a combination of software and / or hardware that can implement a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware, or a combination of software and hardware can also be implemented and conceived.

[0181] Figure 5is a structural block diagram of an intelligent answering device provided by an embodiment of the present application, as shown in the figure, the device comprises: Figure 5

[0182] The acquisition module 501 is configured to acquire multi-round question analysis information of a user and answer analysis information corresponding to each round of question analysis information.

[0183] The determination module 502 is configured to determine an answer key point set and a corresponding answer key point sequence based on all the answer analysis information.

[0184] The fusion module 503 is configured to fuse the multi-round question analysis information to generate global enhanced context information.

[0185] The information generation module 504 is configured to generate guided answering information based on the answer key point set, the answer key point sequence, and the global enhanced context information, the guided answering information comprising guided questioning information and answer information corresponding to the guided questioning information.

[0186] In some embodiments, the determination module 502 is specifically configured to:

[0187] perform semantic granularity division on all the answer analysis information to obtain first, second, and third granularity semantic features, the semantic granularity of the first, second, and third granularity semantic features decreasing in order.

[0188] perform feature fusion on the first, second, and third granularity semantic features to obtain a fusion feature.

[0189] perform a decoding operation on the fusion feature to obtain the answer key point set and the answer key point sequence.

[0190] In some embodiments, the determination module 502 is specifically configured to:

[0191] splice the second and third granularity semantic features to obtain spliced semantic features.

[0192] determine a cross-level association weight between the first granularity semantic feature and the spliced semantic features.

[0193] perform semantic enhancement on the spliced semantic features based on the cross-level association weight to obtain enhanced spliced semantic features.

[0194] perform feature fusion on the enhanced spliced semantic features and the first granularity semantic features to obtain a fusion feature.

[0195] In some embodiments, the determination module 502 is specifically configured to:

[0196] ​perform convolution operation on the fusion features to obtain convolution fusion features;

[0197] perform screening on the fusion features based on the convolution fusion features to obtain screened features;

[0198] perform decoding operation based on the screened features to obtain an answer key point set and an answer key point sequence.

[0199] In some embodiments, the fusion module 503 is specifically configured to:

[0200] perform context fusion on the multi-round question analysis information to obtain global context features;

[0201] perform multi-round context association on the global context features to obtain global enhanced context information.

[0202] In some embodiments, the fusion module 503 is specifically configured to:

[0203] perform time sequence convolution on the current round question analysis information in the multi-round question analysis information and all the round question analysis information before the current round question analysis information to obtain a current round time sequence convolution feature, the current round question analysis information being any round question analysis information in the multi-round question analysis information;

[0204] perform context fusion on the current round time sequence convolution feature and all the round time sequence convolution features before the current round time sequence convolution feature to obtain a current round context feature, the global context features including context features of all the rounds.

[0205] In some embodiments, the fusion module 503 is specifically configured to:

[0206] perform association degree evaluation on the current round context feature and each round context feature before the current round context feature to obtain a feature association weight between the current round context feature and each round context feature before the current round context feature;

[0207] weight all the round context features before the current round context feature based on each feature association weight to obtain weighted context features;

[0208] fuse the weighted context features with the current round context feature to obtain an enhanced context feature of the current round, the global enhanced context information including enhanced context features of all the rounds.

[0209] In some embodiments, the fusion module 503 is further configured to: perform decoding operation on the global enhanced context information to generate thinking intention description information.

[0210] In some embodiments, the information generation module 504 is specifically configured to:

[0211] determine semantic similarity between the answer key point set and the thinking intention description information;

[0212] generate semantic association matching features based on the semantic similarity and the answer key point set;

[0213] perform decoding operations on the semantic association matching features, the answer key point set, and the answer key point sequence to obtain the guided questioning information;

[0214] perform decoding operations on the guided questioning information, the answer key point set, and the answer key point sequence to obtain answer information corresponding to the guided questioning information.

[0215] It should be noted that each of the above modules can be a functional module or a program module, and can be implemented by software or hardware. For the modules implemented by hardware, each of the above modules can be located in the same processor; or each of the above modules can also be located in different processors in any combination.

[0216] In the embodiment, an electronic device is also provided, including a memory and a processor, the memory stores a computer program, and the processor is configured to execute the computer program to perform the steps in any of the above method embodiments.

[0217] Optionally, the electronic device can further include a transmission device and an input / output device, wherein the transmission device is connected with the processor, and the input / output device is connected with the processor.

[0218] Optionally, in the embodiment, the processor can be configured to execute the steps in any of the above method embodiments through the computer program.

[0219] It should be noted that the specific examples in the embodiment can refer to the examples described in the above embodiments and optional implementation manners, which will not be described herein.

[0220] In the description of the embodiments of the present application, it should be understood that the terms "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "center", "top", "bottom", "top", "bottom", "inside", "outside", "inside", "outside", etc. The orientation or positional relationship indicated is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present application and simplifying the description, and does not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore it cannot be understood as a limitation to the present application. Among them, "inside" refers to the inside or enclosed area or space. "Periphery" refers to the area around a particular component or a particular area.

[0221] In the description of the embodiments of the present application, the terms "first", "second", "third", "fourth" are used only to describe purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with "first", "second", "third", "fourth" can be explicitly or implicitly included one or more of the features. In the description of the present application, unless otherwise specified, the meaning of "a plurality of" is two or more.

[0222] In the description of the embodiments of the present application, it should be noted that, unless otherwise explicitly specified and limited, the terms "mounting", "connecting", "connecting", "assembling" should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be directly connected, or indirectly connected through an intermediate medium, or it can be the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0223] In the description of the embodiments of the present application, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0224] In the description of the embodiments of the present application, it should be understood that "-" and "~" represent the range between two values, and the range includes the endpoints. For example: "A-B" represents a range greater than or equal to A and less than or equal to B. "A~B" represents a range greater than or equal to A and less than or equal to B.

[0225] In the description of the embodiments of the present application, the term "and / or" herein is only a description of the association relationship between the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / " herein generally represents that the front and rear associated objects are a "or" relationship.

[0226] Although the embodiments of the present application have been shown and described, those skilled in the art can understand that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. An intelligent answering method, characterized by, The method comprises the following steps: obtaining multi-round question resolution information of a user and answer resolution information corresponding to each round of question resolution information; determining an answer key point set and a corresponding answer key point sequence based on all the answer resolution information; fusing the multi-round question resolution information to generate global enhanced context information; The method of fusing the multi-round question resolution information to generate global enhanced context information comprises the following steps: context fusion is performed on the multi-round question resolution information to obtain global context features; multi-round context association is performed on the global context features to obtain the global enhanced context information; After the multi-round context association on the global context features to obtain the global enhanced context information, the method further comprises the following steps: decoding the global enhanced context information to generate thinking intent description information; Based on the answer key point set, the answer key point sequence and the global enhanced context information, guide type question answering information is generated, and the guide type question answering information comprises guide type question information and answer information corresponding to the guide type question information; The method of generating guide type question answering information based on the answer key point, the answer key point sequence and the global enhanced context information comprises the following steps: determining the semantic similarity between the answer key point set and the thinking intent description information; generating semantic association matching features based on the semantic similarity and the answer key point set; decoding the semantic association matching features, the answer key point set and the answer key point sequence to obtain the guide type question information; decoding the guide type question information, the answer key point set and the answer key point sequence to obtain the answer information corresponding to the guide type question information.

2. The intelligent answering method of claim 1, wherein, The method of determining an answer key point set and a corresponding answer key point sequence based on all the answer resolution information comprises the following steps: performing semantic granularity division on all the answer resolution information to obtain first granularity semantic features, second granularity semantic features and third granularity semantic features, wherein the semantic granularity of the first granularity semantic features, the second granularity semantic features and the third granularity semantic features decreases gradually; performing feature fusion on the first granularity semantic features, the second granularity semantic features and the third granularity semantic features to obtain fusion features; performing decoding operation on the fusion features to obtain the answer key point set and the answer key point sequence. 3.The intelligent answering method of claim 2, wherein, The method of performing feature fusion on the first granularity semantic features, the second granularity semantic features and the third granularity semantic features to obtain fusion features comprises the following steps: splicing the second granularity semantic features and the third granularity semantic features to obtain spliced semantic features; determining the cross-level association weight between the first granularity semantic features and the spliced semantic features; performing semantic enhancement on the spliced semantic features based on the cross-level association weight to obtain enhanced spliced semantic features; performing feature fusion on the enhanced spliced semantic features and the first granularity semantic features to obtain the fusion features. 4.The intelligent answering method of claim 2, wherein, The decoding operation on the fusion feature obtains the answer key point set and the answer key point sequence, and includes: performing convolution operation on the fusion feature to obtain convolution fusion feature; performing screening on the fusion feature based on the convolution fusion feature to obtain screened feature; performing decoding operation based on the screened feature to obtain the answer key point set and the answer key point sequence.

5. The intelligent answering method of claim 1, wherein, The context fusion of the multiple rounds of question analysis information obtains the global context feature, and includes: performing time sequence convolution on the current round question analysis information in the multiple rounds of question analysis information and all the question analysis information before the current round question analysis information to obtain current round time sequence convolution feature, the current round question analysis information being any round question analysis information in the multiple rounds of question analysis information; performing context fusion on the current round time sequence convolution feature and all the round time sequence convolution features before the current round time sequence convolution feature to obtain current round context feature, the global context feature including all the round context features. 6.The intelligent answering method of claim 5, wherein, The multi-round context association on the global context feature obtains the global enhanced context information, and includes: performing association degree evaluation on the current round context feature and each round context feature before the current round context feature to obtain feature association weight between the current round context feature and each round context feature before the current round context feature; weighting all the round context features before the current round context feature based on each feature association weight to obtain weighted context feature; performing fusion on the weighted context feature and the current round context feature to obtain enhanced context feature of the current round; the global enhanced context information includes enhanced context features of all the rounds.

7. An intelligent answering device, characterized in that It includes: The acquisition module is used for acquiring the multi-round question analysis information of the user and the answer analysis information corresponding to each round question analysis information. The determination module is used for determining the answer key point set and the corresponding answer key point sequence based on all the answer analysis information. The fusion module is used for fusing the multiple rounds of question analysis information to generate global enhanced context information. The fusion module is specifically used for: performing context fusion on the multiple rounds of question analysis information to obtain global context feature. The fusion module is specifically used for: performing multi-round context association on the global context feature to obtain the global enhanced context information. The information generation module is used for generating guided answering information based on the answer key point set, the answer key point sequence and the global enhanced context information, the guided answering information including guided questioning information and answer information corresponding to the guided questioning information. The information generation module is specifically used for: determining semantic similarity between the answer key point set and the thinking intention description information. The information generation module is specifically used for: generating semantic association matching feature based on the semantic similarity and the answer key point set. The information generation module is specifically used for: performing decoding operation on the semantic association matching feature, the answer key point set and the answer key point sequence to obtain the guided questioning information. ​ The decoding operation is performed on the guided questioning information, the answer key point set and the answer key point sequence to obtain answer information corresponding to the guided questioning information.

Images