Intelligent question answering method and device

Abstract

The invention relates to the technical field of natural language processing, in particular to an intelligent question answering method and device. The intelligent question answering method comprises the steps of obtaining multiple rounds of 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 answer analysis information; fusing the multi-round problem analysis information to generate global enhanced context information; and generating guided question answering information based on the answer key point set, the answer key point sequence and the global enhancement context information. The thinking-guided question answering method can be provided for students, the independent thinking ability of the students is cultivated, and the comprehensiveness of knowledge point mastering of the students is improved.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to an intelligent question-answering method and apparatus. Background Technology

[0002] With the continuous development of science and technology, intelligent Q&A technology has brought great convenience to students, parents, and teachers. Currently, the Q&A method typically involves students inputting questions into an electronic device, which then uses semantic recognition to identify keywords in the question and directly matches the answer based on those keywords.

[0003] Current Q&A methods only provide direct answers to a single question, failing to provide students with coherent thinking guidance based on the logical relationships between knowledge points in their questions. They also fail to transform students' hesitations and fragmented information into knowledge bottlenecks in the learning process, making it difficult to uncover students' potential needs during the questioning process.

[0004] Therefore, providing a way to answer questions in order to guide students to think coherently is a problem that needs to be solved. Summary of the Invention

[0005] To address the aforementioned problems in the prior art, this application provides an intelligent question-answering method, apparatus, electronic device, and storage medium to solve the problem that question-answering methods in related technologies cannot provide students with coherent thinking guidance.

[0006] Firstly, this embodiment provides an intelligent question-answering method, including: Obtain the user's multi-round question analysis information and the corresponding answer analysis information for each round of question analysis; Based on all the answer analysis information, determine the set of key points for the answer and the corresponding sequence of key points for the answer; By fusing the problem analysis information from multiple rounds, global enhanced contextual information is generated; Guided Q&A information is generated based on the set of key answer points, the sequence of key answer points, and the global enhanced contextual information. The guided Q&A information includes guided question information and the answer information corresponding to the guided question information.

[0007] In some embodiments, determining the set of answer key points and the corresponding sequence of answer key points based on all answer parsing information includes: All answer parsing information is semantically divided 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 decreases progressively. The first granularity semantic feature, the second granularity semantic feature, and the third granularity semantic feature are fused to obtain the fused feature; The fused features are decoded to obtain the set of answer key points and the sequence of answer key points.

[0008] In some embodiments, the feature fusion of the first granularity semantic feature, the second granularity semantic feature, and the third granularity semantic feature to obtain fused features includes: The second granularity semantic feature is concatenated with the third granularity semantic feature to obtain the concatenated semantic feature; Determine the cross-level association weight between the first granular semantic feature and the concatenated semantic feature; Based on the cross-level association weights, the spliced ​​semantic features are semantically enhanced to obtain enhanced spliced ​​semantic features. The enhanced splicing semantic features are fused with the first granular semantic features to obtain the fused features.

[0009] In some embodiments, the decoding operation on the fused features to obtain the set of answer key points and the sequence of answer key points includes: Perform a convolution operation on the fused features to obtain convolutional fused features; The fused features are filtered based on the convolutional fusion features to obtain the filtered features; Based on the filtered features, a decoding operation is performed to obtain the set of answer key points and the sequence of answer key points.

[0010] In some embodiments, fusing the multi-round question parsing information to generate globally enhanced contextual information includes: Contextual fusion is performed on the problem parsing information from multiple rounds to obtain global contextual features; The global context features are subjected to multiple rounds of contextual association to obtain the global enhanced contextual information.

[0011] In some embodiments, the contextual fusion of the multiple rounds of question parsing information to obtain global contextual features includes: The current round problem parsing information in the multi-round problem parsing information is temporally convolved with the problem parsing information of all rounds preceding the current round problem parsing information to obtain the current round temporal convolution feature, where the current round problem parsing information is any round problem parsing information in the multi-round problem parsing information; The current round temporal convolutional feature is fused with the temporal convolutional features of all rounds preceding the current round temporal convolutional feature to obtain the current round contextual feature. The global contextual feature includes the contextual features of all rounds.

[0012] In some embodiments, the step of performing multi-round contextual association on the global contextual features to obtain the global enhanced contextual information includes: The correlation between the current round context feature and each round context feature preceding the current round context feature is evaluated to obtain the feature correlation weight between the current round context feature and each round context feature preceding the current round context feature; Based on the association weight of each feature, all round context features preceding the current round context feature are weighted to obtain a weighted context feature; The weighted context features are fused with the current round context features to obtain the enhanced context features for the current round; The global augmented context information includes augmented context features for all rounds.

[0013] In some embodiments, after performing multi-round contextual association on the global contextual features to obtain the global enhanced contextual information, the method further includes: The global enhanced contextual information is decoded to generate a description of the thought intent.

[0014] In some embodiments, generating guided question-and-answer information based on the answer key points, the answer key point sequence, and the globally enhanced contextual information includes: Determine the semantic similarity between the set of key points in the answer and the information describing the thinking intent; Based on the semantic similarity and the set of key points of the answer, semantic association matching features are generated; The semantic association matching features, the set of answer key points, and the sequence of answer key points are decoded to obtain the guided question information. The guided question information, the set of key answer points, and the sequence of key answer points are decoded to obtain the answer information corresponding to the guided question information.

[0015] Secondly, this embodiment provides an intelligent question-answering device, including: The acquisition module is used to acquire the user's multi-round question analysis information and the corresponding answer analysis information for each round of question analysis information; The determination module is used to determine the set of key points for the answer and the corresponding sequence of key points for the answer based on all answer parsing information. The fusion module is used to fuse the problem parsing information from multiple rounds to generate global enhanced contextual information; The information generation module is used to generate guided Q&A information based on the set of key answer points, the sequence of key answer points, and the global enhanced contextual information. The guided Q&A information includes guided question information and answer information corresponding to the guided question information.

[0016] Thirdly, this embodiment provides an electronic device including 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 described in the first aspect and any embodiment above.

[0017] Fourthly, this embodiment provides a storage medium storing a computer program that, when executed by a processor, implements the intelligent question-answering method described in the first aspect and any embodiment.

[0018] Compared with related technologies, the intelligent Q&A method provided in this application determines the set of key answer points and the corresponding sequence of key answer points from the user's multi-round answer analysis information, rather than extracting key answer points from a single question. This avoids the omission of key information and ensures the completeness of the answer and the logical order of key points. Furthermore, the multi-round question analysis information is fused, making it easier to understand the global question from the fused information and to obtain the user's potential needs and knowledge gaps from the global question, providing support for accurate Q&A in the subsequent question-and-answer process. Finally, the set of key answer points, the sequence of key answer points, and global enhanced contextual information are combined to provide guided questions based on the logical order of key points, thus providing students with coherent thinking guidance. Moreover, the global enhanced contextual information can extract students' knowledge gaps and potential needs from fragmented information in their expressions, improving the comprehensiveness of the guided Q&A content. By providing answers to students through the set of all key question points, the comprehensiveness of students' knowledge mastery is improved.

[0019] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects and advantages of this application more readily apparent. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a hardware structure block diagram of a terminal for an intelligent question-answering method provided in an embodiment of this application; Figure 2This is a flowchart of an intelligent question-answering method provided in an embodiment of this application; Figure 3 This is a flowchart illustrating the generation of fusion features provided in an embodiment of this application; Figure 4 This is a flowchart illustrating the generation of globally enhanced contextual information, provided in an embodiment of this application. Figure 5 This is a structural block diagram of an intelligent question-answering device provided in an embodiment of this application. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] The method embodiments provided in this example can be executed on a terminal, computer, or similar computing device. For example, it can run on a terminal. Figure 1 This is a hardware structure block diagram of a terminal for an intelligent question-answering method provided in an embodiment of this application. For example... Figure 1 As shown, a terminal may include one or more ( Figure 1 Only one is shown in the diagram. A processor 102 and a memory 104 for storing data are also included. The processor 102 may be, but is not limited to, a microprocessor (MCU) or a field-programmable gate array (FPGA). The terminal may also include a transmission device 106 for communication functions and an input / output device 108. Those skilled in the art will understand that… Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the terminal described above. For example, the terminal may also include components that are larger than... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown are illustrated.

[0023] The memory 104 can be used to store computer programs, such as application software programs and modules, like the computer program corresponding to the intelligent question-answering method in this embodiment. The processor 102 executes various functional applications and data processing by running the computer programs stored in the memory 104, thereby implementing the aforementioned method. The memory 104 may include high-speed random access memory and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some embodiments, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the terminal via a network. Embodiments of the aforementioned network include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0024] The transmission device 106 is used to receive or send data via a network. This network includes a wireless network provided by the terminal's communication provider. In one embodiment, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another embodiment, the transmission device 106 can be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0025] With the continuous development of science and technology, intelligent Q&A technology has brought great convenience to students, parents and teachers.

[0026] The current Q&A method usually involves students inputting questions into an electronic device, which then uses semantic recognition to identify keywords in the question and directly matches the answer based on those keywords.

[0027] Current Q&A methods only provide direct answers to a single question, failing to provide students with coherent thinking guidance based on the logical relationships between knowledge points in their questions. They also fail to transform students' hesitations and fragmented information into knowledge bottlenecks in the learning process, making it difficult to uncover students' potential needs during the questioning process.

[0028] Therefore, in intelligent Q&A scenarios, how to provide a Q&A method to guide students to think coherently is a problem that needs to be solved.

[0029] This embodiment provides an intelligent question-answering method. Figure 2 This is a flowchart of an intelligent question-answering method provided in an embodiment of this application. The executing entity of this method can be an electronic device. Optionally, the electronic device can be a server or a terminal device, but this application is not limited thereto. Specifically, as shown... Figure 2As shown, the process includes the following steps: Step S201: Obtain the user's multi-round question analysis information and the answer analysis information corresponding to each round of question analysis information.

[0030] For example, in an intelligent Q&A scenario, users typically input the questions they need to answer into an electronic device via voice, text, or handwriting. The electronic device can then provide the corresponding answers using conventional intelligent algorithms. The user is usually a student, and during this process, because the student's knowledge of the knowledge points is not comprehensive, there may be more than one question that needs to be answered. Therefore, the student will input the questions that need to be answered in multiple rounds, and the electronic device will output multiple corresponding answers.

[0031] Furthermore, the electronic device can collect multi-round question information input by students, as well as answer information output by the electronic device for each question. It then parses the collected question and answer information, whereby the answer information may include problem-solving strategies for the corresponding question. The collected question and answer information can then be parsed separately to obtain multi-round question parsing information and answer parsing information for each round of question parsing. Specifically, the collected question and answer information can be text data, which is then parsed using methods such as word segmentation and format standardization to obtain the user's multi-round question parsing information and answer parsing information for each round of question parsing.

[0032] Step S202: Determine the set of key points for the answer and the corresponding sequence of key points for the answer based on all answer parsing information.

[0033] For example, key points are extracted from all answer parsing information to obtain a set of key points and the corresponding key point sequence.

[0034] Specifically, the core terms are extracted from all answer analysis information, such as concepts, data, proper nouns, and logical connectors. Further, duplicate words are removed from all keywords in the answer analysis information, semantically repetitive words are merged, and words containing related terms are filtered. The extracted and deduplicated keywords are then organized into a static set without repetition or omission according to "categorization logic," ultimately yielding the set of key answer points.

[0035] Furthermore, the keywords in the set of key points of the answer are sorted according to logical deduction or cognitive habits to obtain the sequence of key points of the answer, thus ensuring the completeness, accuracy and logic of the key points of the answer.

[0036] Specifically, taking the math problem of "finding the area of ​​a triangle" as an example, the key points of the answer obtained through the above method include: "A right triangle, the longest side is the hypotenuse, the legs include a=4, b=5, the formula for the area of ​​a triangle is S=½×a×b, substituting the data, S=10." Further, following the logical deduction order, the following sequence of key points can be determined: Key point sequence 1: Determine the triangle type: a right triangle, the longest side is the hypotenuse, its legs include a=4, b=5; Key point sequence 2: Recall the formula for the area of ​​a right triangle: S=½×a×b; Key point sequence 3: Substitute the data: a=4, b=5, to get S=½×4×5; Key point sequence 4: Draw the conclusion: S=10.

[0037] Step S203: The information from multiple rounds of question analysis is fused to generate global enhanced contextual information.

[0038] For example, core attributes in each round of question parsing information can be labeled, and the correlation between multiple rounds of question parsing information can be determined based on the core attributes. The multiple rounds of question parsing information can then be fused based on the correlation between them to obtain global enhanced contextual information.

[0039] Specifically, for the first round question, "What is the Pythagorean theorem?", the core attribute in the question analysis information is: the definition of the Pythagorean theorem is a² + b² = c², the relationship between the hypotenuse and the legs of a right triangle; for the second round question, "Use the Pythagorean theorem to find the leg a, (b=3, c=5)", the core attribute in the question analysis information is: substituting into the formula a² = 5² - 3², the calculation steps are a² = 16, a = 4; when the user adds in the second round question, "I need to confirm whether c is the hypotenuse first", the core attribute in the corresponding question analysis information is: the user has doubts about "hypotenuse determination", and this detail needs to be strengthened during integration; for the third round question, "If it is an obtuse triangle, can the Pythagorean theorem still be used?", the core attribute in the corresponding question analysis information is: the user extends from "right triangle" to "obtuse triangle", and the scope of application of the theorem needs to be supplemented.

[0040] Further determine the correlation between the core attributes: the definition, applicable conditions and scope of the Pythagorean theorem, and explore potential user needs based on this correlation. If c is not the hypotenuse, it is an obtuse triangle, and the corresponding cosine theorem for an obtuse triangle can be added (c²=a²+b²-2abcosC, where C is an obtuse angle).

[0041] Finally, based on this correlation, the information from multiple rounds of problem analysis will be integrated. Specifically, the globally enhanced contextual information obtained is as follows: the Pythagorean theorem formula (a² + b² = c²) is applied to right triangles. The longest side in a right triangle is the hypotenuse. An example of calculating the side length of a right triangle is given (a = 4, b = 3, c = 5). The Pythagorean theorem cannot be directly applied to obtuse triangles. Obtuse triangles require the use of the cosine theorem (c² = a² + b² - 2abcosC, where C is an obtuse angle).

[0042] Step S204: Generate guided Q&A information based on the set of key answer points, the sequence of key answer points, and global enhanced contextual information.

[0043] The guided Q&A information includes guided question information and the corresponding answer information.

[0044] For example, the set of key answer points and the sequence of key answer points are combined with globally enhanced contextual information to guide users to think about the key points sequentially through questioning. Specifically, the electronic device can output guided question information: You can first recall the formula of the Pythagorean theorem (a² + b² = c²), where a and b are the legs and c is the hypotenuse. What problems 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 the square of the hypotenuse in the Pythagorean theorem is equal to the sum of the squares of the two legs (c² = a² + b²). If the legs are substituted as hypotenuses, the calculation result will be incorrect; the basis for determining the hypotenuse is that "the longest side in a right triangle is the hypotenuse". Further thinking prompts are output: If c is not the hypotenuse, how should the side length c be calculated? After the user answers the thinking prompt or after a preset time, the corresponding answer is given: Apply the Law of Cosines (c² = a² + b² - 2abcosC, where C is an obtuse angle) to solve for the side length c. This approach transforms users from "passively receiving" to "actively understanding," while also enhancing their thinking abilities by combining logical information and potential needs within the overall context.

[0045] In the above implementation process, the key points set and sequence of answers are determined from the user's multi-round answer analysis information, rather than extracting key points from a single question. This avoids the omission of key information and ensures the completeness of the answer and the logical order of key points. Furthermore, the multi-round question analysis information is fused to facilitate the understanding of the global question. The global question contains key information from all questions, allowing for the identification of the user's potential needs and knowledge gaps, providing support for accurate answers in subsequent question-and-answer sessions. Finally, based on the combination of the key point set, key point sequence, and global enhanced contextual information, guiding questions are given according to the logical order of key points, providing users with coherent thinking guidance. Moreover, the global enhanced contextual information allows for the extraction of the user's knowledge gaps and potential needs from fragmented information in the user's expression, further enabling answers to the user's questions based on the set of key points from all questions, improving the comprehensiveness of the user's knowledge mastery.

[0046] In some embodiments, determining the set of answer key points and the corresponding sequence of answer key points based on all answer parsing information may include the following steps: Step 1: Divide all answer parsing information into semantic granularities to obtain first-granularity semantic features, second-granularity semantic features, and third-granularity semantic features. The semantic granularity of the first-granularity semantic features, second-granularity semantic features, and third-granularity semantic features decreases progressively.

[0047] For example, all answer parsing information is abstracted into three levels based on semantic granularity: first-granularity semantic features, second-granularity semantic features, and third-granularity semantic features. The semantic granularity of these three levels decreases progressively. In other words, three different levels of abstraction are obtained: the first-granularity semantic feature is a high-level abstraction feature, the second-granularity semantic feature is a medium-level abstraction feature, and the third-granularity semantic feature is a low-level abstraction feature.

[0048] Specifically, analyzing the original text: "To find the area of ​​a right triangle, you need to first understand the formula, then substitute the data of the right-angled sides to calculate, and finally get the area as 6", we can divide all the answer analysis information according to semantic granularity and get the first granularity semantic feature: calculate the area of ​​a right triangle by substituting the formula.

[0049] The first-level semantic feature is broken down into two second-level semantic features: 1. Determine the formula for the area of ​​a right triangle (S=½×a×b); 2. Substitute the right-angled side data (a=3, b=4) to calculate the area.

[0050] The second-level semantic features are decomposed into three third-level semantic features: 1. Area formula parameter definition (a and b are the right-angled sides, S is the area); 2. Known data (a=3, b=4); 3. Calculation process (½×3×4=6).

[0051] Step 2: Perform feature fusion on the first-granularity semantic features, the second-granularity semantic features, and the third-granularity semantic features to obtain fused features.

[0052] Furthermore, the semantic features at three different granularities are transformed into vector forms, and the three vectors are fused using an attention mechanism. Specifically, the BERT model can be used to encode the three semantic features at different granularities into first-granularity word vectors, second-granularity word vectors, and third-granularity word vectors. Then, an attention mechanism is used to assign corresponding weights to the word vectors at each of the three granularities. For example, the first-granularity word vector is assigned the highest attention weight (e.g., 0.4) because it represents the global framework; the second-granularity word vector is assigned a medium weight (e.g., 0.35) because it represents the logical core; and the third-granularity word vector is assigned a low weight (e.g., 0.25) because it provides detailed support. In addition, to preserve the correlation between levels, the local weights of "key correlation information" (e.g., the correspondence between formulas and data) can be increased. Finally, through weighted summation and residual concatenation, the word vectors at the three granularities are fused into a "fused feature vector," and key correlation information is added. That is, the fused feature is: first-granularity word vector × 0.4 + second-granularity word vector × 0.35 + third-granularity word vector × 0.25 + increment of key correlation information.

[0053] Step 3: Decode the fused features to obtain the set of answer key points and the sequence of answer key points.

[0054] 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.

[0055] 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.

[0056] 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).

[0057] 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.

[0058] 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: Step S301: Concatenate the second-granularity semantic features with the third-granularity semantic features to obtain concatenated semantic features.

[0059] For example, semantic granularity is divided for all answer parsing information, specifically as follows: in, X For all answer analysis information, H To transform answer parsing information into complete semantic features, BiLSTM, a bidirectional long short-term memory network, is used. These are the first-granularity semantic features, the second-granularity semantic features, and the third-granularity semantic features, respectively. The dimension selection matrices for the first, second, and third granular semantic features are respectively, all of which are diagonal matrices with diagonal elements of 0 or 1, and satisfy the following conditions: , where I is the identity matrix.

[0060] Furthermore, the second-granularity semantic features are concatenated with the third-granularity semantic features using the concatenation function Concat to obtain the concatenated semantic features. .

[0061] Step S302: Determine the cross-level association weights between the first-granularity semantic features and the concatenated semantic features.

[0062] The attention mechanism is used to determine the cross-level association weights between the first-granularity semantic features and the concatenated semantic features. A ,Right now ,in, Attention This is an attention mechanism.

[0063] Step S303: Perform semantic enhancement on the concatenated semantic features based on cross-level association weights to obtain enhanced concatenated semantic features.

[0064] Furthermore, based on the value projection matrix Furthermore, cross-level association weights are used to semantically enhance the concatenated semantic features, resulting in enhanced concatenated semantic features. As an example, enhancing the semantic features of splicing. It can be determined by the following expression: Step S304: The enhanced splicing semantic features are fused with the first-granularity semantic features to obtain fused features.

[0065] Furthermore, based on the structured fusion matrix, the enhanced concatenated semantic features are fused with the first-granularity semantic features to obtain the fused features. Specifically, if the structured fusion matrix is... Then the fusion features It can be determined by the following expression: In the above implementation process, the second-granularity semantic features and the third-granularity semantic features are concatenated to obtain concatenated semantic features. An attention mechanism is used to determine the cross-level association weight between the first-granularity semantic features and the concatenated semantic features. By using the cross-level association weight, key information closely related to the concatenated semantic features and the first-granularity semantic features can be identified. Then, the concatenated semantic features are semantically enhanced according to the cross-level association weight to obtain enhanced concatenated semantic features. Finally, the enhanced concatenated semantic features and the first-granularity semantic features are fused to obtain fused features. This feature fusion process is not an indiscriminate feature concatenation fusion, but a targeted fusion based on semantic relevance. It preserves the semantic relevance of each level, reduces redundant information, and avoids the problem-solving logic break caused by isolated hierarchical features.

[0066] In some embodiments, decoding the fused features to obtain a set of answer key points and a sequence of answer key points may include the following steps: Step 1: Perform a convolution operation on the fused features to obtain convolutional fused features.

[0067] For example, convolutional fusion features can be obtained by performing convolution operations on the fused features using convolutional layers. Specifically, convolutional fusion features... Z It can be determined by the following expression: Wherein, Conv1D is a one-dimensional convolutional layer. The kernel size is 3, and the sigmoid function is used. It should be noted that this embodiment uses a one-dimensional convolutional layer with a kernel size of 3 as an example. In practical applications, the convolutional layer can also be a two-dimensional or three-dimensional convolutional layer, and the kernel size can be 4 or 5. The convolutional layer dimension and kernel size can be adaptively adjusted according to the actual data, and are not limited here.

[0068] Step 2: Filter the fused features based on the convolutional fusion features to obtain the filtered features.

[0069] Step 3: Decode the filtered features to obtain the set of answer key points and the sequence of answer key points.

[0070] Furthermore, the fused features are filtered through convolutional fusion to obtain filtered features, and the filtered features are decoded to obtain the set of answer key points and the corresponding answer key point sequence.

[0071] Specifically, the set of key points for the answer. Y And the corresponding sequence of key points in the answer. Seq It can be determined by the following expression: in, Y For the set of key points of the answer, Seq For the sequence of key points in the answer, Decoder For Transformer block decoder operations, It is a multilayer perceptron.

[0072] In the above implementation process, convolution operation is performed on the fused features to obtain convolutional fused features, and the fused features are then filtered based on the convolutional fused features to obtain filtered features. Redundant semantic information is filtered through feature filtering. Finally, the set of answer key points and the corresponding answer key point sequence are generated through decoding operation, thereby realizing accurate processing of the entire process from feature extraction to result generation, and completely solving the three major problems of hierarchical extraction, association preservation, and ordered generation in the scenario.

[0073] In some embodiments, fusing multi-turn question resolution information to generate globally enhanced contextual information may include the following steps: Step 1: Perform contextual fusion on the multi-round question analysis information to obtain global contextual features.

[0074] For example, a NER model can be used to extract the core entities of the question parsing information, a relation extraction (RE) model can be used to identify the relationships between entities, and a pre-trained language model can be used to encode each round of questions into semantic vectors, thereby obtaining deep semantic information from the multi-round question information. The pre-trained language model can be a BERT model, a RoBERTa model, etc., and is not limited here.

[0075] Furthermore, an attention mechanism can be used to achieve dynamic weight fusion. Specifically, weights are assigned to the semantic vectors of each round based on the correlation between the question information in each round and the overall question. Then, the local weights of core entities in the semantic vectors are increased. Finally, all semantic vectors are weighted and summed to obtain global contextual features, thereby transforming fragmented single questions into structured and logical global contextual features, and realizing the integration of multi-round question information.

[0076] Step 2: Perform multi-round context association on the global context features to obtain global enhanced context information.

[0077] For example, correlation information is extracted from the global context features, and global enhanced context information is obtained based on the correlation information. Specifically, the correlation information may include logical information, demand information, and extended information. Logical information may be information representing causal or dependency relationships in the global context features; demand information may be information representing demands in the global context features; and extended information may be information representing potential user needs in the global context features. Further, the logical information, demand information, and extended information are correlated with the global context features to obtain the global enhanced context information.

[0078] In addition, text generation models such as GPT and T5 can be used to decode the globally enhanced contextual information into structured text, such as category lists and paragraphs, to ensure that the machine-generated contextual features are consistent with the format of manually compiled text, which is convenient for subsequent use.

[0079] In the above implementation process, contextual fusion of multi-round question analysis information is performed, which transforms fragmented single questions into structured and logical global contextual features. Multi-round contextual association is then performed on the global contextual features, so that the resulting global enhanced contextual information is associated with the logical thread of all question information, as well as the user's potential needs and extended knowledge, thereby improving the accuracy and completeness of the contextual information.

[0080] In some embodiments, contextual fusion of multi-round question parsing information to obtain global contextual features may include the following steps: Step 1: Perform temporal convolution on the current round of problem analysis information and all rounds of problem analysis information before the current round of problem analysis information to obtain the temporal convolution feature of the current round. The current round of problem analysis information is any round of problem analysis information in the multi-round problem analysis information.

[0081] For example, any round of problem parsing information from multiple rounds can be determined as the current round of problem parsing information. Temporal convolutional networks have long-term temporal modeling capabilities and can fully capture the temporal semantic dependencies between multiple rounds of problems. Therefore, the current round of problem parsing information can be temporally convolved with all rounds of problem parsing information before the current round of problem parsing information through a temporal convolutional network to obtain the temporal convolutional features of the current round.

[0082] Specifically, if the temporal convolution feature of the k-th round is Then the temporal convolution features of the kth round It can be determined by the following expression: TCN stands for Temporal Convolutional Network. These are the problem analysis information for rounds 1, 2, ..., k, respectively.

[0083] Step 2: Perform contextual fusion of the current round's temporal convolutional features with the temporal convolutional features of all rounds prior to the current round's temporal convolutional features to obtain the current round's contextual features. The global contextual features include the contextual features of all rounds.

[0084] Furthermore, the temporal convolutional features of the current round are concatenated with the temporal convolutional features of all rounds preceding the current round to obtain concatenated temporal convolutional features. Then, the concatenated convolutional features are fused with context through a context fusion matrix to obtain the context features of the current round.

[0085] Specifically, if the contextual features of the k-th round are Then the context features of the kth round It can be determined by the following expression: in, is the context fusion matrix, are the temporal convolutional features of the 1st, 2nd, …, k-th rounds respectively. Therefore, the context features of all rounds can be determined in this way.

[0086] In the above implementation process, temporal convolution is performed on the current problem parsing information and all previous rounds of problem parsing information, effectively capturing the temporal semantic dependency relationships of multi-round problems. The temporal convolutional features of each round of problems are extracted layer by layer from the problem parsing information data from the 1st round to the k-th round, and then the temporal convolutional features of each round are concatenated and processed through the context fusion matrix to obtain the context features covering the complete information of multi-round conversations, avoiding the defect of ignoring temporal associations in the traditional single-round feature extraction method, and providing a comprehensive context basis for subsequent accurate intention understanding.

[0087] In some of the embodiments, multi-round context association is performed on the global context features to obtain global enhanced context information. As an example, the global enhanced context information can be obtained through the process as Figure 4 shown, Figure 4 is a flowchart for generating global enhanced context information provided by an embodiment of the present application. Specifically, as Figure 4 shown, the process includes the following steps: Step S401: Evaluate the association degree between the context feature of the current round and each previous round of context features before the context feature of the current round, and obtain the feature association weights between the context feature of the current round and each previous round of context features before the context feature of the current round.

[0088] Exemplarily, the context feature of the current round and each previous round of context features before the context feature of the current round are concatenated to obtain the concatenated context feature, and then the association degree of the concatenated context feature is evaluated through a linear function and a non-linear activation function to obtain the feature association weights between the context feature of the current round and each previous round of context features before the context feature of the current round.

[0089] Specifically, if the context feature of the k-th round is , and any previous round of context feature before the context feature of the k-th round is , where i < k, then the feature association weight between and can be determined through the following expression: where Linear is the linear function and Sigmoid is the non-linear activation function. Therefore, the feature association weights between the context feature of the k-th round and each previous round of context features before the context feature of the k-th round can be determined in this way.

[0090] Step S402: Based on the association weight of each feature, weight all previous round context features to obtain weighted context features.

[0091] Step S403: Fuse the weighted context features with the current round context features to obtain the enhanced context features of the current round; the global enhanced context information includes the enhanced context features of all rounds.

[0092] Furthermore, based on the association weight of each feature, all previous rounds of context features are weighted and then fused with the current round of context features to obtain the enhanced context features of the current round.

[0093] 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: This method was used to further identify the enhanced contextual features of all rounds.

[0094] 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.

[0095] 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.

[0096] 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.

[0097] Specifically, the information describing the thinking intentions in round k. It can be determined by the following expression: in, DecoderThis 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.

[0098] 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.

[0099] In some embodiments, generating guided Q&A information based on answer key points, answer key point sequences, and globally enhanced contextual information may include the following steps: Step 1: Determine the semantic similarity between the set of key points in the answer and the information describing the thought process.

[0100] For example, the semantic similarity between the set of answer key points and the thinking intention description information in each round can be determined using a cosine similarity function. Specifically, the semantic similarity between the set of answer key points and the thinking intention description information in the k-th round... It can be determined by the following expression: Where Sim is the cosine similarity calculation, Embed is the embedding function, and the semantic similarity between the set of key points in the answer and the descriptive information of the thinking intention in the k-th round is... It can be expressed using matrices.

[0101] Step 2: Generate semantic association matching features based on semantic similarity and the set of key points of the answer.

[0102] Furthermore, semantic association matching features for each round are generated based on the product of the semantic similarity and the set of answer key points. Specifically, the semantic association matching features for the k-th round... It can be determined by the following expression: Step 3: Decode the semantic association matching features, the set of answer key points, and the sequence of answer key points to obtain guided question information.

[0103] Step 4: Decode the guided question information, the set of key answer points, and the sequence of key answer points to obtain the answer information corresponding to the guided question information.

[0104] As an example, guided question-and-answer information includes guided question information and corresponding answer information, while semantic association matching features include semantic association matching features from all rounds. Decoding the semantic association matching features from all rounds, the set of answer key points, and the sequence of answer key points yields the guided question information; decoding the guided question information, the set of answer key points, and the sequence of answer key points yields the answer information corresponding to the guided question information.

[0105] Specifically, semantic association matching features of all wheels The set of key answer points Y and the sequence of key answer points Seq are input into the decoder of the Transformer block to obtain guided questioning information, i.e., guided questioning information. Question It can be determined by the following expression: Furthermore, guiding questions will be used to convey information. Question The set of key answer points Y and the sequence of key answer points Seq are input into the decoder of the Transformer block to obtain the answer information corresponding to the guided question information, i.e., the guided question information. Question Corresponding answer information Answer It can be determined by the following expression: The Transformer block decoder decodes semantic association matching features, answer key point set, and answer key point sequence into structured text information, thereby obtaining guided question information and corresponding answer information.

[0106] In the above implementation process, the semantic similarity between the set of key answer points and the information describing the thinking intention is determined based on cosine similarity. Then, semantic association matching features are generated based on the semantic similarity and the set of key answer points. This allows the subsequently generated question-and-answer information to effectively integrate multi-round question information and corresponding answer information. The semantic association matching features, the set of key answer points, and the sequence of key answer points are decoded to generate text-based guided question information and corresponding answer information. The text information can effectively guide students to think independently.

[0107] It should be noted that the steps shown in the above process or in the flowchart of the accompanying figures 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 may be executed in a different order than that shown here.

[0108] This embodiment also provides an intelligent question-answering device, which is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. The terms "module," "unit," "subunit," etc., used below refer to combinations of software and / or hardware that implement a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0109] Figure 5 This is a structural block diagram of an intelligent question-answering device provided in an embodiment of this application, such as... Figure 5 As shown, the device includes: The acquisition module 501 is used to acquire the user's multi-round question analysis information and the answer analysis information corresponding to each round of question analysis information; Module 502 is used to determine the set of key points for the answer and the corresponding sequence of key points for the answer based on all answer parsing information. The fusion module 503 is used to fuse multi-round question parsing information to generate global enhanced contextual information; The information generation module 504 is used to generate guided Q&A information based on the set of key answer points, the sequence of key answer points, and global enhanced contextual information. The guided Q&A information includes guided question information and the answer information corresponding to the guided question information.

[0110] In some embodiments, the determining module 502 is specifically used for: The semantic granularity of all answer parsing information is divided into first-granularity semantic features, second-granularity semantic features and third-granularity semantic features. The semantic granularity of the first-granularity semantic features, second-granularity semantic features and third-granularity semantic features decreases step by step. The first-granularity semantic features, the second-granularity semantic features, and the third-granularity semantic features are fused to obtain the fused features; The fused features are decoded to obtain the set of answer key points and the sequence of answer key points.

[0111] In some embodiments, the determining module 502 is specifically used for: The second-granularity semantic features are concatenated with the third-granularity semantic features to obtain the concatenated semantic features; Determine the cross-level association weights between the first-granularity semantic features and the concatenated semantic features; Semantic enhancement is performed on the concatenated semantic features based on cross-level association weights to obtain enhanced concatenated semantic features. The enhanced splicing semantic features are fused with the first-granularity semantic features to obtain the fused features.

[0112] In some embodiments, the determining module 502 is specifically used for: Perform a convolution operation on the fused features to obtain convolutionally fused features; The fused features are filtered based on the convolutional fusion features to obtain the filtered features; Decoding is performed based on the filtered features to obtain the set of answer key points and the sequence of answer key points.

[0113] In some embodiments, the fusion module 503 is specifically used for: Contextual fusion is performed on multi-round question analysis information to obtain global contextual features; Multi-round contextual association is performed on global contextual features to obtain global enhanced contextual information.

[0114] In some embodiments, the fusion module 503 is specifically used for: The current round of problem analysis information in the multi-round problem analysis information is temporally convolved with the problem analysis information of all rounds before the current round of problem analysis information to obtain the temporal convolution feature of the current round. The current round of problem analysis information is any round of problem analysis information in the multi-round problem analysis information. The contextual features of the current round are fused with the contextual features of all rounds prior to the current round's temporal convolutional features to obtain the current round's contextual features. The global contextual features include the contextual features of all rounds.

[0115] In some embodiments, the fusion module 503 is specifically used for: The correlation between the current round context feature and each round context feature preceding the current round context feature is evaluated to obtain the feature correlation weight between the current round context feature and each round context feature preceding the current round context feature; Based on the association weight of each feature, all round context features before the current round context feature are weighted to obtain the weighted context feature; The weighted contextual features are fused with the contextual features of the current round to obtain the enhanced contextual features of the current round; the global enhanced contextual information includes the enhanced contextual features of all rounds.

[0116] In some embodiments, the fusion module 503 is also used to: decode the global enhanced contextual information to generate a description of the thought intent.

[0117] In some embodiments, the information generation module 504 is specifically used for: Determine the semantic similarity between the set of key points in the answer and the information describing the thought process; Semantic association matching features are generated based on semantic similarity and the set of key points in the answer; Decoding operations are performed on semantic association matching features, the set of key points for answers, and the sequence of key points for answers to obtain guided question information; The guided question information, the set of key points for the answer, and the sequence of key points for the answer are decoded to obtain the answer information corresponding to the guided question information.

[0118] It should be noted that the above modules can be functional modules or program modules, and can be implemented through software or hardware. For modules implemented through hardware, the above modules can reside in the same processor; or the above modules can be located in different processors in any combination.

[0119] This embodiment also provides an electronic device including a memory and a processor, the memory storing a computer program and the processor being configured to run the computer program to perform the steps in any of the above method embodiments.

[0120] Optionally, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor and the input / output device is connected to the processor.

[0121] Optionally, in this embodiment, the processor may be configured to execute the steps of any of the above method embodiments via a computer program.

[0122] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated in this embodiment.

[0123] In the description of the embodiments of the present invention, it should be understood that the terms "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "center," "top," "bottom," "top," "bottom," "inner," "outer," "inner side," and "outer side," etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the accompanying drawings and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. "Inner side" refers to the interior or enclosed area or space. "Outer perimeter" refers to the area surrounding a specific component or specific area.

[0124] In the description of embodiments of the present invention, the terms "first," "second," "third," and "fourth" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first," "second," "third," or "fourth" may explicitly or implicitly include one or more of that feature. In the description of the present invention, unless otherwise stated, "a plurality of" means two or more.

[0125] In the description of the embodiments of the present invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," "joining," and "assembly" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0126] In the description of embodiments of the present invention, specific features, structures, materials or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0127] In the description of the embodiments of the present invention, it should be understood that "-" and "~" represent a range between two numerical values, and this range includes the endpoints. For example, "AB" 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.

[0128] In the description of embodiments of the present invention, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0129] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent question-answering method, characterized in that, include: Obtain the user's multi-round question analysis information and the corresponding answer analysis information for each round of question analysis; Based on all the answer analysis information, determine the set of key points for the answer and the corresponding sequence of key points for the answer; By fusing the problem analysis information from multiple rounds, global enhanced contextual information is generated; Guided Q&A information is generated based on the set of key answer points, the sequence of key answer points, and the global enhanced contextual information. The guided Q&A information includes guided question information and the answer information corresponding to the guided question information.

2. The intelligent question-answering method according to claim 1, characterized in that, The process of determining the set of key answer points and the corresponding sequence of key answer points based on all answer parsing information includes: All answer parsing information is semantically divided 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 decreases progressively. The first granularity semantic feature, the second granularity semantic feature, and the third granularity semantic feature are fused to obtain the fused feature; The fused features are decoded to obtain the set of answer key points and the sequence of answer key points.

3. The intelligent question-answering method according to claim 2, characterized in that, The feature fusion of the first granularity semantic feature, the second granularity semantic feature, and the third granularity semantic feature to obtain the fused feature includes: The second granularity semantic feature is concatenated with the third granularity semantic feature to obtain the concatenated semantic feature; Determine the cross-level association weight between the first granular semantic feature and the concatenated semantic feature; Based on the cross-level association weights, the spliced ​​semantic features are semantically enhanced to obtain enhanced spliced ​​semantic features. The enhanced splicing semantic features are fused with the first granular semantic features to obtain the fused features.

4. The intelligent question-answering method according to claim 2, characterized in that, The decoding operation on the fused features to obtain the set of answer key points and the sequence of answer key points includes: Perform a convolution operation on the fused features to obtain convolutional fused features; The fused features are filtered based on the convolutional fusion features to obtain the filtered features; Based on the filtered features, a decoding operation is performed to obtain the set of answer key points and the sequence of answer key points.

5. The intelligent question-answering method according to claim 1, characterized in that, The process of fusing the problem parsing information from multiple rounds to generate globally enhanced contextual information includes: Contextual fusion is performed on the problem parsing information from multiple rounds to obtain global contextual features; The global context features are subjected to multiple rounds of contextual association to obtain the global enhanced contextual information.

6. The intelligent question-answering method according to claim 5, characterized in that, The contextual fusion of the problem parsing information from multiple rounds to obtain global contextual features includes: The current round problem parsing information in the multi-round problem parsing information is temporally convolved with the problem parsing information of all rounds preceding the current round problem parsing information to obtain the current round temporal convolution feature, where the current round problem parsing information is any round problem parsing information in the multi-round problem parsing information; The current round temporal convolutional feature is fused with the temporal convolutional features of all rounds preceding the current round temporal convolutional feature to obtain the current round contextual feature. The global contextual feature includes the contextual features of all rounds.

7. The intelligent question-answering method according to claim 6, characterized in that, The process of performing multi-round contextual association on the global contextual features to obtain the global enhanced contextual information includes: The correlation between the current round context feature and each round context feature preceding the current round context feature is evaluated to obtain the feature correlation weight between the current round context feature and each round context feature preceding the current round context feature; Based on the association weight of each feature, all round context features preceding the current round context feature are weighted to obtain a weighted context feature; The weighted context features are fused with the current round context features to obtain the enhanced context features for the current round; The global augmented context information includes augmented context features for all rounds.

8. The intelligent question-answering method according to claim 5, characterized in that, After performing multi-round contextual association on the global contextual features to obtain the global enhanced contextual information, the method further includes: The global enhanced contextual information is decoded to generate a description of the thought intent.

9. The intelligent question-answering method according to claim 8, characterized in that, The generation of guided Q&A information based on the answer key points, the answer key point sequence, and the globally enhanced contextual information includes: Determine the semantic similarity between the set of key points in the answer and the information describing the thinking intent; Based on the semantic similarity and the set of key points of the answer, semantic association matching features are generated; The semantic association matching features, the set of answer key points, and the sequence of answer key points are decoded to obtain the guided question information. The guided question information, the set of key answer points, and the sequence of key answer points are decoded to obtain the answer information corresponding to the guided question information.

10. An intelligent question-answering device, characterized in that, include: The acquisition module is used to acquire the user's multi-round question analysis information and the answer analysis information corresponding to each round of question analysis information; The determination module is used to determine the set of key points for the answer and the corresponding sequence of key points for the answer based on all answer parsing information. The fusion module is used to fuse the problem parsing information from multiple rounds to generate global enhanced contextual information; The information generation module is used to generate guided Q&A information based on the set of key answer points, the sequence of key answer points, and the global enhanced contextual information. The guided Q&A information includes guided question information and answer information corresponding to the guided question information.

Images