Intelligent question-answering method, device and computer equipment

Abstract

The present application involves big data technology and discloses a method for intelligent question answering, including: obtaining a question statement raised by a user; extracting the corresponding keywords in the question statement and obtaining the classification information of the keywords; judging whether a node matching the classification information of the keywords can be found on the knowledge tree according to the classification information of the keywords; if not, performing intent recognition on the question statement to obtain the focus information corresponding to the question statement; determining the node matching the question statement on the knowledge tree according to the focus information corresponding to the question statement; and performing intelligent question answering according to the node matching the question statement on the knowledge tree as the answer node. By establishing a knowledge tree with a simple structure, matching the corresponding node in the knowledge tree according to the question statement raised by the user as the answer node corresponding to the answer to the question statement, and replying the knowledge content corresponding to the answer node to the user, intelligent question answering in special fields with relatively small or extremely small data volumes can be achieved.

Description

Technical Field

[0001] The present application relates to the field of big data, and in particular to methods, devices and computer equipment for intelligent question answering. Background Art

[0002] The intelligent question-answering system can accept questions input by users in natural language, and return a relatively concise and accurate answer after steps such as question analysis, information retrieval and answer extraction. The intelligent question-answering system provides an option for developing enterprise problem-solving platforms, providing enterprises with automated services in various vertical business levels such as personnel, finance, marketing, and technology, which not only saves a lot of manpower and improves problem-solving efficiency, but also accumulates valuable knowledge assets for enterprises.

[0003] Existing intelligent question-answering systems are generally implemented by investing a lot of manpower and material resources to form a complex knowledge graph, or by training a deep neural network with a large amount of standard data. However, the inventors realized that for special fields with small or extremely small data volumes, intelligent question-answering cannot be achieved through knowledge graphs or deep neural networks. Summary of the invention

[0004] The main purpose of this application is to solve the technical problem that intelligent question answering cannot be achieved through knowledge graphs or deep neural networks in special fields where the amount of data is relatively small or extremely small.

[0005] This application proposes a method for intelligent question answering, including:

[0006] Get the question sentence raised by the user;

[0007] Extracting the keywords corresponding to the question sentence and obtaining classification information of the keywords;

[0008] Determine whether a node matching the classification information of the keyword can be found on the knowledge tree according to the classification information of the keyword;

[0009] If not, then perform intent recognition on the question sentence to obtain focus information corresponding to the question sentence;

[0010] Determining a node on the knowledge tree that matches the question statement according to the focus point information corresponding to the question statement;

[0011] Intelligent question and answer is performed based on the nodes on the knowledge tree that match the question statement as answer nodes.

[0012] Preferably, the step of extracting the corresponding keywords in the question sentence and obtaining classification information of the keywords includes:

[0013] After removing stop words from the question sentence, perform word segmentation to obtain a word segmentation combination corresponding to the question sentence;

[0014] Determining whether there is a designated word in the word segmentation combination that is different from the terminology rules of the preset keyword dictionary;

[0015] If so, the designated words are replaced one by one with standardized words corresponding to the designated words to obtain a modified word segmentation combination, wherein the standardized words are included in the keyword dictionary;

[0016] Extracting characteristic words corresponding to the specified part of speech in the modified participle combination as the keyword;

[0017] The classification information corresponding to the keyword is determined in the keyword dictionary.

[0018] Preferably, the step of determining whether a node matching the classification information of the keyword can be found on the knowledge tree according to the classification information of the keyword includes:

[0019] Obtaining keyword tables corresponding to each node of the knowledge tree and classification information corresponding to each keyword table, to obtain a classification information set corresponding to the knowledge tree;

[0020] Determining whether the classification information of the keyword is included in the classification information set;

[0021] If so, determining that according to the classification information of the keyword, a node matching the classification information of the keyword can be found on the knowledge tree;

[0022] According to the keyword search algorithm, a node matching the classification information of the keyword is determined.

[0023] Preferably, the step of determining the node matching the classification information of the keyword according to the keyword search algorithm includes:

[0024] Obtaining vocabulary weights corresponding to all vocabulary in a specified keyword table corresponding to a specified node, wherein the specified node is any one of all nodes in the knowledge tree;

[0025] Calculating a matching value between the designated node and the keyword according to the vocabulary weight;

[0026] According to the calculation process of the matching value between the designated node and the keyword, respectively calculate the matching value between all nodes in the knowledge tree and the keyword;

[0027] The node corresponding to the maximum matching value is used as the node matching the keyword.

[0028] Preferably, the step of performing intent recognition on the question sentence to obtain focus information corresponding to the question sentence includes:

[0029] Extracting semantic features of the question sentence;

[0030] Inputting the semantic features of the question sentence into a text classification neural network to obtain a coarse-grained information classification corresponding to the question sentence;

[0031] The focus point information corresponding to the question statement is determined according to the coarse-grained information classification.

[0032] Preferably, the step of performing intelligent question answering based on the node on the knowledge tree that matches the question statement as the answer node includes:

[0033] Determine whether the answer node has any associated child nodes;

[0034] If so, obtaining the knowledge content of the child node;

[0035] Using the knowledge content corresponding to the answer node as the answer to the question statement, and using the knowledge content of the sub-node as the knowledge guide in the intelligent question answering;

[0036] The answer to the question statement and the knowledge guidance are synchronously fed back to the terminal corresponding to the user.

[0037] Preferably, the step of obtaining the question statement raised by the user includes:

[0038] Get the input statement entered by the user;

[0039] Input the input sentence into the Bert model to obtain the prediction value corresponding to each word in the input sentence;

[0040] According to the prediction value corresponding to the designated character, determining whether the designated character is a wrong character, wherein the designated character is any one of all the characters in the input sentence;

[0041] If yes, obtain the pinyin corresponding to the specified character;

[0042] According to the pinyin corresponding to the designated character, determine the candidate character with the highest pinyin similarity as the correction character of the designated character;

[0043] After correcting all the typos in the input sentence, the question sentence raised by the user is obtained.

[0044] The present application also provides a device for intelligent question answering, comprising:

[0045] An acquisition module is used to acquire question statements raised by users;

[0046] An extraction module, used to extract the corresponding keywords in the question sentence and obtain classification information of the keywords;

[0047] A judgment module, used to judge whether a node matching the classification information of the keyword can be found on the knowledge tree according to the classification information of the keyword;

[0048] An intention recognition module, for performing intention recognition on the question statement to obtain focus information corresponding to the question statement if a node matching the classification information of the keyword cannot be found on the knowledge tree;

[0049] A determination module, configured to determine a node on the knowledge tree that matches the question statement based on the focus point information corresponding to the question statement;

[0050] As a module, it is used to perform intelligent question and answer based on the nodes on the knowledge tree that match the question statement as answer nodes.

[0051] The present application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0052] The present application also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the above method are implemented.

[0053] The present application establishes a knowledge tree with a simple structure. The knowledge tree is derived from a root node, and according to the association relationship of the knowledge structure, sub-nodes and leaf nodes connected to the root node are established in sequence. Each node has fixed elements: node position, node description, node keywords, node attributes (for example, definition, process, method, etc.), and node sub-nodes. According to the question statement raised by the user, the keyword classification information in the question statement or the user's focus information can be determined, and then the corresponding node in the knowledge tree can be matched as the answer node corresponding to the answer to the question statement, and the knowledge content corresponding to the answer node can be replied to the user, so as to realize intelligent question and answer in special fields with small or extremely small data volume. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 A flowchart of a smart question-answering method according to an embodiment of the present application;

[0055] Figure 2 A schematic diagram of a knowledge tree structure according to an embodiment of the present application;

[0056] Figure 3 A schematic diagram of the process of step S2 of an embodiment of the present application;

[0057] Figure 4 A schematic diagram of the process of step S3 of an embodiment of the present application;

[0058] Figure 5 A schematic diagram of the process of step S34 of an embodiment of the present application;

[0059] Figure 6 A schematic diagram of the process of step S4 of an embodiment of the present application;

[0060] Figure 7 A schematic diagram of the process of step S5 of an embodiment of the present application;

[0061] Figure 8 A schematic diagram of the process of step S6 of an embodiment of the present application;

[0062] Fig. 9 A schematic diagram of the process of step S1 of an embodiment of the present application;

[0063] Fig.10 A schematic diagram of the structure of an intelligent question-answering device according to an embodiment of the present application;

[0064] Fig.11 A schematic diagram of the internal structure of a computer device according to an embodiment of the present application. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0066] Reference Figure 1 The intelligent question answering method of this embodiment includes:

[0067] S1: Get the question sentence raised by the user.

[0068] The questions raised by the user can be obtained through the voice receiving system or the text receiving system, and the sentences corresponding to the questions can be saved.

[0069] S2: extracting the corresponding keywords in the question sentence and obtaining classification information of the keywords.

[0070] After removing stop words from the question sentence, the keywords of the question sentence are retained, and then the corresponding category annotation information is searched in the keyword dictionary according to the keywords to obtain the classification information of the keywords. There are generally some words without practical meaning in the question sentence, such as auxiliary words or adverbs such as "what", "maybe", "of", etc., which are called stop words. Stop words themselves have no effect on the accuracy of the classification algorithm, but if there are a large number of them, it will seriously affect the calculation speed, so stop words will be deleted in advance.

[0071] S3: Determine whether a node matching the classification information of the keyword can be found on the knowledge tree based on the classification information of the keyword.

[0072] The structure of the knowledge tree of this application is as follows Figure 2 As shown, it is led out by a root node, and each node is established in sequence according to the knowledge structure system. Each node has fixed elements, including: node position, node description, node keywords, node attributes (for example: definition, process, method, etc.), and node child nodes. The structure of the above knowledge tree is constructed by traversing the knowledge materials, searching for knowledge point content, and generating nodes corresponding to each knowledge point content according to the knowledge structure. There is no need to design, edit and build a complex traditional graph database. The application effect is very good for fields with only a small amount of data.

[0073] Based on the keywords of the question statement and the keyword table corresponding to each node, a keyword search algorithm is used to match nodes to determine the node corresponding to the question statement on the knowledge tree, and the knowledge of the node is used as the answer to the question statement.

[0074] S4: If not, then perform intent recognition on the question statement to obtain focus information corresponding to the question statement.

[0075] The intention recognition process of this application extracts semantic features of question statements based on the Bert model, and obtains the user's focus information through a text classification neural network. The semantic features of question statements are extracted based on the Bert model, and only coarse-grained information classification is used. Not only is the initial manual labeling workload low, but it can also obtain the user's intention without over-relying on the possible results of the text classification neural network. For example: the question statement is "What is the bidding process?" or "How to bid?", the semantic features of the question statement are extracted based on the Bert model, and the user's focus information is obtained through the text classification neural network: "process" or "how to proceed".

[0076] S5: Determine a node on the knowledge tree that matches the question statement according to the focus point information corresponding to the question statement.

[0077] The obtained focus point information is compared with the node attributes of each node, and the nodes that are consistent with each other are retained.

[0078] S6: Perform intelligent question and answer based on the nodes on the knowledge tree that match the question statement as answer nodes.

[0079] The present application establishes a knowledge tree with a simple structure. The knowledge tree is derived from a root node, and according to the association relationship of the knowledge structure, sub-nodes and leaf nodes connected to the root node are established in sequence. Each node has fixed elements: node position, node description, node keywords, node attributes (for example, definition, process, method, etc.), and node sub-nodes. According to the question statement raised by the user, the keyword classification information in the question statement or the user's focus information can be determined, and then the corresponding node in the knowledge tree can be matched as the answer node corresponding to the answer to the question statement, and the knowledge content corresponding to the answer node can be replied to the user, so as to realize intelligent question and answer in special fields with small or extremely small data volume.

[0080] Reference Figure 3 The step S2 of extracting the keywords corresponding to the question statement and obtaining the classification information of the keywords includes:

[0081] S21: Segment the question sentence to obtain a segmentation combination corresponding to the question sentence;

[0082] S22: removing stop words from the word segmentation combination to obtain a first modified word segmentation combination;

[0083] S23: determining whether there is a designated word in the first modified word segmentation combination that is different from the terminology rule of the preset keyword dictionary;

[0084] S24: If yes, then replace the designated vocabulary with standardized vocabulary corresponding to the designated vocabulary one by one to obtain a second modified word combination, wherein the standardized vocabulary is included in the keyword dictionary;

[0085] S25: extracting characteristic words corresponding to the specified part of speech in the second modified participle combination as the keyword;

[0086] S26: Determine the classification information corresponding to the keyword in the keyword dictionary.

[0087] When extracting keywords from question statements, this application will segment the question statements, rewrite the syntax and replace the vocabulary. This includes unifying the uppercase and lowercase letters into a preset unified format, standardizing the replacement of professional terms, replacing synonyms, etc., so that the keywords of the question statements are the same as the terminology rules in the keyword dictionary, so as to more accurately locate the classification information of the keywords. The word segmentation technology used in the embodiment of this application is the pkuseg word segmentation technology; the syntax rewriting is to detect the colloquial vocabulary in the user's question statement and rewrite the colloquial vocabulary into standardized vocabulary, so that the questioning method is converted into a more unified and standard form. For example: "How to do **" will be rewritten as "the implementation method of **", "what is the use of ***" will be rewritten as "the role of ***", etc. Vocabulary replacement includes standard word replacement and synonym replacement. Standard word replacement is achieved by establishing a standard word list, and each standard word will map one or several common similar words. For example: "opening bids" and "bidding" will correspond to "bidding", "process", "steps", "procedures", etc. will correspond to "process", etc. The principle of synonym replacement is basically the same as the standard word replacement mentioned above. When replacing, the abbreviation or English of the proper noun is replaced with the proper noun in Chinese, for example, "2g" corresponds to "Government Business Center". After word segmentation and rewriting, the word segmentation combination of the question sentence mainly retains the characteristic words with entity meaning in the question sentence, including nouns and verbs, so as to determine the classification information of the keywords in the question sentence through these characteristic words. For example, according to the characteristic words, the corresponding category annotation information is searched in the keyword dictionary to obtain the classification information of the keywords in the question sentence, so as to accurately find the nodes on the knowledge tree structure.

[0088] Reference Figure 4 The step S3 of judging whether a node matching the classification information of the keyword can be found on the knowledge tree according to the classification information of the keyword includes:

[0089] S31: Obtain the keyword tables corresponding to each node of the knowledge tree, and the classification information corresponding to each keyword table, to obtain a classification information set corresponding to the knowledge tree;

[0090] S32: Determine whether the classification information of the keyword is included in the classification information set;

[0091] S33: If yes, then determine whether a node matching the classification information of the keyword can be found on the knowledge tree according to the classification information of the keyword;

[0092] S34: Determine, according to the keyword search algorithm, a node matching the classification information of the keyword.

[0093] In the embodiment of the present application, each node is associated with a keyword table, and the keyword table stores words representing the knowledge content on the node. For example, if the knowledge content on the node is the bidding content, the words included in the keyword table of this node are: bidding, tendering, bidding, benchmarking, tender documents, etc., and the classification information corresponding to the keyword table is the financial bidding category. The nodes of the knowledge tree of the present application have different classification information corresponding to the nodes according to the different keyword tables corresponding to the nodes. By summarizing the classification information corresponding to all nodes, a classification information set corresponding to the knowledge tree is formed. If the classification information of the keyword is included in the classification information set, it means that the node of the question sentence can be determined by matching the keywords.

[0094] Reference Figure 5 The step S34 of determining the node matching the classification information of the keyword according to the keyword search algorithm comprises:

[0095] S341: Obtaining vocabulary weights corresponding to all words in a specified keyword table corresponding to a specified node;

[0096] S342: Calculating a matching value between the designated node and the keyword according to the vocabulary weight;

[0097] S343: Calculate the matching values ​​of all nodes in the knowledge tree and the keyword respectively according to the calculation process of the matching value of the designated node and the keyword;

[0098] S344: The node corresponding to the maximum matching value is used as the node matching the keyword.

[0099] In the embodiment of the present application, the vocabulary in the keyword table corresponding to each node has different weights, and the stronger the vocabulary feature, the higher the vocabulary weight. The keywords of the question statement and the vocabulary in the keyword table corresponding to each node respectively calculate the matching value of the keywords of the question statement and the keyword table of each node respectively, sort them according to the size of the matching value, and use the node with the largest matching value as the answer node of the question statement, and use the knowledge content corresponding to the answer node as the answer to the question statement. The knowledge content of the nodes with other matching values ​​will not be presented in the answer. The vocabulary weight in the keyword table in the embodiment of the present application realizes vocabulary weight matching by setting three vocabulary levels and the weight ranges corresponding to the three vocabulary levels. Among the above three vocabulary levels, the first is the subject word, which can directly reflect the user's question focus, such as: red card, yellow card, three letters, back to back, etc.; the second is non-theme words, such as: definition, process, temporary, difference and other general words; the third is other words, such as: neutral words such as enterprise, equipment, clauses, etc. The subject word has the highest weight, the non-theme word has the lowest weight, and the other words have a medium weight.

[0100] Reference Figure 6The step S4 of performing intent recognition on the question statement to obtain the focus point information corresponding to the question statement includes:

[0101] S41: extracting semantic features of the question sentence;

[0102] S42: inputting the semantic features of the question sentence into a text classification neural network to obtain a coarse-grained information classification corresponding to the question sentence;

[0103] S43: Determine the focus point information corresponding to the question statement according to the coarse-grained information classification.

[0104] In an embodiment of the present application, the semantic features of the question sentence are extracted through the Bert model, and then the extracted semantic features are input into the text classification neural network to obtain a coarse-grained information classification. The text classification neural network of the present application is a model obtained by training based on coarse-grained annotation data. Through the Bert model, the sentence vector value of the question sentence is obtained as the semantic feature of the question sentence, and then the sentence vector value is input into the fully connected layer of the text classification neural network to obtain the question category of the question sentence, and determine which type the question category of the question sentence belongs to, that is, to obtain the user's focus information, and select the corresponding node structure in the knowledge tree to search for knowledge nodes. For example, if the focus information is "process", the node structure with the node attribute of process is selected for search.

[0105] Reference Figure 7 The step S5 of determining the node on the knowledge tree that matches the question statement according to the focus point information corresponding to the question statement comprises:

[0106] S51: Compare the focus point information with the node attributes of each node of the knowledge tree one by one, and obtain the matching value between the question statement and each node respectively;

[0107] S52: Use the node corresponding to the maximum matching value as the node on the knowledge tree that matches the question statement.

[0108] This application combines semantic feature analysis of question statements to obtain the focus information corresponding to the question statements, and by comparing the focus information with node attributes, more appropriate node knowledge is determined as the answer to the question. This application can combine keyword search and intent recognition to optimize keyword search results, so that its search results can be improved from multiple choices to three choices, two choices, or even directly give precise answers. In addition, by combining keyword search and intent recognition, the accuracy of the answer is higher, or even if the answer cannot be given, a clear prompt can be given. Compared with intelligent question and answer that relies entirely on deep neural networks, intelligent question and answer is more accurate and efficient.

[0109] The intelligent question-and-answer system of the present application does not require excessive data preparation. It only needs to edit the knowledge materials in the vertical field into a unique knowledge tree structure, so that a better intelligent question-and-answer experience can be achieved within the vertical field. In other embodiments of the present application, when the user inputs a question statement to be asked to the intelligent question-and-answer system, the question statement can also be directly subjected to coarse-grained intent recognition to obtain the intent analysis result, and the answer can be matched according to the intent analysis result, omitting the keyword classification matching process to save the processing flow of the intelligent question-and-answer and improve the feedback efficiency.

[0110] Reference Figure 8 The step S6 of performing intelligent question answering based on the node on the knowledge tree that matches the question statement as the answer node includes:

[0111] S61: Determine whether the answer node has any associated child node;

[0112] S62: If yes, obtain the knowledge content of the child node;

[0113] S63: Using the knowledge content corresponding to the answer node as an answer to the question statement, and using the knowledge content of the sub-node as a knowledge guide in intelligent question answering;

[0114] S64: Synchronously feeding back the answer to the question statement and the knowledge guidance to the terminal corresponding to the user.

[0115] In the embodiment of the present application, if the matched answer node has child nodes, it is considered that the matched answer node includes multiple child nodes, such as including Figure 2 Node 1, node 1-1 and node 1-2 in the answer node will only retain node 1 according to the principle that the root node is given priority for knowledge nodes with the same path. Nodes 1-1 and 1-2 will be displayed in a knowledge-guided manner and will not be directly retained in the answer. The structure of the knowledge tree of the present application stores the association relationship between knowledge nodes, thereby having better knowledge guidance and knowledge association capabilities. When the question statement asked by the user is answered, not only can the knowledge content of the current answer node be provided as the answer to the question statement, but also the related knowledge nodes can be pushed out. Especially in multi-fork processes, the intelligent question-answering system of the present application has a good knowledge guidance effect.

[0116] Reference Fig. 9 The step S1 of obtaining the question statement raised by the user comprises:

[0117] S11: Obtaining an input statement input by a user;

[0118] S12: Input the input sentence into the Bert model to obtain the prediction value corresponding to each word in the input sentence;

[0119] S13: judging whether the designated character is a wrong character according to the prediction value corresponding to the designated character, wherein the designated character is any one of all the characters in the input sentence;

[0120] S14: If yes, obtain the pinyin corresponding to the specified character;

[0121] S15: determining, according to the pinyin corresponding to the designated character, a candidate character with the highest pinyin similarity as a correction character for the designated character;

[0122] S16: After correcting all the typos in the input sentence, the question sentence raised by the user is obtained.

[0123] The present application corrects typos in user input sentences to improve the accuracy of keyword classification or determination of user focus information. For example, in the process of semantic feature extraction and obtaining user focus information through text classification neural network, errors are identified and corrected in text to improve the accuracy of classification. The typo correction process of the present application is as follows: the user's input sentence is input into the Bert model to obtain three or more prediction values ​​corresponding to each word in the input sentence. If all the prediction values ​​are the same, it is a correct word, otherwise it is judged as a typo. If it is a typo, the original word corresponding to the prediction value is matched by pinyin similarity, and the candidate word with the highest pinyin similarity is selected as the correct word after correction.

[0124] Reference Fig.10 , the intelligent question-answering device of an embodiment of the present application includes:

[0125] The acquisition module 1 is used to acquire the question statement raised by the user.

[0126] The questions raised by the user can be obtained through the voice receiving system or the text receiving system, and the sentences corresponding to the questions can be saved.

[0127] Extraction module 2 is used to extract the corresponding keywords in the question sentence and obtain the classification information of the keywords.

[0128] After removing stop words from the question sentence, the keywords of the question sentence are retained, and then the corresponding category annotation information is searched in the keyword dictionary according to the keywords to obtain the classification information of the keywords. There are generally some words without practical meaning in the question sentence, such as auxiliary words or adverbs such as "what", "maybe", "of", etc., which are called stop words. Stop words themselves have no effect on the accuracy of the classification algorithm, but if there are a large number of them, it will seriously affect the calculation speed, so stop words will be deleted in advance.

[0129] The judgment module 3 is used to judge whether a node matching the classification information of the keyword can be found on the knowledge tree according to the classification information of the keyword.

[0130] The structure of the knowledge tree of the present application is shown in 2, which is derived from a root node, and each node is established in sequence according to the knowledge structure system. Each node has fixed elements, including: node position, node description, node keywords, node attributes (for example: definition, process, method, etc.), and node child nodes. The structure of the above knowledge tree is constructed by traversing the knowledge materials, searching for knowledge point content, and generating nodes corresponding to each knowledge point content according to the knowledge structure. There is no need to design, edit and build a complex traditional graph database, and the application effect is very good for fields with only a small amount of data.

[0131] Based on the keywords of the question statement and the keyword table corresponding to each node, a keyword search algorithm is used to match nodes to determine the node corresponding to the question statement on the knowledge tree, and the knowledge of the node is used as the answer to the question statement.

[0132] The intention recognition module 4 is used to perform intention recognition on the question statement if a node matching the classification information of the keyword cannot be found on the knowledge tree to obtain the focus point information corresponding to the question statement.

[0133] The intention recognition process of this application extracts semantic features of question statements based on the Bert model, and obtains the user's focus information through a text classification neural network. The semantic features of question statements are extracted based on the Bert model, and only coarse-grained information classification is used. Not only is the initial manual labeling workload low, but it can also obtain the user's intention without over-relying on the possible results of the text classification neural network. For example: the question statement is "What is the bidding process?" or "How to bid?", the semantic features of the question statement are extracted based on the Bert model, and the user's focus information is obtained through the text classification neural network: "process" or "how to proceed".

[0134] The determination module 5 is used to determine the node on the knowledge tree that matches the question statement according to the focus point information corresponding to the question statement.

[0135] The obtained focus point information is compared with the node attributes of each node, and the nodes that are consistent with each other are retained.

[0136] As module 6, it is used to perform intelligent question and answer based on the nodes on the knowledge tree that match the question statement as answer nodes.

[0137] The present application establishes a knowledge tree with a simple structure. The knowledge tree is derived from a root node, and according to the association relationship of the knowledge structure, sub-nodes and leaf nodes connected to the root node are established in sequence. Each node has fixed elements: node position, node description, node keywords, node attributes (for example, definition, process, method, etc.), and node sub-nodes. According to the question statement raised by the user, the keyword classification information in the question statement or the user's focus information can be determined, and then the corresponding node in the knowledge tree can be matched as the answer node corresponding to the answer to the question statement, and the knowledge content corresponding to the answer node can be replied to the user, so as to realize intelligent question and answer in special fields with small or extremely small data volume.

[0138] Furthermore, the extraction module 2 comprises:

[0139] A word segmentation unit, used for segmenting the question sentence to obtain a word segmentation combination corresponding to the question sentence;

[0140] An obtaining unit is used to remove stop words from the word combination to obtain a first modified word combination;

[0141] A first judging unit, configured to judge whether there is a designated word in the first modified word segmentation combination that is different from a terminology rule of a preset keyword dictionary;

[0142] a replacement unit, configured to replace, if there is a designated vocabulary different from the terminology rule of the preset keyword dictionary, the designated vocabulary with a standardized vocabulary corresponding to the designated vocabulary in a one-to-one correspondence, to obtain a second modified participle combination, wherein the standardized vocabulary is included in the keyword dictionary;

[0143] A first extraction unit is used to extract a characteristic word corresponding to a specified part of speech in the second modified participle combination as the keyword;

[0144] The first determining unit is used to determine the classification information corresponding to the keyword in the keyword dictionary.

[0145] When extracting keywords from question statements, this application will segment the question statements, rewrite the syntax and replace the vocabulary. This includes unifying the uppercase and lowercase letters into a preset unified format, standardizing the replacement of professional terms, replacing synonyms, etc., so that the keywords of the question statements are the same as the terminology rules in the keyword dictionary, so as to more accurately locate the classification information of the keywords. The word segmentation technology used in the embodiment of this application is the pkuseg word segmentation technology; the syntax rewriting is to detect the colloquial vocabulary in the user's question statement and rewrite the colloquial vocabulary into standardized vocabulary, so that the questioning method is converted into a more unified and standard form. For example: "How to do **" will be rewritten as "the implementation method of **", "what is the use of ***" will be rewritten as "the role of ***", etc. Vocabulary replacement includes standard word replacement and synonym replacement. Standard word replacement is achieved by establishing a standard word list, and each standard word will map one or several common similar words. For example: "opening bids" and "bidding" will correspond to "bidding", "process", "steps", "procedures", etc. will correspond to "process", etc. The principle of synonym replacement is basically the same as the standard word replacement mentioned above. When replacing, the abbreviation or English of the proper noun is replaced with the proper noun in Chinese, for example, "2g" corresponds to "Government Business Center". After word segmentation and rewriting, the word segmentation combination of the question sentence mainly retains the characteristic words with entity meaning in the question sentence, including nouns and verbs, so as to determine the classification information of the keywords in the question sentence through these characteristic words. For example, according to the characteristic words, the corresponding category annotation information is searched in the keyword dictionary to obtain the classification information of the keywords in the question sentence, so as to accurately find the nodes on the knowledge tree structure.

[0146] Furthermore, the judging module 3 includes:

[0147] A first acquisition unit is used to acquire a keyword table corresponding to each node of the knowledge tree, and classification information corresponding to each keyword table, to obtain a classification information set corresponding to the knowledge tree;

[0148] A second judgment unit, used to judge whether the classification information of the keyword is included in the classification information set;

[0149] a determination unit, configured to determine, if included in the classification information set, whether a node matching the classification information of the keyword can be found on the knowledge tree according to the classification information of the keyword;

[0150] The second determining unit is used to determine the node matching the classification information of the keyword according to the keyword search algorithm.

[0151] In the embodiment of the present application, each node is associated with a keyword table, and the keyword table stores words representing the knowledge content on the node. For example, if the knowledge content on the node is the bidding content, the words included in the keyword table of this node are: bidding, tendering, bidding, benchmarking, tender documents, etc., and the classification information corresponding to the keyword table is the financial bidding category. The nodes of the knowledge tree of the present application have different classification information corresponding to the nodes according to the different keyword tables corresponding to the nodes. By summarizing the classification information corresponding to all nodes, a classification information set corresponding to the knowledge tree is formed. If the classification information of the keyword is included in the classification information set, it means that the node of the question sentence can be determined by matching the keywords.

[0152] Furthermore, the second determining unit includes:

[0153] An acquisition subunit is used to acquire vocabulary weights corresponding to all vocabulary in a specified keyword table corresponding to a specified node;

[0154] A first calculation subunit, configured to calculate a matching value between the designated node and the keyword according to the vocabulary weight;

[0155] A second calculation subunit, configured to calculate the matching values ​​of all nodes in the knowledge tree and the keyword respectively according to the calculation process of the matching value of the designated node and the keyword;

[0156] As a subunit, it is used to use the node corresponding to the maximum matching value as the node matching the keyword.

[0157] In the embodiment of the present application, the vocabulary in the keyword table corresponding to each node has different weights, and the stronger the vocabulary feature, the higher the vocabulary weight. The keywords of the question statement and the vocabulary in the keyword table corresponding to each node respectively calculate the matching value of the keywords of the question statement and the keyword table of each node respectively, sort them according to the size of the matching value, and use the node with the largest matching value as the answer node of the question statement, and use the knowledge content corresponding to the answer node as the answer to the question statement. The knowledge content of the nodes with other matching values ​​will not be presented in the answer. The vocabulary weight in the keyword table in the embodiment of the present application realizes vocabulary weight matching by setting three vocabulary levels and the weight ranges corresponding to the three vocabulary levels. Among the above three vocabulary levels, the first is the subject word, which can directly reflect the user's question focus, such as: red card, yellow card, three letters, back to back, etc.; the second is non-theme words, such as: definition, process, temporary, difference and other general words; the third is other words, such as: neutral words such as enterprise, equipment, clauses, etc. The subject word has the highest weight, the non-theme word has the lowest weight, and the other words have a medium weight.

[0158] Furthermore, the intention recognition module 4 includes:

[0159] A second extraction unit, used to extract semantic features of the question sentence;

[0160] An input unit, used for inputting the semantic features of the question sentence into a text classification neural network to obtain a coarse-grained information classification corresponding to the question sentence;

[0161] The third determining unit is used to determine the focus point information corresponding to the question sentence according to the coarse-grained information classification.

[0162] In an embodiment of the present application, the semantic features of the question sentence are extracted through the Bert model, and then the extracted semantic features are input into the text classification neural network to obtain a coarse-grained information classification. The text classification neural network of the present application is a model obtained by training based on coarse-grained annotation data. Through the Bert model, the sentence vector value of the question sentence is obtained as the semantic feature of the question sentence, and then the sentence vector value is input into the fully connected layer of the text classification neural network to obtain the question category of the question sentence, and determine which type the question category of the question sentence belongs to, that is, to obtain the user's focus information, and select the corresponding node structure in the knowledge tree to search for knowledge nodes. For example, if the focus information is "process", the node structure with the node attribute of process is selected for search.

[0163] Further, module 5 is determined, including:

[0164] A comparison unit, used to compare the focus point information with the node attributes of each node of the knowledge tree one by one, and obtain a matching value between the question statement and each node;

[0165] The first unit is used to use the node corresponding to the maximum matching value as the node on the knowledge tree that matches the question statement.

[0166] This application combines semantic feature analysis of question statements to obtain the focus information corresponding to the question statements, and by comparing the focus information with node attributes, more appropriate node knowledge is determined as the answer to the question. This application can combine keyword search and intent recognition to optimize keyword search results, so that its search results can be improved from multiple choices to three choices, two choices, or even directly give precise answers. In addition, by combining keyword search and intent recognition, the accuracy of the answer is higher, or even if the answer cannot be given, a clear prompt can be given. Compared with intelligent question and answer that relies entirely on deep neural networks, intelligent question and answer is more accurate and efficient.

[0167] The intelligent question-and-answer system of the present application does not require excessive data preparation. It only needs to edit the knowledge materials in the vertical field into a unique knowledge tree structure, so that a better intelligent question-and-answer experience can be achieved within the vertical field. In other embodiments of the present application, when the user inputs a question statement to be asked to the intelligent question-and-answer system, the question statement can also be directly subjected to coarse-grained intent recognition to obtain the intent analysis result, and the answer can be matched according to the intent analysis result, omitting the keyword classification matching process to save the processing flow of the intelligent question-and-answer and improve the feedback efficiency.

[0168] Furthermore, as module 6, it includes:

[0169] A third judgment unit, used to judge whether the answer node has an associated child node;

[0170] A second acquisition unit, configured to acquire the knowledge content of the associated child node if there is one;

[0171] The second unit is used to use the knowledge content corresponding to the answer node as the answer to the question statement, and use the knowledge content of the sub-node as the knowledge guide in the intelligent question answering;

[0172] A feedback unit is used to synchronously feed back the answer to the question statement and the knowledge guidance to the terminal corresponding to the user.

[0173] In the embodiment of the present application, if the matched answer node has child nodes, it is considered that the matched answer node includes multiple child nodes, such as including Figure 2 Node 1, node 1-1 and node 1-2 in the answer node will only retain node 1 according to the principle that the root node is given priority for knowledge nodes with the same path. Nodes 1-1 and 1-2 will be displayed in a knowledge-guided manner and will not be directly retained in the answer. The structure of the knowledge tree of the present application stores the association relationship between knowledge nodes, thereby having better knowledge guidance and knowledge association capabilities. When the question statement asked by the user is answered, not only can the knowledge content of the current answer node be provided as the answer to the question statement, but also the related knowledge nodes can be pushed out. Especially in multi-fork processes, the intelligent question-answering system of the present application has a good knowledge guidance effect.

[0174] Further, the acquisition module 1 includes:

[0175] A third acquisition unit, used to acquire an input sentence input by a user;

[0176] A fourth acquisition unit, used for inputting the input sentence into a Bert model to acquire a prediction value corresponding to each word in the input sentence;

[0177] a fourth judgment unit, configured to judge whether the designated character is a wrong character according to a prediction value corresponding to the designated character, wherein the designated character is any one of all characters in the input sentence;

[0178] A fifth acquisition unit, for acquiring the pinyin corresponding to the designated character if it is a wrong character;

[0179] A fourth determining unit, configured to determine, according to the pinyin corresponding to the designated character, a candidate character with the highest pinyin similarity as a correction character for the designated character;

[0180] The correction unit is used to correct all the wrong characters in the input sentence to obtain the question sentence raised by the user.

[0181] The present application corrects typos in user input sentences to improve the accuracy of keyword classification or determination of user focus information. For example, in the process of semantic feature extraction and obtaining user focus information through text classification neural network, errors are identified and corrected in text to improve the accuracy of classification. The typo correction process of the present application is as follows: the user's input sentence is input into the Bert model to obtain three or more prediction values ​​corresponding to each word in the input sentence. If all the prediction values ​​are the same, it is a correct word, otherwise it is judged as a typo. If it is a typo, the original word corresponding to the prediction value is matched by pinyin similarity, and the candidate word with the highest pinyin similarity is selected as the correct word after correction.

[0182] Reference Fig.11 In an embodiment of the present application, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Fig.11 As shown. The computer device includes a processor, a memory, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store all data required for the process of intelligent question and answer. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the method of intelligent question and answer is implemented.

[0183] The above-mentioned processor executes the above-mentioned intelligent question and answer method, including: obtaining a question statement raised by a user; extracting corresponding keywords in the question statement, and obtaining classification information of the keywords; judging whether a node matching the classification information of the keywords can be found on the knowledge tree based on the classification information of the keywords; if not, performing intent recognition on the question statement to obtain focus information corresponding to the question statement; determining a node on the knowledge tree that matches the question statement based on the focus information corresponding to the question statement; and performing intelligent question and answer based on the node on the knowledge tree that matches the question statement as the answer node.

[0184] The above-mentioned computer device establishes a knowledge tree with a simple structure. The knowledge tree is led by a root node, and according to the association relationship of the knowledge structure, sub-nodes and leaf nodes connected to the root node are established in sequence. Each node has fixed elements: node position, node description, node keywords, node attributes (for example: definition, process, method, etc.), node sub-nodes, and can determine the keyword classification information or user focus information in the question statement according to the question statement raised by the user, and then match the corresponding node in the knowledge tree as the answer node corresponding to the answer to the question statement, and reply the knowledge content corresponding to the answer node to the user, so as to realize intelligent question and answer in special fields with small or extremely small data volume.

[0185] In one embodiment, the processor extracts the keywords corresponding to the question sentence and obtains the classification information of the keywords, including: performing segmentation on the question sentence after removing stop words to obtain a segmentation combination corresponding to the question sentence; judging whether there is a designated vocabulary different from the terminology rules of a preset keyword dictionary in the segmentation combination; if so, replacing the designated vocabulary with standardized vocabulary corresponding to the designated vocabulary one by one to obtain a revised segmentation combination, wherein the standardized vocabulary is included in the keyword dictionary; extracting the feature vocabulary corresponding to the designated part of speech in the revised segmentation combination as the keyword; and determining the classification information corresponding to the keyword in the keyword dictionary.

[0186] In one embodiment, the processor determines whether a node matching the classification information of the keyword can be found on the knowledge tree based on the classification information of the keyword, including: obtaining a keyword table corresponding to each node of the knowledge tree, and classification information corresponding to each keyword table, to obtain a classification information set corresponding to the knowledge tree; determining whether the classification information of the keyword is included in the classification information set; if so, determining that a node matching the classification information of the keyword can be found on the knowledge tree based on the classification information of the keyword; and determining a node matching the classification information of the keyword based on a keyword search algorithm.

[0187] In one embodiment, the processor determines the step of a node matching the classification information of the keyword based on a keyword search algorithm, including: obtaining the vocabulary weights corresponding to all words in a specified keyword table corresponding to the specified node, wherein the specified node is any one of all nodes in the knowledge tree; calculating the matching value between the specified node and the keyword based on the vocabulary weights; calculating the matching values ​​between all nodes in the knowledge tree and the keyword respectively according to the calculation process of the matching value between the specified node and the keyword; and taking the node corresponding to the maximum matching value as the node matching the keyword.

[0188] In one embodiment, the processor performs intent recognition on the question statement to obtain the focus information corresponding to the question statement, including: extracting semantic features of the question statement; inputting the semantic features of the question statement into a text classification neural network to obtain a coarse-grained information classification corresponding to the question statement; and determining the focus information corresponding to the question statement based on the coarse-grained information classification.

[0189] In one embodiment, the processor uses the node on the knowledge tree that matches the question statement as the answer node for intelligent question answering, including: determining whether the answer node has an associated child node; if so, obtaining the knowledge content of the child node; using the knowledge content corresponding to the answer node as the answer to the question statement, and using the knowledge content of the child node as the knowledge guidance in intelligent question answering; and synchronously feeding back the answer to the question statement and the knowledge guidance to the terminal corresponding to the user.

[0190] In one embodiment, the step of the processor obtaining a question statement raised by a user includes: obtaining an input statement input by the user; inputting the input statement into a Bert model to obtain a prediction value corresponding to each character in the input statement; judging whether the specified character is a wrong character according to the prediction value corresponding to the specified character, wherein the specified character is any one of all the characters in the input statement; if so, obtaining the pinyin corresponding to the specified character; determining a candidate character with the highest pinyin similarity according to the pinyin corresponding to the specified character as a correction character for the specified character; after correcting all the wrong characters in the input statement, obtaining the question statement raised by the user.

[0191] Those skilled in the art will understand that Fig.11 The structure shown in is merely a block diagram of a portion of the structure related to the present application solution and does not constitute a limitation on the computer device to which the present application solution is applied.

[0192] An embodiment of the present application also provides a computer-readable storage medium having a computer program stored thereon, and a method for implementing intelligent question and answering when the computer program is executed by a processor, comprising: obtaining a question statement raised by a user; extracting corresponding keywords in the question statement, and obtaining classification information of the keywords; determining whether a node matching the classification information of the keywords can be found on a knowledge tree based on the classification information of the keywords; if not, performing intent recognition on the question statement to obtain focus information corresponding to the question statement; determining a node on the knowledge tree matching the question statement based on the focus information corresponding to the question statement; and performing intelligent question and answering based on the node on the knowledge tree matching the question statement as an answer node.

[0193] The above-mentioned computer-readable storage medium establishes a knowledge tree with a simple structure. The knowledge tree is derived from a root node, and according to the association relationship of the knowledge structure, sub-nodes and leaf nodes connected to the root node are established in sequence. Each node has fixed elements: node position, node description, node keywords, node attributes (for example: definition, process, method, etc.), and node sub-nodes. According to the question statement raised by the user, the keyword classification information in the question statement or the user's focus information can be determined, and then the corresponding node in the knowledge tree is matched as the answer node corresponding to the answer to the question statement, and the knowledge content corresponding to the answer node is replied to the user, so as to realize intelligent question and answer in special fields with small or extremely small data volume.

[0194] In one embodiment, the processor extracts the keywords corresponding to the question sentence and obtains the classification information of the keywords, including: performing segmentation on the question sentence after removing stop words to obtain a segmentation combination corresponding to the question sentence; judging whether there is a designated vocabulary different from the terminology rules of a preset keyword dictionary in the segmentation combination; if so, replacing the designated vocabulary with standardized vocabulary corresponding to the designated vocabulary one by one to obtain a revised segmentation combination, wherein the standardized vocabulary is included in the keyword dictionary; extracting the feature vocabulary corresponding to the designated part of speech in the revised segmentation combination as the keyword; and determining the classification information corresponding to the keyword in the keyword dictionary.

[0195] In one embodiment, the processor determines whether a node matching the classification information of the keyword can be found on the knowledge tree based on the classification information of the keyword, including: obtaining a keyword table corresponding to each node of the knowledge tree, and classification information corresponding to each keyword table, to obtain a classification information set corresponding to the knowledge tree; determining whether the classification information of the keyword is included in the classification information set; if so, determining that a node matching the classification information of the keyword can be found on the knowledge tree based on the classification information of the keyword; and determining a node matching the classification information of the keyword based on a keyword search algorithm.

[0196] In one embodiment, the processor determines the step of a node matching the classification information of the keyword based on a keyword search algorithm, including: obtaining the vocabulary weight corresponding to all words in a specified keyword table corresponding to the specified node, wherein the specified node is any one of all nodes in the knowledge tree; calculating the matching value between the specified node and the keyword based on the vocabulary weight; calculating the matching values ​​between all nodes in the knowledge tree and the keyword respectively according to the calculation process of the matching value between the specified node and the keyword; and taking the node corresponding to the maximum matching value as the node matching the keyword.

[0197] In one embodiment, the processor performs intent recognition on the question statement to obtain the focus information corresponding to the question statement, including: extracting semantic features of the question statement; inputting the semantic features of the question statement into a text classification neural network to obtain a coarse-grained information classification corresponding to the question statement; and determining the focus information corresponding to the question statement based on the coarse-grained information classification.

[0198] In one embodiment, the processor uses the node on the knowledge tree that matches the question statement as the answer node for intelligent question answering, including: determining whether the answer node has an associated child node; if so, obtaining the knowledge content of the child node; using the knowledge content corresponding to the answer node as the answer to the question statement, and using the knowledge content of the child node as the knowledge guidance in intelligent question answering; and synchronously feeding back the answer to the question statement and the knowledge guidance to the terminal corresponding to the user.

[0199] In one embodiment, the step of the processor obtaining a question statement raised by a user includes: obtaining an input statement input by the user; inputting the input statement into a Bert model to obtain a prediction value corresponding to each character in the input statement; judging whether the specified character is a wrong character according to the prediction value corresponding to the specified character, wherein the specified character is any one of all the characters in the input statement; if so, obtaining the pinyin corresponding to the specified character; determining a candidate character with the highest pinyin similarity according to the pinyin corresponding to the specified character as a correction character for the specified character; after correcting all the wrong characters in the input statement, obtaining the question statement raised by the user.

[0200] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing related hardware through a computer program, and the above-mentioned computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided in this application and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0201] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the existence of other identical elements in the process, device, article or method including the element.

[0202] The above description is only a preferred embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for intelligent question answering, characterized in that: include: Get the question sentence raised by the user; Extracting the keywords corresponding to the question sentence and obtaining classification information of the keywords; Determine whether a node matching the classification information of the keyword can be found on the knowledge tree according to the classification information of the keyword; If not, then perform intent recognition on the question sentence to obtain focus information corresponding to the question sentence; Extracting semantic features of the question sentence based on the Bert model, using coarse-grained information classification, and obtaining the user's focus information through a text classification neural network; Determining a node on the knowledge tree that matches the question statement according to the focus point information corresponding to the question statement; Perform intelligent question and answer based on the node on the knowledge tree that matches the question statement as the answer node; The step of determining whether a node matching the classification information of the keyword can be found on the knowledge tree according to the classification information of the keyword comprises: Obtaining keyword tables corresponding to each node of the knowledge tree and classification information corresponding to each keyword table, to obtain a classification information set corresponding to the knowledge tree; Determining whether the classification information of the keyword is included in the classification information set; If so, determining that according to the classification information of the keyword, a node matching the classification information of the keyword can be found on the knowledge tree; According to the keyword search algorithm, a node matching the classification information of the keyword is determined.

2. The intelligent question-answering method according to claim 1, characterized in that: The step of extracting the keywords corresponding to the question sentence and obtaining the classification information of the keywords includes: After removing stop words from the question sentence, perform word segmentation to obtain a word segmentation combination corresponding to the question sentence; Determining whether there is a designated word in the word segmentation combination that is different from the terminology rules of the preset keyword dictionary; If so, the designated words are replaced one by one with standardized words corresponding to the designated words to obtain a modified word segmentation combination, wherein the standardized words are included in the keyword dictionary; Extracting characteristic words corresponding to the specified part of speech in the modified participle combination as the keyword; The classification information corresponding to the keyword is determined in the keyword dictionary.

3. The intelligent question-answering method according to claim 1, characterized in that: The step of determining the node matching the classification information of the keyword according to the keyword search algorithm includes: Obtaining vocabulary weights corresponding to all vocabulary in a specified keyword table corresponding to a specified node, wherein the specified node is any one of all nodes in the knowledge tree; Calculating a matching value between the designated node and the keyword according to the vocabulary weight; According to the calculation process of the matching value between the designated node and the keyword, respectively calculate the matching value between all nodes in the knowledge tree and the keyword; The node corresponding to the maximum matching value is used as the node matching the keyword.

4. The intelligent question-answering method according to claim 1, characterized in that: The step of performing intent recognition on the question sentence to obtain focus information corresponding to the question sentence includes: Extracting semantic features of the question sentence; Inputting the semantic features of the question sentence into a text classification neural network to obtain a coarse-grained information classification corresponding to the question sentence; The focus point information corresponding to the question statement is determined according to the coarse-grained information classification.

5. The intelligent question-answering method according to claim 1, characterized in that: The step of performing intelligent question answering based on the node on the knowledge tree that matches the question statement as the answer node comprises: Determine whether the answer node has any associated child nodes; If so, obtaining the knowledge content of the child node; Using the knowledge content corresponding to the answer node as the answer to the question statement, and using the knowledge content of the sub-node as the knowledge guide in the intelligent question answering; The answer to the question statement and the knowledge guidance are synchronously fed back to the terminal corresponding to the user.

6. The intelligent question-answering method according to claim 1, characterized in that: The step of obtaining the question statement raised by the user includes: Get the input statement entered by the user; Input the input sentence into the Bert model to obtain the prediction value corresponding to each word in the input sentence; According to the prediction value corresponding to the designated character, determining whether the designated character is a wrong character, wherein the designated character is any one of all the characters in the input sentence; If yes, obtain the pinyin corresponding to the specified character; According to the pinyin corresponding to the designated character, determine the candidate character with the highest pinyin similarity as the correction character of the designated character; After correcting all the typos in the input sentence, the question sentence raised by the user is obtained.

7. An intelligent question-answering device, characterized in that: include: An acquisition module is used to acquire question statements raised by users; An extraction module, used to extract the corresponding keywords in the question sentence and obtain classification information of the keywords; A judgment module, used to judge whether a node matching the classification information of the keyword can be found on the knowledge tree according to the classification information of the keyword; An intention recognition module, for performing intention recognition on the question statement to obtain focus information corresponding to the question statement if a node matching the classification information of the keyword cannot be found on the knowledge tree; Extracting semantic features of the question sentence based on the Bert model, using coarse-grained information classification, and obtaining the user's focus information through a text classification neural network; A determination module, configured to determine a node on the knowledge tree that matches the question statement based on the focus point information corresponding to the question statement; As a module, used for performing intelligent question and answer based on the node on the knowledge tree that matches the question statement as the answer node; The step of determining whether a node matching the classification information of the keyword can be found on the knowledge tree according to the classification information of the keyword comprises: Obtaining keyword tables corresponding to each node of the knowledge tree and classification information corresponding to each keyword table, to obtain a classification information set corresponding to the knowledge tree; Determining whether the classification information of the keyword is included in the classification information set; If so, determining that according to the classification information of the keyword, a node matching the classification information of the keyword can be found on the knowledge tree; According to the keyword search algorithm, a node matching the classification information of the keyword is determined.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images