Technology consultation intelligent reply method and system based on big data mining

Abstract

The invention relates to the technical field of big data mining, in particular to a technical consultation intelligent reply method and system based on big data mining, and the system comprises a data collection module, a data preprocessing module, a user demand analysis module, a knowledge graph construction module, an intelligent reply module and an effect evaluation module. The data preprocessing module adopts a cleaning mode of combining a regular expression and a box plot algorithm and a word segmentation labeling technology of Bi-LSTM and CRF models, can effectively remove data noise, process abnormal values and accurately process unstructured data, and provides a high-quality data basis for subsequent analysis and intelligent reply; according to the method, the user demand model is constructed through the clustering analysis and association rule algorithm, potential demands of the user can be deeply mined, the knowledge graph is combined, the system can more accurately understand user problems, and a foundation is laid for generating accurate replies.

Description

Technical Field

[0001] The present invention relates to the field of big data mining technology, and in particular to a method and system for intelligently replying to technical consultations based on big data mining. Background Art

[0002] In today's era of rapid development of information technology, technical consulting services are playing an increasingly important role in the development of businesses and individuals. With the rapid development of information technology, the demand for technical consulting services is growing. In the technical consulting process, how to efficiently process massive amounts of data and provide users with accurate and intelligent responses has become a key issue. Existing technical consulting methods and systems, such as an industry consulting method and system based on a big data platform with announcement number CN117829798A, collect various industry consulting related data based on the big data platform, wherein the various industry consulting related data include text data, voice data and image data, and extract information from the various industry consulting related data based on natural language processing technology and computer vision processing technology to obtain key information of various industry consulting, which can achieve the effect of autonomous data mining and greatly reduce the workload of manual data structuring, thereby solving the defects of the existing technology of wasting a lot of manpower and requiring professional experience. At the same time, by inputting the key information of various industry consulting into the self-learning model for training, an industry consulting model is obtained, and the industry consulting to be responded to is processed based on the industry consulting model, thereby realizing industry consulting based on the big data platform and achieving the effect of autonomous response to consultation.

[0003] Although the above applications have utilized big data or artificial intelligence technology to a certain extent, they still have shortcomings in the depth of data mining and the accuracy of intelligent responses. Existing intelligent response systems for technical consultations often use relatively simple cleaning and processing methods in the data preprocessing stage. They are unable to effectively remove noise and outliers in the data, and it is difficult to accurately process unstructured data, resulting in the accuracy and efficiency of subsequent data analysis and intelligent responses being affected. For example, when processing user consultation text data, some systems have low accuracy in word segmentation and part-of-speech tagging, which leads to deviations in the understanding of user needs, thereby affecting the quality of responses. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for intelligent reply to technical consultation based on big data mining to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions: The intelligent reply method for technical consultation based on big data mining includes the following steps: Collect historical user technical consultation data, industry-related technical documents, and competitor consultation data through web crawlers and API interfaces, clean the collected data, remove noise and duplicate data, and perform word segmentation and part-of-speech tagging preprocessing to convert unstructured data into structured data; Using the K-means clustering algorithm, we perform cluster analysis on the pre-processed user consultation data based on the keywords and semantic features of the user consultation questions, divide the user groups, and use the Apriori algorithm to explore the relationship between user consultation questions and technical solutions, and build a user demand model; Extract technical concepts, terminology, principles, and application cases from technical documents, academic papers, and industry reports. Use named entity recognition technology to identify knowledge entities, and use relationship extraction technology to determine the relationships between entities. This builds a technical knowledge graph centered on industry and technical fields. When a user asks a consultation question, we analyze the question to extract key information, determine the user's group and potential needs based on the user demand model, generate candidate response content based on the knowledge graph, use the neural network model to optimize the ranking of candidate responses, and present the response in natural language. Collect user feedback such as satisfaction evaluation of responses and further consultation needs, analyze the feedback using big data mining technology, evaluate the accuracy, practicality and completeness of responses, and update user demand models, knowledge graphs and intelligent response algorithms based on the evaluation results.

[0006] Preferably, in the data cleaning step, regular expression matching and outlier detection algorithms are used to remove data containing a large number of meaningless characters or format errors.

[0007] Preferably, in the knowledge graph construction step, the relationships between knowledge entities include inheritance relationships, association relationships, and application relationships, and the knowledge graph is stored in a graph database in a graph structure.

[0008] The technical consultation intelligent response system based on big data mining includes: The data collection module is used to interact with external data sources through web crawlers and API interfaces to collect historical user technical consultation data, industry technical documents, and competitor consultation data; The data preprocessing module uses regular expression matching and outlier detection algorithms to clean the collected data, and uses word segmentation tools to perform word segmentation and part-of-speech tagging preprocessing to convert unstructured data into structured data; The user demand analysis module uses the K-means clustering algorithm and Apriori algorithm to analyze the pre-processed user consultation data and build a user demand model; The knowledge graph construction module extracts knowledge information from multiple sources and uses named entity recognition and relationship extraction technology to build a technical knowledge graph centered on industry technology areas; The intelligent reply module analyzes user consultation questions, combines user demand models and knowledge graphs to generate candidate replies, uses a neural network model to optimize the ranking of candidate replies, and presents natural language replies; The effect evaluation module collects user feedback information, analyzes the feedback and evaluates the response effect, and updates the parameters and algorithms of each module of the system based on the evaluation results.

[0009] Preferably, the data acquisition module supports real-time data acquisition and periodic updating.

[0010] Preferably, the intelligent reply module is further provided with a semantic similarity calculation unit for calculating the semantic similarity between the candidate replies and the user questions to assist in sorting optimization.

[0011] Preferably, the data preprocessing module uses a regular expression pattern matching rule library and a box plot outlier detection algorithm to clean the collected data, and uses a word segmentation tool composed of a bidirectional long short-term memory network combined with a conditional random field model to perform word segmentation and part-of-speech tagging preprocessing to convert unstructured data into structured data, as follows: Data cleaning: Regular expression matching cleaning: Build a regular expression pattern matching rule library and define matching patterns for common format errors, meaningless characters, illegal characters, etc. in the data.

[0012] Preferably, it also includes: box plot outlier detection and cleaning; Boxplot outlier detection and cleaning: For numerical data, the boxplot principle is used to detect outliers; Assume the data set is , first calculate the lower quartile of the data set and upper quartile , interquartile range , the outlier determination formula is: Lower limit:

[0013] Upper limit:

[0014] in, The outlier determination coefficient is usually 1.5 or 3. or greater than The data are identified as outliers and cleaned.

[0015] Preferably, it also includes: word segmentation and part-of-speech tagging; Word segmentation and part-of-speech tagging: A bidirectional long short-term memory network (Bi-LSTM) is combined with a conditional random field (CRF) model to form a word segmentation tool; The Bi-LSTM model can make full use of context information for the input sequence. , the forward hidden state at time t and the backward hidden state The calculation is as follows:

[0016] Concatenate the forward and backward hidden states to get , input it into the CRF layer for part-of-speech tagging; the CRF layer calculates the state transition probability and emission probability to obtain the optimal part-of-speech tagging sequence , where the state transition probability matrix Indicates that from the label Transfer to label The probability of emission probability matrix Indicates the label at position t Generate observations The probability of is , and the optimal labeling sequence is solved by the Viterbi algorithm.

[0017] Compared with the prior art, the present invention has the following beneficial effects: The data preprocessing module of the present invention adopts a cleaning method that combines regular expressions with box plot algorithms, as well as word segmentation and tagging technology of Bi-LSTM and CRF models. It can effectively remove data noise, handle outliers, and accurately process unstructured data, providing a high-quality data foundation for subsequent analysis and intelligent response. The present invention constructs a user demand model through cluster analysis and association rule algorithm, which can deeply explore the user's potential needs. Combined with the knowledge graph, it enables the system to understand user problems more accurately and lay the foundation for generating accurate responses. Based on optimized data and accurate understanding of needs, the intelligent reply module of the present invention can quickly generate high-quality replies that meet user needs, thereby improving user satisfaction and technical consulting service efficiency. The effect evaluation module of the present invention optimizes the system according to user feedback, so that the system can continuously adapt to new demands and technological developments and maintain good performance and competitiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic diagram of the process of the present invention; Figure 2 It is a schematic diagram of the framework of the present invention. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention. Example 1

[0020] See also Figure 1-2 ,The intelligent reply method for technical consultation based on big data mining ,includes the following steps: We collect historical user technical consultation data, industry-related technical documents, and competitor consultation data through web crawlers and API interfaces. We clean the collected data and use regular expression matching and outlier detection algorithms to remove data containing a large number of meaningless characters or format errors, remove noise data and duplicate data, and perform word segmentation and part-of-speech tagging preprocessing to convert unstructured data into structured data. Using the K-means clustering algorithm, we perform cluster analysis on the pre-processed user consultation data based on the keywords and semantic features of the user consultation questions, divide the user groups, and use the Apriori algorithm to explore the relationship between user consultation questions and technical solutions, and build a user demand model; Extract technical concepts, terminology, principles, and application cases from technical documents, academic papers, and industry reports. Use named entity recognition to identify knowledge entities. Use relationship extraction to determine the relationships between entities. Build a technical knowledge graph centered on industry technology. The relationships between knowledge entities include inheritance, association, and application. The knowledge graph is stored in a graph database using a graph structure. When a user asks a consultation question, we analyze the question to extract key information, determine the user's group and potential needs based on the user demand model, generate candidate response content based on the knowledge graph, use the neural network model to optimize the ranking of candidate responses, and present the response in natural language. Collect user feedback such as satisfaction evaluation of responses and further consultation needs, analyze the feedback using big data mining technology, evaluate the accuracy, practicality and completeness of responses, and update user demand models, knowledge graphs and intelligent response algorithms based on the evaluation results. Example 2

[0021] The technical consultation intelligent response system based on big data mining includes: The data collection module is used to interact with external data sources through web crawlers and API interfaces to collect user technical consultation historical data, industry technical documents, and competitor consultation data.

[0022] The data preprocessing module uses regular expression matching and outlier detection algorithms to clean the collected data, and uses word segmentation tools to perform word segmentation and part-of-speech tagging preprocessing to convert unstructured data into structured data.

[0023] The data preprocessing module uses a regular expression pattern matching rule library and a boxplot outlier detection algorithm to clean the collected data. It uses a word segmentation tool composed of a bidirectional long short-term memory network (Bi-LSTM) combined with a conditional random field (CRF) model to perform word segmentation and part-of-speech tagging preprocessing, converting unstructured data into structured data. The details are as follows: Data cleaning: Regular expression matching cleaning: Build a regular expression pattern matching rule library and define matching patterns for common format errors, meaningless characters, illegal characters, etc. in the data.

[0024] Boxplot outlier detection and cleaning: For numerical data, the boxplot principle is used to detect outliers. Suppose the data set is , first calculate the lower quartile of the data set and upper quartile , interquartile range The outlier determination formula is: Lower limit:

[0025] Upper limit:

[0026] in, The outlier determination coefficient is usually 1.5 or 3. or greater than The data are identified as outliers and cleaned.

[0027] Word segmentation and part-of-speech tagging: The word segmentation tool is constructed by combining the Bi-LSTM network with the Conditional Random Field (CRF) model. The Bi-LSTM model can make full use of context information and , the forward hidden state at time t and the backward hidden state The calculation is as follows:

[0028] Concatenate the forward and backward hidden states to get , and input it into the CRF layer for part-of-speech tagging. The CRF layer calculates the state transition probability and emission probability to obtain the optimal part-of-speech tagging sequence , where the state transition probability matrix Indicates that from the label Transfer to label The probability of emission probability matrix Indicates the label at position t Generate observations The probability of is , and the optimal labeling sequence is solved by the Viterbi algorithm.

[0029] The user needs analysis module uses an improved K-means++ clustering algorithm to initialize cluster centers, combines TF-IDF (term frequency-inverse document frequency) to calculate keyword weights, and performs cluster analysis on pre-processed user consultation data. It also uses the FP-Growth algorithm to mine frequent item sets and association rules between user consultation questions and technical solutions, and construct a user needs model. The details are as follows: Randomly select a sample point from the data set as the first cluster center . For each unselected sample point , calculate the shortest distance to the selected cluster center ,in is the number of selected cluster centers. Select sample points As the next cluster center, Probability of being selected and is proportional to .

[0030] Repeat the above steps until K cluster centers are selected.

[0031] K-means clustering iterative calculation: For each sample point , calculate its distance to each cluster center Euclidean distance ,Will Assign to the cluster to which the nearest cluster center belongs, where n is the data dimension.

[0032] Recalculate the cluster center of each cluster ,in For the clusters, is the number of sample points in the cluster. Repeat the above steps until the cluster center no longer changes or the maximum number of iterations is reached.

[0033] Calculate keyword weights based on TF-IDF: Term Frequency (TF) calculation: ,in It is In the document The number of times a word appears, It is The sum of the occurrences of all words in a document.

[0034] Inverse Document Frequency (IDF) calculation: , where N is the total number of documents, is the number of documents containing the jth word. Keyword weight calculation: , keyword weights are used to assist in determining the core features of user consultation questions and optimize clustering results.

[0035] Association rule mining based on FP-Growth algorithm: Construct FP-Tree: Scan the data set, count the frequency of occurrence of each item, and delete items whose support is lower than the minimum support threshold min_sup. For each transaction in the transaction database, we sort the items in descending order of frequency and construct an FP-Tree. Starting from the root node, for each item in the transaction, if the item exists on the current path, we increase its count by 1; otherwise, we create a new node and connect it to the current path. Mining frequent itemsets: Starting from the leaf nodes of the FP-Tree, the conditional FP-Tree is recursively generated. For each node, a conditional pattern base ending with the node is generated, and then the conditional FP-Tree is constructed based on the conditional pattern base.

[0036] The process of mining frequent itemsets is repeated on the conditional FP-Tree until no new frequent itemsets can be found. Generate association rules: For each frequent itemset , calculate all its non-empty subsets Y, and calculate the association rules Confidence ,in and are the supports of frequent itemsets X and Y respectively. The association rules with confidence greater than the minimum confidence threshold min_conf are retained, so as to mine the association between user consultation questions and technical solutions and build a user demand model.

[0037] The knowledge graph construction module extracts knowledge information from multiple sources and uses named entity recognition and relationship extraction technology to build a technical knowledge graph with industry technology fields as the core.

[0038] The intelligent reply module analyzes user consultation questions, combines user demand models and knowledge graphs to generate candidate replies, uses a neural network model to optimize the ranking of candidate replies, and presents natural language replies; The intelligent reply module uses a BERT-based semantic understanding model to deeply analyze user consultation questions. Combining the user demand model with the knowledge graph, it uses a generative pre-trained model (such as an improved version of the GPT series) to generate candidate replies. It then uses a neural network model based on a multi-head attention mechanism to perform ranking optimization and ultimately present the replies in natural language. The details are as follows: (1) Analysis of user consultation questions BERT-based semantic encoding: Translating user inquiry questions As the input of the BERT model, the BERT model processes the problem through a multi-layer bidirectional Transformer encoder to obtain the contextual semantic representation of each word. , its final semantic vector representation The calculation is as follows: in, Represents the output of the BERT model The hidden state vector of each word. Key information extraction: Using the sequence tagging model based on the attention mechanism, the semantic vector output by BERT : in, It is calculated by the attention function The key information words are filtered out through attention weights to complete the analysis of user consultation questions. (2) Candidate response generation Combining the user groups and potential needs determined by the user demand model and the relevant knowledge in the knowledge graph, the parsed user questions and supplementary information are used as inputs for the generative pre-training model. Taking the improved GPT-based model as an example, candidate responses are generated through autoregression. In the generation process, the probability distribution of the next word is predicted based on the generated words. For the word at position t , and its probability is calculated as follows: Optimizing candidate response ranking A neural network model based on a multi-head attention mechanism is used to optimize the ranking of candidate responses. The semantic vector of the user's question is transformed into and the semantic vector of each candidate response (k=1, 2, ..., K, K is the number of candidate replies) as the input of the multi-head attention mechanism. Multi-head attention calculation: For each attention head , calculate the query vector , key vector Sum vector : ,in, , , Is a trainable weight matrix. Calculate the attention score : in, Is the dimension of the key vector. The attention weight is obtained by the softmax function Finally, the output of each attention head is obtained : Fusion of multi-head attention outputs: concatenate the outputs of all attention heads and perform linear transformation to obtain the final semantic representation :

[0039] Where h is the number of attention heads, is a trainable weight matrix. By calculating the similarity between O and the user's question semantic vector, the candidate responses are ranked and the optimal response is selected and presented to the user in natural language.

[0040] The effect evaluation module collects user feedback information, analyzes the feedback and evaluates the response effect, and updates the parameters and algorithms of each module of the system based on the evaluation results.

[0041] In the fuzzy comprehensive evaluation method, the criteria for classifying the review set must be determined based on the actual application scenario and the characteristics of user feedback. The following provides specific criteria for classifying the review set into "excellent," "good," "fair," and "poor" from both quantitative and qualitative perspectives: (1) Classification criteria based on quantitative indicators: Response accuracy Excellent (90-100 points): The response fully meets the user's inquiry question, covers all key knowledge points, contains no errors, omissions or misleading information, and the cited data, technical principles, etc. are accurate.

[0042] Good (75-89 points): The response content is generally consistent with the user's inquiry question, including most of the key knowledge points. There are a few vague expressions or minor errors that do not affect the core content, such as the inaccurate explanation of some minor concepts.

[0043] Medium (60-74 points): The reply content is somewhat related to the user's inquiry question, but some key knowledge points are missing, there are many errors or unclear expressions, which may cause some difficulty for users to understand.

[0044] Poor (0-59 points): The response has low relevance to the user's inquiry question, key knowledge points are seriously missing, and there is a lot of erroneous or misleading information, which cannot provide effective help to users.

[0045] (2) Response practicality: Excellent (90-100 points): The solutions and suggestions provided in the reply are highly practical and can directly help users solve practical problems. Users can implement them without having to refer to other materials.

[0046] Good (75-89 points): The solutions and suggestions provided in the reply are somewhat feasible, but users may need to combine them with other information or make appropriate adjustments before implementation. The implementation effect is good.

[0047] Medium (60-74 points): The solutions and suggestions provided in the reply are generally practical. Users need to spend more energy and time to understand, adjust and try them out, and the implementation effect is limited.

[0048] Poor (0-59 points): The solutions and suggestions provided in the reply are not practical, cannot help users solve practical problems, and may even provide the wrong direction.

[0049] (3) Completeness of response: Excellent (90-100 points): The response fully covers all aspects of the user's inquiry, including the background, principles, multiple possible solutions, precautions, etc., without any omissions.

[0050] Good (75-89 points): The response content basically covers the main aspects of the user's consultation issue. There are a few omissions in minor aspects, which do not affect the overall understanding and solution of the problem.

[0051] Medium (60-74 points): The response only covers part of the user's inquiry and omits important information, which may prevent the user from fully understanding and resolving the issue.

[0052] Poor (0-59 points): The response only covers a small part of the user's inquiry question, and a large amount of important content is missing, which cannot meet the user's need for a complete understanding of the problem.

[0053] (4) Classification criteria based on qualitative description: Excellent: User feedback clearly indicates great satisfaction, praising the responses as accurate, practical, and complete, resolving all their questions and problems and even exceeding their expectations. The user may actively recommend the system.

[0054] Good: Users express satisfaction with the feedback and acknowledge that the response has solved the problem to a certain extent. However, they also make some small suggestions or hope for further additions, believing that there is room for improvement.

[0055] Medium: User feedback is average, pointing out that the reply has some problems, such as being inaccurate, not practical, or incomplete, and needs further improvement, but does not completely deny the value of the reply.

[0056] Poor: User feedback indicates dissatisfaction, clearly pointing out that the response contains serious errors, is impractical, or seriously lacks key information, cannot solve the problem, and even misleads the user. The user requests a new reply or expresses disappointment with the system.

[0057] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent replying to technical consultation based on big data mining, characterized in that: The following steps are involved: Collect historical user technical consultation data, industry-related technical documents, and competitor consultation data through web crawlers and API interfaces, clean the collected data, remove noise and duplicate data, and perform word segmentation and part-of-speech tagging preprocessing to convert unstructured data into structured data; Using the K-means clustering algorithm, we perform cluster analysis on the pre-processed user consultation data based on the keywords and semantic features of the user consultation questions, divide the user groups, and use the Apriori algorithm to explore the relationship between user consultation questions and technical solutions, and build a user demand model; Extract technical concepts, terminology, principles, and application cases from technical documents, academic papers, and industry reports. Use named entity recognition technology to identify knowledge entities, and use relationship extraction technology to determine the relationships between entities. This builds a technical knowledge graph centered on industry and technical fields. When a user asks a consultation question, we analyze the question to extract key information, determine the user's group and potential needs based on the user demand model, generate candidate response content based on the knowledge graph, use the neural network model to optimize the ranking of candidate responses, and present the response in natural language. Collect user feedback such as satisfaction evaluation of responses and further consultation needs, analyze the feedback using big data mining technology, evaluate the accuracy, practicality and completeness of responses, and update user demand models, knowledge graphs and intelligent response algorithms based on the evaluation results.

2. The intelligent reply method for technical consultation based on big data mining according to claim 1 is characterized in that: In the data cleaning step, regular expression matching and outlier detection algorithms are used to remove data containing a large number of meaningless characters or format errors.

3. The intelligent reply method for technical consultation based on big data mining according to claim 1 is characterized in that: In the knowledge graph construction step, the relationships between knowledge entities include inheritance relationships, association relationships, and application relationships, and the knowledge graph is stored in a graph database in a graph structure.

4. The intelligent response system for technical consultation based on big data mining is characterized by: include: The data collection module is used to interact with external data sources through web crawlers and API interfaces to collect historical user technical consultation data, industry technical documents, and competitor consultation data; The data preprocessing module uses regular expression matching and outlier detection algorithms to clean the collected data, and uses word segmentation tools to perform word segmentation and part-of-speech tagging preprocessing to convert unstructured data into structured data; The user demand analysis module uses the K-means clustering algorithm and Apriori algorithm to analyze the pre-processed user consultation data and build a user demand model; The knowledge graph construction module extracts knowledge information from multiple sources and uses named entity recognition and relationship extraction technology to build a technical knowledge graph centered on industry technology areas; The intelligent reply module analyzes user consultation questions, combines user demand models and knowledge graphs to generate candidate replies, uses a neural network model to optimize the ranking of candidate replies, and presents natural language replies; The effect evaluation module collects user feedback information, analyzes the feedback and evaluates the response effect, and updates the parameters and algorithms of each module of the system based on the evaluation results.

5. The technical consultation intelligent reply system based on big data mining according to claim 1 is characterized in that: The data acquisition module supports real-time data acquisition and periodic updating.

6. The intelligent reply system for technical consultation based on big data mining according to claim 5 is characterized in that: The intelligent reply module is also provided with a semantic similarity calculation unit for calculating the semantic similarity between the candidate reply and the user question to assist in sorting optimization.

7. The intelligent reply system for technical consultation based on big data mining according to claim 5 is characterized in that: The data preprocessing module uses a regular expression pattern matching rule library and a boxplot outlier detection algorithm to clean the collected data, and uses a word segmentation tool composed of a bidirectional long short-term memory network combined with a conditional random field model to perform word segmentation and part-of-speech tagging preprocessing to convert unstructured data into structured data, as follows: Data cleaning: Regular expression matching cleaning: Build a regular expression pattern matching rule library and define matching patterns for common format errors, meaningless characters, illegal characters, etc. in the data.

8. The intelligent reply system for technical consultation based on big data mining according to claim 5 is characterized in that: Also includes: Boxplot outlier detection and cleaning; Boxplot outlier detection and cleaning: For numerical data, the boxplot principle is used to detect outliers; Assume the data set is , first calculate the lower quartile of the data set and upper quartile , interquartile range , the outlier determination formula is: Lower limit: Upper limit: in, The outlier determination coefficient is usually 1.5 or 3. or greater than The data are identified as outliers and cleaned.

9. The intelligent reply system for technical consultation based on big data mining according to claim 5 is characterized in that: Also includes: Word segmentation and part-of-speech tagging; Word segmentation and part-of-speech tagging: A bidirectional long short-term memory network (Bi-LSTM) is combined with a conditional random field (CRF) model to form a word segmentation tool; The Bi-LSTM model can make full use of context information for the input sequence. , the forward hidden state at time t and the backward hidden state The calculation is as follows: Concatenate the forward and backward hidden states to get , input it into the CRF layer for part-of-speech tagging; The CRF layer calculates the state transition probability and emission probability to obtain the optimal part-of-speech tagging sequence , where the state transition probability matrix Indicates that from the label Transfer to label The probability of emission probability matrix Indicates the label at position t Generate observations The probability of is , and the optimal labeling sequence is solved by the Viterbi algorithm.

Images