Semantic analysis method based on intelligent AI

Abstract

The invention discloses a semantic analysis method based on intelligent AI, relates to the technical field of semantic analysis, and solves the technical problem that errors exist in analyzed semantics due to the fact that a unified analysis mode is adopted for analysis. According to the method, the to-be-analyzed problem input by the user is subjected to keyword extraction and combined arrangement to obtain the reduced problem, the problem core can be focused, irrelevant information interference can be reduced, a more accurate basis is provided for subsequent matching and analysis, non-special problems and special problems can be more accurately distinguished by adopting multiple matching modes and threshold judgment, and the user experience is improved. Different processing methods are adopted for different types of problems, the pertinence and accuracy of semantic analysis are improved, different processing methods are adopted for special results and non-special results, and single-type problems and multi-type problems are processed respectively, so that the method can adapt to various complex problem situations and has high adaptability.

Description

Technical Field

[0001] The present invention relates to the technical field of semantic parsing, and in particular to a semantic parsing method based on intelligent AI. Background Art

[0002] With the rapid development of information technology, natural language processing is increasingly being used in various fields. As one of the important tasks of natural language processing, semantic parsing aims to understand the semantics of natural language sentences and convert them into a machine-understandable form for subsequent processing and application.

[0003] In recent years, with the continuous advancement of artificial intelligence technology, deep learning methods have achieved remarkable results in semantic parsing. Deep learning models can automatically learn the feature representation of language and process natural language end-to-end, improving the accuracy and efficiency of semantic parsing. At the same time, the development of technologies such as knowledge graphs and pre-trained language models has also provided new ideas and methods for semantic parsing.

[0004] According to Chinese patent publication number CN111382253A, a semantic parsing method and a semantic parser are disclosed, the method comprising: receiving a natural sentence to be parsed; determining a standard sentence corresponding to the natural sentence to be parsed; inputting the standard sentence into a pre-trained naive semantic parser to obtain a logical expression corresponding to the standard sentence.

[0005] However, when some existing AI semantic parsing is used, different input problems will have different parsing methods. If a unified parsing method is used for parsing, semantic errors will occur in the parsing. Summary of the invention

[0006] In view of the deficiencies in the prior art, the present invention provides a semantic parsing method based on intelligent AI, which solves the problem of errors in the parsed semantics caused by using a unified parsing method.

[0007] To achieve the above objectives, the present invention is implemented by the following technical solutions: a semantic parsing method based on intelligent AI, the method specifically comprising the following steps:

[0008] Step S01, extracting keywords from the problem to be analyzed, and preprocessing the problem to be analyzed according to the extracted keywords to obtain a reduced problem;

[0009] Step S02, matching the problem to be analyzed with the obtained reduced problem to obtain a pre-selected result, and making a conventional judgment on the problem to be analyzed based on the pre-selected result to generate a judgment result;

[0010] Step S03, analyzing the problem to be analyzed corresponding to the obtained special result, and obtaining segmentation information by segmenting the problem to be analyzed, and generating analysis information by performing semantic analysis based on the segmentation information;

[0011] Step S04, respectively analyzing the unsolved problems corresponding to the obtained non-special results and special results, analyzing the historical data, and screening the historical data in combination with the problem fields corresponding to the unsolved problems to obtain analysis regularity information;

[0012] Step S05: display the obtained semantic information to the corresponding operator.

[0013] As a further solution of the present invention: the specific method of obtaining the reduction problem in step S01 is:

[0014] The problem to be analyzed is obtained, and then the keywords in the problem to be analyzed are obtained according to the identification features, and the extracted keywords are combined and sorted to obtain a reduced problem.

[0015] As a further solution of the present invention: the specific method of generating the judgment result in step S02 is:

[0016] Match the obtained reduced problem with the database to obtain the pre-selected results, and label them as i, where i=1, 2, ..., j, where j represents the number of pre-selected results;

[0017] At the same time, the matching ratio between the reduced problem and the pre-selected result i is calculated, and the mean of the sum of all matching ratios is calculated, recorded as the matching mean, and the matching mean is compared with the threshold.

[0018] As a further solution of the present invention: the specific method of comparing the matching mean with the threshold in step S02 is:

[0019] When the matching mean is greater than the threshold, it means that the reduction problem is a non-special problem and a non-special result is generated. Conversely, when the matching mean is less than the threshold, it means that the reduction problem is a special problem and a special result is generated. The non-special result and the special result are combined to obtain the judgment result.

[0020] As a further solution of the present invention: the specific method of analyzing the problem to be solved corresponding to the special result in step S03 is:

[0021] The problem to be analyzed is segmented to obtain common features in the problem to be analyzed. The problem to be analyzed is segmented using the common features as segmentation points to obtain segmentation information, and the segmentation information is labeled as a, where a=1, 2, ..., b, where b represents the number of segmentation information, and a conventional judgment is performed on the segmentation information.

[0022] As a further solution of the present invention: the specific method of performing conventional judgment on the segmentation information in step S03 is:

[0023] Get any set of segmentation information as the target object, then get all the historical records within time t, calculate the occurrence ratio of the target object in the historical records, and compare the occurrence ratio with the preset value;

[0024] If the occurrence ratio is greater than the preset value, the target object is classified as regular information. Conversely, if the occurrence ratio is less than the preset value, the target object is classified as irregular information.

[0025] As a further solution of the present invention: the specific method of generating parsing information by performing semantic analysis according to the segmentation information in step S03 is:

[0026] The conventional information obtained by classification is not processed. The unconventional information obtained by classification is semantically translated to obtain unconventional translated information, and the unconventional translated information is combined with the conventional information to obtain combined information. The combined information is then matched and parsed with the database to generate parsed information.

[0027] As a further solution of the present invention: the specific method of analyzing the problem to be solved corresponding to the non-special result in step S04 is:

[0028] Obtain historical data, and classify different semantic analysis problems in the historical data to obtain the same type of processing information, then obtain the problem domain corresponding to the problem to be analyzed, and match the problem domain of the problem to be analyzed with the same type of processing information.

[0029] As a further solution of the present invention: the specific method of matching the problem domain of the problem to be analyzed with the same type of processing information in step S04 is:

[0030] Obtain the problem domain of the problem to be analyzed, identify the type of the problem domain, and perform secondary classification on the problem to be analyzed to obtain a single type of problem or multiple types of problems, and generate the problem type;

[0031] Then, the problem types obtained by classification are matched with the processing information of the same type, and the target processing information is generated. The number of problem domain types in the target processing information is identified to obtain single-type problems and multiple-type problems, and analysis is performed based on the number of different problem domain types.

[0032] As a further solution of the present invention: the specific method of analyzing the number of different problem domain types in step S04 is:

[0033] When the target processing information is a single type of question, the problem processing method corresponding to the target processing information is recorded as the standard analysis method, and semantic analysis is performed on the problem to be parsed according to the standard analysis method to obtain semantic information;

[0034] When the target processing information is a multi-type problem, the problem domains corresponding to the problem to be analyzed are obtained and numbered as o in the order of appearance, where o=1, 2, ..., p, and p represents the number of problem domains. Then, semantic analysis of the corresponding processing information of the same type is performed in sequence according to the numbering order o, and preprocessing information is obtained. At the same time, the obtained preprocessing information is subjected to secondary processing, and the specific secondary processing method is as follows:

[0035] The obtained preprocessed information is combined in the order of the labels to obtain preprocessed reorganized information, and the preprocessed reorganized information is deleted, the word overlap of the corresponding preprocessed information in the preprocessed reorganized information is calculated, and the word overlap is compared with the standard value;

[0036] If the vocabulary overlap is greater than the standard value, the corresponding preprocessing information will be marked. Otherwise, if the vocabulary overlap is less than the standard value, it will not be processed. At the same time, for the marked preprocessing information, the preprocessing information with large labels is semantically deleted based on the preprocessing information corresponding to the small labels to obtain semantic information.

[0037] The present invention provides a semantic parsing method based on intelligent AI. Compared with the prior art, it has the following beneficial effects:

[0038] The present invention extracts keywords and combines and organizes the questions to be analyzed input by the user to obtain reduced questions, which can focus on the core of the question, reduce the interference of irrelevant information, and provide a more accurate basis for subsequent matching and analysis. It adopts multiple matching methods and threshold judgments, and can more accurately distinguish between non-special questions and special questions. Different processing methods are adopted for different types of questions, which improves the pertinence and accuracy of semantic analysis. Different processing methods for special results and non-special results, as well as separate processing of single-type questions and multiple types of questions, make it possible to adapt to various complex problem situations and have strong adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a flow chart of the steps of the present invention;

[0040] Figure 2 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention.

[0042] For example, see Figure 1 , the present application provides a semantic parsing method based on intelligent AI, the method specifically comprises the following steps:

[0043] In step S01, keywords in the problem to be analyzed are extracted, and the problem to be analyzed is preprocessed according to the extracted keywords to obtain a reduced problem.

[0044] First, the question to be analyzed is obtained, and the question to be analyzed here is input by the user, and there are many ways of input, including voice input and text input. Then, the keywords in the question to be analyzed are obtained according to the recognition features, and the recognition features here are represented as corresponding subject, predicate and object terms or words. The specific keywords are represented as corresponding subjects, predicates and objects in the question to be analyzed, and the extracted keywords are combined and sorted to obtain the reduced question.

[0045] In a specific example, for example, the user inputs the question "Xiao Ming is happily flying a kite in the park." Through analysis, the subject is "Xiao Ming", the predicate is "flying a kite", and the object is "kite" (here, according to the specific analysis situation, "flying a kite happily in the park" can be regarded as a more complex combination of predicate and object). After extracting these keywords, they are combined and sorted to obtain the reduced question "Xiao Ming is flying a kite".

[0046] For example, a user asks a question using voice input: "Beautiful flowers bloom in the sun." Analysis shows that the subject is "flowers" and the predicate is "blooming." After combination and organization, the question can be reduced to "flowers blooming."

[0047] Next, in step S02, the problem to be analyzed is matched with the obtained reduced problem to obtain a pre-selected result, and a conventional judgment is performed on the problem to be analyzed based on the pre-selected result to generate a judgment result.

[0048] The obtained reduced problem is matched with the database, and the database here is pre-established. At the same time, the database can be established by using natural language text data collected from different fields through web crawlers, data set downloads, etc., and the obtained matching results are recorded as pre-selected results. The matching method here can be matched with the database through keyword matching. The keywords in the reduced problem are matched with the problems in the database respectively. If the problems in the database contain all the keywords in the reduced problem, it is considered that the match is successful. At the same time, the combination of keywords in the reduced problem is considered and matched with the problems in the database. Different keyword combination weights can be set to improve the accuracy of the match.

[0049] At the same time, the obtained pre-selected results are labeled as i, and i=1, 2, ..., j, where j represents the number of pre-selected results. Then, the matching ratio between the reduced problem and the pre-selected result i is calculated, and the matching ratio here represents the ratio of the number of data in the reduced problem to the number of data in the pre-selected results. At the same time, they are calculated in sequence. For example, the matching ratio between the reduced problem and the pre-selected result labeled 1 is calculated, and then the matching ratio between the pre-selected result labeled 2 and the reduced problem is calculated, and so on. The mean of the sum of all matching ratios is calculated and recorded as the matching mean. At the same time, the matching mean is compared with the threshold, and the specific value of the threshold here is set by the operator according to experience. When the matching mean is greater than the threshold, it indicates that the reduced problem is a non-special problem and a non-special result is generated. Conversely, when the matching mean is less than the threshold, it indicates that the reduced problem is a special problem and a special result is generated.

[0050] Next, in step S03, the problem to be analyzed corresponding to the obtained special result is analyzed, and segmentation information is obtained by segmenting the problem to be analyzed, and semantic analysis is performed based on the segmentation information to generate analysis information.

[0051] For the problem to be analyzed that is judged as a special result, the problem to be analyzed is segmented to obtain common features in the problem to be analyzed, and the common features here represent common words or common characters obtained through big data analysis. At the same time, the problem to be analyzed is segmented with the common features as the segmentation points to obtain segmentation information, and the segmentation information here includes a single character or a word group, and the segmentation information is labeled as a, and a=1, 2, ..., b, where b represents the number of segmentation information, and then the segmentation information is routinely judged as follows:

[0052] Get any set of segmented information as the target object, then get all the historical records within time t, where the historical records represent all the parsed records within time t, and calculate the occurrence ratio of the target object in the historical records, and compare the occurrence ratio with a preset value, and the specific value of the preset value is set by the operator. If the occurrence ratio is greater than the preset value, the target object is classified as regular information, otherwise, if the occurrence ratio is less than the preset value, the target object is classified as irregular information;

[0053] And so on, make regular judgments on all segmentation information;

[0054] At the same time, the conventional information obtained by classification is not processed, and the unconventional information obtained by classification is further semantically translated to obtain unconventional translation information, and the unconventional translation information is combined with the conventional information to obtain combined information, and then the combined information is matched and parsed with the database to generate parsed information.

[0055] The specific parsing method here is to establish a corresponding language model, and the process of establishing the language model is as follows:

[0056] First, collect corpus. Collect a large amount of natural language text data from multiple sources, including news articles, blogs, novels, academic papers, etc. Then perform data annotation. For specific semantic parsing tasks, the corpus needs to be annotated. The annotation content can include part-of-speech tagging, named entity recognition, semantic role labeling, etc. Annotation can be done by professional annotators, or automatic annotation tools can be used in combination with manual proofreading.

[0057] At the same time, appropriate methods and models are selected, including rule-based methods (analyzing the grammatical structure and semantic rules of the language and formulating corresponding parsing rules), statistical-based methods (using statistical machine learning algorithms, such as hidden Markov models (HMM), conditional random fields (CRF), etc., to train the corpus and learn the statistical laws of the language) and deep learning-based methods (deep learning models, such as recurrent neural networks (RNN), long short-term memory networks (LSTM), convolutional neural networks (CNN), etc., can automatically learn the feature representation of language and perform end-to-end semantic parsing of natural language).

[0058] Next, the collected corpus is preprocessed by cleaning, segmenting, removing stop words and other preprocessing operations (existing natural language processing toolkits such as NLTK, spaCy, etc. can be used for data preprocessing). According to the selected method and model, the annotated corpus is used for training (deep learning frameworks such as TensorFlow, PyTorch, etc. can be used for model training) and finally the corresponding language model is established.

[0059] Next, in step S04, the problem to be solved corresponding to the obtained non-special result is analyzed, and the analysis regularity information is obtained by analyzing the historical data and screening the historical data in combination with the problem field corresponding to the problem to be solved.

[0060] First, historical data is obtained, and different semantic analysis problems in the historical data are classified to obtain the same type of processing information. The classification here is based on the problem domain corresponding to the corresponding problem. There are a large number of semantic analysis problems in different fields such as science and technology, finance, and education in the historical data. After classification, science and technology problem processing information, financial problem processing information, and education problem processing information are formed. Then, the problem to be analyzed is obtained, and the problem domain corresponding to the problem to be analyzed is obtained. The problem domain here is obtained through machine learning, and the problem domain of the problem to be analyzed is matched with the same type of processing information. The specific matching method is as follows:

[0061] Obtain the problem domain of the problem to be analyzed, identify the type of the problem domain, and perform secondary classification on the problem to be analyzed to obtain a single type of problem or multiple types of problems, and generate a problem type. Here, the classification is based on the type of the problem domain. For example, if the problem to be analyzed only involves one domain, the corresponding problem to be analyzed is a single type of problem. If the problem to be analyzed involves two or more domains, the corresponding problem to be analyzed is a multiple type of problem. Now there is a problem to be analyzed, "What are the applications of the latest artificial intelligence technology in the medical field?" Through machine learning algorithm analysis, it is determined that the problem domain of this problem to be analyzed is the intersection of science and technology and medicine. Therefore, the problem to be analyzed is classified into multiple types of problems. Then, the problem type obtained by classification is matched with the same type of processing information, and the target processing information is generated;

[0062] When the target processing information is a single type of question, the problem processing method corresponding to the target processing information is recorded as the standard analysis method, and the problem processing method here represents the semantic matching method for the problem to be parsed, and the semantic information is obtained by performing semantic analysis on the problem to be parsed according to the standard analysis method;

[0063] Finally, in step S05, the obtained semantic information is displayed to the corresponding operator.

[0064] Embodiment 2, as Embodiment 2 of the present invention, is implemented on the basis of Embodiment 1, and differs from Embodiment 1 in that:

[0065] In step S04, when the target processing information is a multi-type problem, the problem domains corresponding to the problem to be analyzed are obtained and numbered as o in the order of appearance, and o=1, 2, ..., p, and p represents the number of problem domains. Then, semantic analysis of the corresponding processing information of the same type is performed in sequence according to the numbering order o, and preprocessing information is obtained. At the same time, the obtained preprocessing information is subjected to secondary processing, and the specific secondary processing method is as follows:

[0066] The obtained preprocessing information is combined in the order of the labels to obtain the preprocessing reorganization information, and the preprocessing reorganization information is deleted at the same time, and the vocabulary overlap of the corresponding preprocessing information in the preprocessing reorganization information is calculated, and the vocabulary overlap here is to calculate the number of words that appear in the two texts together, and then divide it by the total number of words in the two texts to obtain the vocabulary overlap. Since the preprocessing reorganization information is obtained by combining multiple groups of preprocessing information, one group of preprocessing information is used as the analysis object, and the vocabulary overlap is calculated by the permutation and combination method, and then the calculated vocabulary overlap is compared with the standard value, and the standard value here is represented as a value for judging whether the two groups of preprocessing information are highly semantically similar;

[0067] If the vocabulary overlap is greater than the standard value, the corresponding preprocessing information will be marked. Otherwise, if the vocabulary overlap is less than the standard value, it will not be processed. At the same time, for the marked preprocessing information, the preprocessing information corresponding to the small number is used as the standard to perform semantic deletion on the large-number preprocessing information to obtain semantic information. Similarly, the same processing is performed on all marked preprocessing information.

[0068] In a specific example, for example, assume that there are three sets of preprocessing information, labeled P1, P2, and P3. The preprocessing reorganization information is "P1+P2+P3". First, calculate the vocabulary overlap of P1 and P2. Assuming that the number of common words is 5, the total vocabulary of P1 is 10, and the total vocabulary of P2 is 12, then the vocabulary overlap is 5 / (10+12-5)≈0.385. Set the standard value to 0.35. Since 0.385 is greater than 0.35, P1 and P2 are marked. Then calculate the vocabulary overlap of P1 and P3. Assuming that the number of common words is 4, the total vocabulary of P1 is 10, and the total vocabulary of P3 is 15, then the vocabulary overlap is 4 / (10+15-4)≈0.229. Since 0.229 is less than the standard value, P1 and P3 are not processed. Then, P2 is semantically deleted based on P1, and the unnecessary parts of P2 that are repeated with P1 are removed to obtain more accurate semantic information. Similarly, if there are other marked preprocessing information combinations, they are also processed in the same way.

[0069] Finally, in step S05, the obtained semantic information is displayed to the corresponding operator.

[0070] Embodiment 3, as the embodiment 3 of the present invention, focuses on combining the implementation processes of embodiment 1 and embodiment 2 for implementation.

[0071] Some of the data in the above formulas are dimensionless and numerically calculated. Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0072] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A semantic parsing method based on intelligent AI, characterized in that: The method specifically comprises the following steps: Step S01, extracting keywords from the problem to be analyzed, and preprocessing the problem to be analyzed according to the extracted keywords to obtain a reduced problem; Step S02, matching the problem to be analyzed with the obtained reduced problem to obtain a pre-selected result, and making a conventional judgment on the problem to be analyzed based on the pre-selected result to generate a judgment result; Step S03, analyzing the problem to be analyzed corresponding to the obtained special result, and obtaining segmentation information by segmenting the problem to be analyzed, and generating analysis information by performing semantic analysis based on the segmentation information; Step S04, respectively analyzing the unsolved problems corresponding to the obtained non-special results and special results, analyzing the historical data, and screening the historical data in combination with the problem fields corresponding to the unsolved problems to obtain analysis regularity information; Step S05: display the obtained semantic information to the corresponding operator.

2. The semantic parsing method based on intelligent AI according to claim 1, characterized in that: The specific method of obtaining the reduction problem in step S01 is: The problem to be analyzed is obtained, and then the keywords in the problem to be analyzed are obtained according to the identification features, and the extracted keywords are combined and sorted to obtain a reduced problem.

3. The semantic parsing method based on intelligent AI according to claim 1, characterized in that: The specific method of generating the judgment result in step S02 is: Match the obtained reduced problem with the database to obtain the pre-selected results, and label them as i, where i=1, 2, ..., j, where j represents the number of pre-selected results; At the same time, the matching ratio between the reduced problem and the pre-selected result i is calculated, and the mean of the sum of all matching ratios is calculated, recorded as the matching mean, and the matching mean is compared with the threshold.

4. The semantic parsing method based on intelligent AI according to claim 3, characterized in that: The specific method of comparing the matching mean with the threshold in step S02 is: When the matching mean is greater than the threshold, it means that the reduction problem is a non-special problem and a non-special result is generated. Conversely, when the matching mean is less than the threshold, it means that the reduction problem is a special problem and a special result is generated. The non-special result and the special result are combined to obtain the judgment result.

5. The semantic parsing method based on intelligent AI according to claim 1, characterized in that: The specific method of analyzing the problem to be analyzed corresponding to the special result in step S03 is: The problem to be analyzed is segmented to obtain common features in the problem to be analyzed. The problem to be analyzed is segmented using the common features as segmentation points to obtain segmentation information, and the segmentation information is labeled as a, where a=1, 2, ..., b, where b represents the number of segmentation information, and a conventional judgment is performed on the segmentation information.

6. The semantic parsing method based on intelligent AI according to claim 5, characterized in that: The specific method of performing conventional judgment on the segmentation information in step S03 is: Get any set of segmentation information as the target object, then get all the historical records within time t, calculate the occurrence ratio of the target object in the historical records, and compare the occurrence ratio with the preset value; If the occurrence ratio is greater than the preset value, the target object is classified as regular information. Conversely, if the occurrence ratio is less than the preset value, the target object is classified as irregular information.

7. The semantic parsing method based on intelligent AI according to claim 1, characterized in that: The specific method of generating parsing information by performing semantic analysis according to the segmentation information in step S03 is: The conventional information obtained by classification is not processed. The unconventional information obtained by classification is semantically translated to obtain unconventional translated information, and the unconventional translated information is combined with the conventional information to obtain combined information. The combined information is then matched and parsed with the database to generate parsed information.

8. The semantic parsing method based on intelligent AI according to claim 1, characterized in that: The specific method of analyzing the unsolved problem corresponding to the non-special result in step S04 is: Obtain historical data, and classify different semantic analysis problems in the historical data to obtain the same type of processing information, then obtain the problem domain corresponding to the problem to be analyzed, and match the problem domain of the problem to be analyzed with the same type of processing information.

9. The semantic parsing method based on intelligent AI according to claim 8, characterized in that: The specific method of matching the problem domain of the problem to be analyzed with the same type of processing information in step S04 is: Obtain the problem domain of the problem to be analyzed, identify the type of the problem domain, and perform secondary classification on the problem to be analyzed to obtain a single type of problem or multiple types of problems, and generate the problem type; Then, the problem types obtained by classification are matched with the processing information of the same type, and the target processing information is generated. The number of problem domain types in the target processing information is identified to obtain single-type problems and multiple-type problems, and analysis is performed based on the number of different problem domain types.

10. The semantic parsing method based on intelligent AI according to claim 9, characterized in that: The specific method of analyzing the number of different problem domain types in step S04 is: When the target processing information is a single type of question, the problem processing method corresponding to the target processing information is recorded as the standard analysis method, and semantic analysis is performed on the problem to be parsed according to the standard analysis method to obtain semantic information; When the target processing information is a multi-type problem, the problem domains corresponding to the problem to be analyzed are obtained and numbered as o in the order of appearance, where o=1, 2, ..., p, and p represents the number of problem domains. Then, semantic analysis of the corresponding processing information of the same type is performed in sequence according to the numbering order o, and preprocessing information is obtained. At the same time, the obtained preprocessing information is subjected to secondary processing, and the specific secondary processing method is as follows: The obtained preprocessed information is combined in the order of the labels to obtain preprocessed reorganized information, and the preprocessed reorganized information is deleted, the word overlap of the corresponding preprocessed information in the preprocessed reorganized information is calculated, and the word overlap is compared with the standard value; If the vocabulary overlap is greater than the standard value, the corresponding preprocessing information will be marked. Otherwise, if the vocabulary overlap is less than the standard value, it will not be processed. At the same time, for the marked preprocessing information, the preprocessing information with large labels is semantically deleted based on the preprocessing information corresponding to the small labels to obtain semantic information.

Images