Aviation safety accident report analysis method based on topic modeling and word co-occurrence network

Abstract

The invention belongs to the field of natural language processing, particularly relates to an aviation safety accident report analysis method based on topic modeling and a word co-occurrence network, and aims to solve the problems that an existing topic modeling method is limited in understanding ability in aviation safety accident text analysis and cannot quantitatively reveal key causes. The method comprises the steps that a report text is acquired and preprocessed; based on a preset semantic fusion enhanced topic modeling engine keyword, obtaining a semantic fusion enhanced feature vector, and extracting a topic structure by adopting an improved deep embedding clustering model; and constructing an aviation safety accident word co-occurrence network based on a dynamic cosine similarity threshold value, performing network analysis, and determining key risk factors. According to the method, structured topic mining and deep learning semantic extraction are dynamically fused, and a network analysis method is combined, so that deep analysis and risk identification of the aviation safety accident report are realized, and the limitation of a traditional method on capturing text deep semantic information and subtle semantic difference is broken through.

Description

Technical Field

[0001] The present invention belongs to the technical field of natural language processing and data analysis, and specifically relates to an aviation safety accident report analysis method and system based on topic modeling and word co-occurrence network. Background Art

[0002] Aviation safety directly impacts the safety of lives and property. In-depth analysis of historical aviation accident data and the identification of risk factors and their underlying connections are crucial for preventing aviation accidents and optimizing aviation safety management decisions. Aviation safety data comes from diverse sources, with text data being one of the most important formats. The U.S. National Transportation Safety Board (NTSB) collects a large number of aviation accident reports, detailing the accident process, personnel, aircraft status, weather conditions, and possible causes. These safety reports are a core source of information for exploring aviation safety hazards and analyzing the causes of flight accidents, providing the industry with a data foundation for mining past accident information and learning lessons. However, this unstructured or semi-structured free text makes automated text analysis a difficult task.

[0003] With the development of text mining technology, techniques such as topic analysis and sentiment analysis have been widely used to extract information from various texts. In the field of aviation safety, topic modeling methods such as LDA (Latent Dirichlet Allocation) and STM (Structural Topic Model), leveraging their unsupervised nature, can identify the underlying topic structure of documents at a macro level, assisting in analyzing the primary causes of aviation accidents.

[0004] However, current topic modeling methods still have significant flaws. First, they lack the ability to understand complex texts. Relying solely on word importance to discover topics, they fail to effectively capture the complex syntactic and semantic relationships within the text, significantly reducing the accuracy of topic mining. Second, while topics can provide a macroscopic view of the causes of aviation accidents, they are merely a collection of semantically related risk factors and lack a quantitative description of the influencing factors. This makes it difficult to accurately capture the key factors of aviation accidents and reveal the underlying causes of the accidents. Therefore, a method that can accurately mine key information from aviation accident texts is urgently needed. Summary of the Invention

[0005] To address the aforementioned problems in the prior art, namely, the limited understanding capabilities of existing topic modeling methods in analyzing aviation safety accident texts and their inability to quantitatively reveal key causes, the first aspect of the present invention proposes an aviation safety accident report analysis method based on topic modeling and word co-occurrence networks, comprising the following steps: S100, obtaining the aviation safety accident report text to be processed and performing pre-processing; S200, based on a preset Semantic Fusion-Enhanced Topic Modeling Engine (SF-ETME), extract the topic structure and keywords of each pre-processed aviation safety accident report text, the method is as follows: S210: Determine keywords, topic distribution vectors, and sentence embedding vectors of the aviation safety accident report text, and combine them with an adaptive weighted fusion strategy to obtain a semantically fused enhanced feature vector; S220, using an improved deep embedding clustering model in combination with a contrastive learning method to perform feature learning and cluster analysis on the semantic fusion enhanced feature vector to obtain an optimal topic-word distribution result, thereby determining a topic structure; S300: Based on the extracted topic structure and keywords, an aviation safety accident word co-occurrence network (WCN) is constructed based on the dynamic cosine similarity threshold and network analysis is performed. The importance of each risk node is measured through the node centrality index, and the key risk factors are determined and output as analysis results.

[0006] In some preferred embodiments, the pretreatment is performed by: Based on a pre-built corpus, a standardized process for preprocessing text in the aviation field is constructed, and the aviation safety accident report text is preprocessed according to the standardized process for preprocessing text in the aviation field; wherein the corpus is constructed based on the NTSB aviation safety incident report.

[0007] In some preferred embodiments, the aviation safety accident report text is preprocessed according to a standardized process for text preprocessing in the aviation field, and the method is as follows: S110, performing text cleaning on the aviation safety accident report text, extracting the NTSB aviation accident comprehensive data table from the corpus, and performing association based on the primary key accident number; S120. Based on the combined results, restore the abbreviations of commonly used aviation terms to their full forms; S130. Based on the results of abbreviation restoration, convert English phrases commonly used in the aviation field into multi-word expressions, and use a word segmenter to segment the aviation safety accident text; S140, restoring the word to basic word units based on the word segmentation result; S150. After the restoration is completed, remove the words that appear in all documents in the corpus, including general stop words and stop words in the aviation field.

[0008] In some preferred embodiments, the semantic fusion enhanced feature representation is obtained by: S211, respectively determining the word frequency - inverse document frequency of each of the pre-processed aviation safety accident report texts, and then determining keywords; S212: Input the number of topics K. For a given d-th document containing K topics, based on the pre-set topic popularity covariate X of aviation safety accident reports, i , using the logistic normal model to generate the document-topic prior distribution vector, calculate the probability distribution of each topic in the text, and obtain the probability topic distribution vector; S213. For a given d-th document, generate a sentence embedding vector using the BERT pre-trained model; S214. Based on text features, an adaptive weighted fusion strategy is adopted to dynamically fuse the probabilistic topic distribution vector and the sentence embedding vector to generate a semantically fused enhanced feature vector.

[0009] In some preferred embodiments, a semantic fusion enhanced feature vector is generated by: Calculate the topic specificity H of text features i : Among them, p i,k is the probability of document i on topic k, K is the number of topics; Calculate the content complexity L of text features i : Among them, len(text i ) is the number of word segments in document l, and N is the total number of documents in the corpus; Calculate the semantic clarity C of text features i : in, is the BERT encoding vector of document i, cos is the cosine similarity calculation; The H i , L i 、C i After standardization, weighted fusion is performed to obtain the dynamic fusion weight s i : in w1>w2>w3; The s i Dynamic fusion coefficient α mapped to 0-1 i : The probability topic distribution vector and the sentence embedding vector are weightedly concatenated based on the dynamic fusion coefficient to obtain a semantic fusion enhanced feature vector. Where γ is the STM basic magnification factor, θ d ∈R K is the STM probability topic distribution vector, α d ∈R D BERT sentence embedding vector.

[0010] In some preferred embodiments, the potential topic structure is determined by: S221, constructing a candidate interval of the number of topics K according to a pre-set density constraint of the word co-occurrence network; S222. For each candidate topic number in the candidate interval, calculate the corresponding topic model evaluation index and determine the topic-word distribution result; S223. Based on the model evaluation index of each candidate topic number, a three-axis line comparison chart of topic number-topic model evaluation index is drawn to determine the optimal topic number K. ＊ , the optimal number of topics K ＊ The corresponding topic-word distribution results are used as the potential topic structure.

[0011] In some preferred embodiments, based on the number of candidate topics, the corresponding topic model evaluation index is calculated, and the topic-word distribution result is determined, and the method is as follows: B1. Determine the semantic fusion enhanced feature representation under the number of candidate topics and learn the feature vector based on the autoencoder Constructing a low-dimensional latent space representation of semantically fused enhanced feature vectors; B2. Use the K-means clustering algorithm to cluster the feature vectors after dimensionality reduction, determine the initial cluster center, calculate the soft assignment, and sharpen the soft assignment probability; B3. Use a pseudo-label strategy to mark the positive and negative sample pairs after sharpening the clustering results, construct a loss function for the joint optimization objective, introduce a contrastive learning loss term to improve the loss function of the deep embedding model, train the deep embedding clustering model, and optimize the clustering results; B4. Use word frequency-inverse document frequency to extract the topic words of each cluster and obtain the topic-word distribution result; B5. Calculate topic model evaluation indicators based on the topic-word distribution results; the topic model evaluation indicators include topic consistency, exclusivity index and silhouette coefficient.

[0012] In some preferred embodiments, the loss function of the joint optimization objective is constructed as follows: The training loss functions include reconstruction loss, clustering loss and contrast loss. The weight combination is adjusted through random grid search to determine the weight parameters of each loss function. The parameter configuration that maximizes the silhouette coefficient and clustering purity of the clustering result is selected as the final weight parameter.

[0013] In some preferred embodiments, the key risk factors in the field of aviation safety are determined by: S301: Constructing a word-document co-occurrence matrix based on the topic structure; wherein the rows of the matrix represent documents and the columns represent risk factors; S302: Calculating the cosine similarity between different risk factors based on the word-document co-occurrence matrix; S303: Determine a dynamic cosine similarity threshold, obtain an adjacency matrix, and then construct a generated word co-occurrence network using risk factors as nodes; S304: Configuring attributes for each node in the generated word co-occurrence network; S305: Analyze the constructed risk factor network using a network analysis method, calculate node centrality indicators, and then determine the key factors leading to aviation safety accidents; the node centrality indicators include degree centrality, eigenvector centrality, and betweenness centrality of the node.

[0014] A second aspect of the present invention provides an aviation safety accident report analysis system based on topic modeling and word co-occurrence network, the system comprising: A data acquisition module configured to acquire the aviation safety accident report text to be processed and perform pre-processing; A model building module configured to build a semantic fusion enhanced topic modeling engine; The data extraction module is configured to extract the topic structure and keywords of each aviation safety accident report text after preprocessing based on a preset semantic fusion enhanced topic modeling engine; The data analysis module is configured to construct an aviation safety accident word co-occurrence network based on the extracted topic structure and keywords based on the dynamic cosine similarity threshold and perform network analysis, thereby determining key risk factors and outputting them as analysis results; The method for extracting the subject structure and keywords of each aviation safety accident report text after preprocessing is as follows: Determining keywords, topic distribution vectors, and sentence embedding vectors of the aviation safety accident report text, and combining them with an adaptive weighted fusion strategy to obtain a semantically fused enhanced feature vector; An improved deep embedding clustering model is used in combination with a contrastive learning method to perform feature learning and cluster analysis on the semantic fusion enhanced feature vector to obtain the optimal topic-word distribution result and then determine the topic structure.

[0015] Beneficial effects of the present invention: 1. This invention combines text mining, topic modeling, and network analysis technologies. Through the dynamic fusion of structured topic mining and deep semantic extraction, it efficiently captures the topic structure of documents, reveals the correlation between risk factors affecting aviation safety accidents, and can identify the contextual semantic information and subtle semantic differences of words, thereby identifying the key causes of aviation safety incidents. This greatly improves the accuracy of topic modeling, overcomes the limitations of traditional methods in capturing deep semantic information and subtle semantic differences in text, and enables in-depth analysis and risk identification of aviation safety accident reports. 2. Considering the characteristics of aviation accident reports, such as the excessive number of abbreviations and fixed phrases, the present invention improves the conventional processing flow by adding steps such as abbreviation restoration, multi-word expression (MWE) segmentation, and lemma restoration. The order of each step is also rearranged, achieving efficient and accurate preprocessing of aviation accident texts. 3. This paper uses a probabilistic topic model (STM) to obtain topic distribution vectors for texts, and extracts sentence embedding vectors using a deep language model (BERT). Combined with an adaptive weighted fusion strategy, this method achieves a semantically fused and enhanced feature representation. It also employs an improved deep embedding clustering model, combined with contrastive learning to achieve coordinated optimization of the feature space and clustering structure, thereby accurately extracting the underlying topic structure from aviation accident reports. This method significantly outperforms traditional topic modeling methods in metrics such as silhouette coefficient, semantic consistency, and exclusivity, and possesses superior capabilities for capturing complex semantics and latent topics. 4. The method organically combines topic modeling with word co-occurrence networks. Through a word co-occurrence network based on a dynamic cosine similarity threshold, it displays the intricate connections between aviation accident risk factors. It uses node centrality indicators to measure the importance of each risk node and identify the key causes of aviation accidents. This overcomes the defect of traditional topic modeling methods that it is difficult to deeply and quantitatively mine aviation accident risk factors, and can effectively capture the key factors affecting aviation accidents. It can comprehensively and process-basedly analyze aviation accident reports, obtain the key factors affecting aviation accidents, and provide decision-making support for aviation safety management and accident prevention, thereby reducing accident risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Other features, objects and advantages of the present application will become more apparent upon reading the detailed description of non-limiting embodiments made with reference to the following drawings: Figure 1This is a structural diagram of the semantic fusion enhanced topic modeling engine of the present invention; Figure 2 is a flow chart of an aviation safety accident report analysis method based on topic modeling and word co-occurrence network in a first embodiment of the present invention; Figure 3 is a flow chart of NTSB data extraction according to the first embodiment of the present invention; Figure 4 This is a three-axis line comparison chart of the number of topics and the evaluation index of the topic model in the first embodiment of the present invention; Figure 5 This is a result diagram of the aviation safety accident word co-occurrence network in the first embodiment of the present invention. DETAILED DESCRIPTION

[0017] The present application will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are intended only to illustrate the relevant invention and are not intended to limit the invention. It should also be noted that, for ease of description, only portions relevant to the relevant invention are shown in the accompanying drawings.

[0018] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0019] This paper provides an aviation accident report analysis method based on topic modeling and word co-occurrence networks. This innovative and dynamic integration of structured topic mining and deep learning semantic extraction accurately captures document topics through an improved deep embedding clustering model. While efficiently capturing document topic structure and inter-topic relationships, this method can also identify contextual semantic information and subtle semantic differences within words, significantly improving the accuracy of topic modeling. This method enables comprehensive and streamlined analysis of aviation accident reports, identifying key factors influencing aviation accidents and providing decision support for aviation safety management and accident prevention, thereby reducing accident risks.

[0020] The present invention provides an aviation safety accident report analysis method based on topic modeling and word co-occurrence network, comprising the following steps: S100, obtaining the aviation safety accident report text to be processed and performing pre-processing; S200, based on a preset semantic fusion enhanced topic modeling engine, extracting the topic structure and keywords of each pre-processed aviation safety accident report text, the method is as follows: S210: Determine keywords, topic distribution vectors, and sentence embedding vectors of the aviation safety accident report text, and combine them with an adaptive weighted fusion strategy to obtain a semantically fused enhanced feature vector; S220, using an improved deep embedding clustering model in combination with a contrastive learning method to perform feature learning and cluster analysis on the semantic fusion enhanced feature vector to obtain an optimal topic-word distribution result, thereby determining a topic structure; S300: Based on the extracted topic structure and keywords, an aviation safety accident word co-occurrence network is constructed and network analysis is performed based on the dynamic cosine similarity threshold. The importance of each risk node is measured through the node centrality index, and then the key risk factors are determined and output as analysis results.

[0021] In order to more clearly illustrate the aviation safety accident report analysis method based on topic modeling and word co-occurrence network of the present invention, the following is combined with Figure 1 Each step in the embodiment of the present invention is described in detail.

[0022] The aviation safety accident report analysis method based on topic modeling and word co-occurrence network of the first embodiment of the present invention is as follows: Figure 1 、 2 As shown, the method includes the following steps S100 to S400, each of which is described in detail as follows: S100. Obtain the pending aviation safety accident report text and perform pre-processing: Based on a pre-built corpus, a standardized process for preprocessing text in the aviation field is constructed, and the aviation safety accident report text is preprocessed according to the standardized process for preprocessing text in the aviation field; wherein the corpus is constructed based on NTSB (National Transportation Safety Board) aviation safety incident reports.

[0023] Preferably, the aviation safety accident report text is preprocessed according to the standardized process of text preprocessing in the aviation field, and the aviation text is processed in the following order: text cleaning - abbreviation restoration - multi-word expression segmentation - morphological restoration - stop word removal; the specific method is: S110, text cleaning: performing text cleaning on the aviation safety accident report text, and extracting the NTSB aviation accident comprehensive data table from the corpus, including the accident narrative text "narratives" table, the accident "events" table, and the aircraft "aircraft" table, which are linked by the primary key "ev_id" of the accident number; S120, Abbreviation Restoration: Common aviation abbreviations are restored to their full form; for example, "apch" will be restored to "approach"; S130, Multi-Word Expression (MWE) tokenization: This converts commonly used English phrases in the aviation field into multi-word expressions, converting event narratives from long sentences into lists of words to prepare for text vectorization. The MWETokenizer tokenizer is used to tokenize aviation safety incident text, for example, converting "fuel tank" to "fuel_tank" and "nose landing gear" to "nose_landing_gear." The tokenizer treats multi-word expressions as a whole. S140, lemmatization: Based on the segmentation results, words are restored to their basic lemmas to merge similar words, such as different tenses of the same verb. This has the same goal as converting words to lowercase; for example, "better" will be restored to "good"; S150, stop word removal: After the restoration is completed, remove the words that appear in all documents in the corpus, including general stop words and stop words in the aviation field, such as "aircraft", "airplane", "airport", "flight", etc.

[0024] Lemmatization is the process of determining the word form based on the intended meaning of a word. Unlike stemming, lemmatization results depend on the word's part of speech and meaning within a sentence, making it more effective than stemming.

[0025] In this embodiment, relevant tables are extracted from the NTSB data set and linked by the primary key accident number "ev_id", such as Figure 3 As shown. In the narrative table "narratives", the narr_accf field describes the entire incident in detail in raw text format. In the event table "events", the ev_type field identifies whether the incident is an accident (an accident refers to a situation in which people are killed or seriously injured, or the aircraft suffers significant damage. Events other than accidents refer to situations related to aircraft operations that may affect operational safety). The ev_highest_injury field identifies the number of casualties in the incident. Use ev_id as the primary key to connect the narrative table, event table, and aircraft table, and limit the research object to fixed-wing powered aircraft. Delete records with an empty ev_type to complete the extraction of NTSB data.

[0026] Further preferably, in this embodiment, when performing text cleaning, a string processing tool is used to delete all punctuation marks and numbers except the hyphen "-" in the text; in addition, all English letters are converted to lowercase format.

[0027] Conventional text preprocessing involves cleaning text, case conversion, word segmentation, stop word filtering, and stemming. Taking into account the characteristics of aviation accident reports, such as the excessive number of abbreviations and fixed phrases, this paper improves on this conventional process by adding steps such as abbreviation restoration, multi-word expression (MWE) word segmentation, and lemmatization. The order of these steps is also rearranged, achieving efficient and accurate preprocessing of aviation accident text.

[0028] S200, based on a preset semantic fusion enhanced topic modeling engine, including a first-stage semantic fusion enhanced feature extraction and a second-stage semantic fusion topic modeling, to extract the topic structure and keywords of each pre-processed aviation safety accident report text.

[0029] Preferably, the first stage semantic fusion enhanced feature extraction includes two parts: structured topic mining and deep semantic extraction, which respectively obtain the probabilistic topic distribution vector and sentence embedding vector of the text, and generate dynamic fusion weights considering the topic clarity, content complexity and semantic clarity of the text.

[0030] Preferably, the first stage of the SF-ETME method is to determine the keywords, topic distribution vectors and sentence embedding vectors of the aviation safety accident report text, and combine them with an adaptive weighted fusion strategy to obtain a semantically fused enhanced feature vector, the method of which is: S211, respectively determining the word frequency - inverse document frequency of each of the pre-processed aviation safety accident report texts, and then determining keywords; S212: Input the number of topics K. For a given d-th document containing K topics, based on the pre-set topic popularity covariate X of aviation safety accident reports, i , using the logistic normal model to generate the document-topic prior distribution vector θ d , calculate the probability distribution of each topic in the text and obtain the probability topic distribution vector; θ d =LogisticNormal K-1 (Γ T X d T ,S); Where Γ=[γ1,...,γ K-1 ], k=1,...,K-1, is a hyperparameter, and parameter S is the covariance matrix; S213. For a given d-th document with K topics, use the BERT pre-trained model to generate a sentence embedding vector α d ; α d =BERT(d); Among them, BERT is a pre-trained model all-MiniLM-L6-v2 based on the Transformer architecture; S214. Based on text features, an adaptive weighted fusion strategy is adopted to dynamically fuse the probabilistic topic distribution vector and the sentence embedding vector to generate a semantically fused enhanced feature vector.

[0031] Further preferably, in this embodiment, the method for determining keywords is: A1. Calculate the term frequency (TF) of each word in the preprocessed aviation safety accident report text and normalize it to obtain a normalized term frequency vector. For each document, count the frequency of each word in the document to form a term frequency vector. The term frequency is the number of times a given word appears in the document. Among them, TF ij Indicates the frequency value of the i-th word in the j-th text, n ij Indicates the number of times the i-th word appears in the j-th text, represents the sum of the word frequencies of all words in the jth text; Considering that a small number of words with high word frequency will affect the overall proportion, the word frequency is normalized and then logarithmic processing is performed: A2. Calculate the inverse document frequency (IDF) of each key word in the information. The document frequency (DF) refers to the number of documents in which a word appears in the content. The inverse document frequency (IDF) is the inverse of the document frequency (DF). When i=1,2,...,K, calculate the corresponding IDF respectively i , that is, the inverse text frequency of each key word in the j-th text information is obtained; A3. Combining the normalized term frequency vector and the inverse document frequency, calculating the term frequency-inverse document frequency (TF-IDF). The TF-IDF value is obtained by multiplying the term frequency (TF) by the inverse document frequency (IDF). The larger the TF-IDF value of a term, the more important it is to the corresponding event text. TFIDF ij =TF ij ×IDF i ; A4. Filter out keywords from the aviation safety accident report text based on the TF-IDF value and a preset threshold or sorting rule.

[0032] Further preferably, the semantic fusion enhanced feature vector is generated by: Calculate the topic specificity H of text features i : Among them, p i,k is the probability of document i on topic k, K is the total number of topics; Long texts tend to require deeper semantic understanding, and the content complexity L of text features is calculated. i : Among them, len(text i ) is the number of word segments in document i, and N is the total number of documents in the corpus; Calculate the semantic clarity C of text features i : in, is the BERT encoding vector of document i, cos is the cosine similarity calculation, C i Higher values ​​indicate clearer semantic expression; The H i , L i 、C i After standardization, weighted fusion is performed to obtain the dynamic fusion weight s i : in w1>w2>w3; The s i Dynamic fusion coefficient α mapped to 0-1 i : The probability topic distribution vector and the sentence embedding vector are weightedly concatenated based on the dynamic fusion coefficient to obtain a semantic fusion enhanced feature vector. Where γ is the STM basic magnification factor, θ d ∈R K is the STM probability topic distribution vector, α d ∈R D BERT sentence embedding vector.

[0033] Preferably, in this embodiment, w1 = 0.5, w2 = 0.3, and w3 = 0.2.

[0034] The goal is to enhance the contribution of the STM module when topics are clearly concentrated, the text is short, and semantic confidence is low; and to increase the weight of the BERT module when the text is long, topics are dispersed, the text is long, and semantic confidence is high. This enables the model to automatically select the most effective feature representation for texts with different characteristics. Dynamically weighting the two vectors to fuse creates a semantically enhanced feature vector that comprehensively captures the deep semantic features of the topic and context.

[0035] Optimally, the second stage of semantic fusion topic modeling uses an improved deep embedding clustering model to simultaneously learn data feature representation and cluster assignment through a deep neural network, and combines contrastive learning to achieve collaborative optimization of the feature space and cluster structure. This method achieves the optimal topic-word distribution under the three-dimensional optimization of feature learning, clustering, and contrastive learning, thereby determining the hidden topic structure in the aviation safety accident report text collection.

[0036] Preferably, the second stage of the SF-ETME method: using an improved deep embedding clustering model, combined with a contrastive learning method, to perform feature learning and cluster analysis on the semantic fusion enhanced feature vector to obtain the optimal topic-word distribution result, and then determine the topic structure; the method is: S221, constructing a candidate interval of the number of topics K according to a pre-set density constraint of the word co-occurrence network; S222. For each candidate topic number in the candidate interval, calculate the corresponding topic model evaluation index and determine the topic-word distribution result; S223. Based on the model evaluation indicators of the number of candidate topics, a three-axis line comparison chart of the number of topics-semantic consistency, exclusivity, and silhouette coefficient is drawn to determine the optimal number of topics K. ＊ , the optimal number of topics K ＊ The corresponding topic-word distribution results are used as the potential topic structure.

[0037] Further preferably, the optimal number of topics K is determined ＊ When maintaining the semantic consistency within the topic as much as possible, the exclusivity between topics is also as large as possible, and the silhouette coefficient is as high as possible. According to the density requirements of the word co-occurrence network, a certain range is taken as the range of the number of topics K. For each K in the selected range of topics, topic modeling is performed, and then the best number of topics K is determined by comparing the indicators. ＊ .

[0038] Further preferably, in this embodiment, based on the number of candidate topics, the corresponding topic model evaluation index is calculated, and the topic-word distribution result is determined, and the method is: B1. Determine the semantic fusion enhanced feature representation under the number of candidate topics, pre-train the auto-encoder; learn the feature vector based on the auto-encoder Constructing a low-dimensional latent space representation of semantically fused enhanced feature vectors; B2. Initialize cluster centers: Use the K-means clustering algorithm to cluster the reduced eigenvectors, determine the initial cluster centers, calculate the soft assignments, and sharpen the soft assignment probabilities. B3. Use a pseudo-label strategy to mark the positive and negative sample pairs after sharpening the clustering results, construct a loss function for the joint optimization objective, introduce a contrastive learning loss term to improve the loss function of the deep embedding model, train the deep embedding clustering model, and optimize the clustering results; B4. Use TF-IDF to extract the topic words of each cluster and obtain the topic-word distribution results; B5. Calculate topic model evaluation indicators based on the topic-word distribution results; the topic model evaluation indicators include semantic coherence, exclusivity, and silhouette coefficient.

[0039] Specifically, in this embodiment, the auto-encoder is pre-trained as follows: let the input vector be n is the dimension of the input vector; Encoding process: The input vector x is mapped to a low-dimensional latent space through the encoder to obtain the encoded vector where m <n; z=f enc (x) = σ(W1x+b1); Among them, W1 and b1 are the weight matrix and bias vector of the encoder respectively, and σ is the activation function ReLU; Decoding process: The decoder restores the encoded vector z to the input space to obtain the reconstructed vector Among them, W2 and b2 are the weight matrix and bias vector of the decoder respectively, and σ is the activation function ReLU; Calculate the loss function: Define the loss function Measure the reconstruction vector The difference between it and the input vector x; By minimizing the loss function, the parameters of the encoder and decoder are optimized to make the reconstructed vector as close to the input vector as possible. And the trained autoencoder is used to complete the merged vector Dimensionality reduction: in

[0040] Specifically, in this embodiment, the K-means clustering algorithm is used to cluster the vector z after dimensionality reduction. d For clustering, the algorithm goal is to minimize the sum of squared errors from the sample point to the centroid of the cluster to which it belongs, that is, to optimize the following objective function: Among them, c i Represents data point z i The cluster to which it belongs; is cluster c i The center point of Represents the distance between a data point and the center point of its cluster, measured using Euclidean distance.

[0041] Specifically, soft distribution is calculated to measure the similarity between the sample and the cluster center. In this embodiment, the soft distribution is calculated based on Student's t distribution: Among them, q ij is the probability that sample i belongs to cluster j, and ||·|| is the Euclidean distance.

[0042] Specifically, in this embodiment, the soft assignment probability is sharpened to make it more concentrated; through the sharpening process, the model can more clearly identify the cluster center, and the calculation formula is as follows: where q' ij is the target distribution after sharpening.

[0043] Preferably, in this embodiment, the deep embedding clustering model is trained by: Assume that there are N samples in the current batch, and use the clustering soft assignment results to generate pseudo labels; the positive samples select samples in the same cluster as positive sample pairs, and all samples in the same cluster are positively correlated with each other (P i ={p|c p =c i}), samples from different clusters are automatically regarded as negative samples (N i ={n|c n ≠c i}); The pseudo label of each sample is The definition of the contrast loss function and the gradient update direction of the contrast loss term are as follows: Where τ is the temperature coefficient, τ=0.1, sim ij =z i ·z j / τ.

[0044] In this embodiment, the training loss function consists of three parts: reconstruction loss, clustering loss, and contrast loss. The loss function is as follows: The multi-loss weight parameter determination method α', β', and γ' adjusts the weight combination through random grid search to select the parameter configuration that maximizes the silhouette coefficient and cluster purity of the clustering result. In this embodiment, α'=0.5, β'=0.3, and γ'=0.2.

[0045] Preferably, the clustering loss uses KL divergence, and the calculation method and the gradient update direction of the clustering loss term are as follows: Among them, z i represents the representation of the i-th sample in the latent space, c j represents the position of the j-th cluster center in the latent space.

[0046] By introducing the contrastive learning loss term to improve the loss function of the deep embedding model, compared with the traditional deep embedding model, the improved model simultaneously promotes the aggregation of similar samples in the feature space during training, and keeps heterogeneous samples away from each other. It also adopts a pseudo-label driven strategy combined with a temperature scaling mechanism to solve the dependence of traditional contrastive learning on manual labels.

[0047] Preferably, for each cluster, the TF-IDF value of each word is calculated to extract the most representative topic words; assuming there are K topic clusters and V words; For the document set D in cluster k k , calculate the TF-IDF value of each word w in all documents in the cluster as follows: where |D k | is the number of documents in cluster k; the top five words with high TF-IDF values ​​in each cluster are extracted from the TF-IDF calculation results as topic words.

[0048] Preferably, the topic model evaluation index is calculated as follows: The degree of correlation between different words under the same topic. The higher the consistency of a topic, the closer the vocabulary of the topic is semantically. For a given topic t, the M words most related to the topic are expressed as: The theme consistency is: Where, Represents words in the document The word frequency, Represents words in a document and Number of co-occurrences; The exclusivity index FREX balances word frequency by calculating the weighted harmonic mean of word frequency and exclusivity. Words with high FREX values ​​tend to appear only in a single topic. The calculation method is as follows: Where ECDF is the cumulative empirical distribution function, and w is the preset probability; The silhouette coefficient focuses on the distribution of sample points within a cluster and the degree of separation between clusters. The value range of this indicator is [-1, 1]. The larger the value of the indicator, the more compact the cluster and the better the clustering performance. Conversely, the smaller the value, the worse the clustering performance. The silhouette coefficient S of each sample i The calculation is as follows: The silhouette coefficient S for all samples i Take the average value to get the overall silhouette coefficient of the cluster; where, for each sample i, a i is the average distance between it and other samples in the same cluster, b i is the average distance between it and all samples of the nearest cluster.

[0049] In this embodiment, the range of possible topic number K is 5 to 20, and steps S201 to S306 are repeated to obtain the model evaluation index under each topic number, and a three-axis broken line comparison of topic number-semantic consistency, exclusivity, and silhouette coefficient is drawn. Figure 4 As shown in the figure, it can be clearly seen that when the number of topics is 12, the modeling effect is the best.

[0050] S300: Based on the extracted topic structure and keywords, an aviation safety accident word co-occurrence network is constructed based on a dynamic cosine similarity threshold and network analysis is performed to determine key risk factors and output them as analysis results. The method is as follows: S301: Constructing a word-document co-occurrence matrix; constructing a word-document co-occurrence matrix based on the topic-word distribution results, retaining terms in each topic that can serve as risk factors; wherein the rows of the matrix represent documents, and the columns represent risk factors, and selecting TF-IDF values ​​as elements of the matrix. The TF-IDF value calculation method refers to steps A1-A4; S302: Calculate vector similarity; calculate the cosine similarity between different risk factors (column vectors) based on the word-document co-occurrence matrix; where a cosine similarity of 1 indicates complete similarity and 0 indicates no correlation. Assume that vector A is [A1, A2, ..., A n ], the vector B is [B1,B2,...,B n ], then the cosine similarity between vector A and vector B is: S303: Generate word co-occurrence network; calculate dynamic cosine similarity threshold θ' i , filter out the elements in the similarity matrix that are less than the threshold and obtain the adjacency matrix; take the risk factors as nodes, connect the nodes in the adjacency matrix that have co-occurrence relationships (matrix elements are 1) to generate a word co-occurrence network; S304: Configuring attributes for each node in the generated word co-occurrence network; wherein each node in the word co-occurrence network is labeled with the corresponding risk factor name; risk factor nodes under the same topic are assigned the same color, and the node size is determined by degree centrality; S305: Analyze the constructed risk factor network using a network analysis method, calculate node centrality indicators, and then determine the key factors leading to aviation safety accidents; the node centrality indicators include degree centrality, eigenvector centrality, and betweenness centrality of the node.

[0051] Further preferably, in this embodiment, the dynamic cosine similarity threshold θ'i is: μ′ i =mean(CosSim(w a ,w b )); σ′ i =std(CosSim(w a ,w b )); θ' i =0.5×w' i +0.5×μ' i +0.3×σ' i ; θ min =0.3≤θ i ≤θ max =0.8; Where w a ,w b ∈T i , that is, for each topic T i , obtain the cosine similarity between all risk factors (terms) and calculate the mean μ' iand standard deviation σ' i ;w' i is the weighted average of TF-IDF weights; θ min and θ max It is the upper and lower bounds of the threshold to ensure a reasonable filtering range.

[0052] Further preferably, node size is determined by degree centrality, which measures the degree of connection of a node with all other nodes in the network. For an undirected graph with g nodes, the degree centrality of node i is the total number of direct connections between i and other g-1 nodes, specifically: Where x ij Indicates the connection relationship between node i and node j. If node i and node j are directly connected, x ij =1, otherwise x ij =0.

[0053] In this embodiment, if Figure 5 The figure shows the node attribute configuration results of this embodiment, visually demonstrating the connections and importance of various risk factors. Nodes with the same theme are assigned the same color. Node size is proportional to degree centrality, which measures the degree of connectivity between a node and all other nodes in a network. Larger nodes, or nodes with stronger centrality, typically represent risk factors with significant influence and relevance within the risk network. The figure shows that "(equipment) maintenance" is the most important risk factor in aviation safety, associated with over half of the risk factors. "Wing failure" and "engine failure" also pose potentially serious risks, directly impacting aircraft flight performance and controllability. By identifying these nodes, aviation safety managers can understand which risk factors dominate the network and prioritize measures to mitigate these risks. Focusing resources and attention on nodes with strong centrality can help managers more effectively allocate limited resources and implement risk management measures, ensuring that the most critical and influential factors in the risk network receive timely attention and addressing, thereby reducing potential safety risks.

[0054] Betweenness Centrality measures the importance of a node as a bridge or key path in the network. A high betweenness centrality of a node indicates that it plays an important intermediary role in information transfer or communication between different nodes. Betweenness centrality is: Among them, sd(i,j,k) means that the shortest path from node j to node k passes through node i, that is, i is on the shortest path from j to k.

[0055] Eigenvector centrality assumes that the importance of a node depends on its degree of connection with other important nodes. Eigenvector centrality iteratively calculates the centrality value of a node so that the centrality of a node is proportional to the centrality of its neighboring nodes, specifically: Where c is a proportional constant, a ij =1 if and only if i and j are connected, otherwise 0.

[0056] The present invention organically combines topic modeling with word co-occurrence networks. It uses a word co-occurrence network based on a dynamic cosine similarity threshold to display the intricate connections between aviation accident risk factors. It uses a node centrality index to measure the importance of each risk node and identify the key causes of aviation accidents. This overcomes the defect of traditional topic modeling methods that it is difficult to deeply and quantitatively mine aviation accident risk factors, and can effectively capture the key factors affecting aviation accidents.

[0057] Although the various steps in the above embodiment are described in the above-mentioned order, those skilled in the art will understand that in order to achieve the effect of this embodiment, different steps do not have to be executed in such an order. They can be executed simultaneously (in parallel) or in a reverse order. These simple changes are within the scope of protection of the present invention.

[0058] The aviation safety accident report analysis system based on topic modeling and word co-occurrence network according to the second embodiment of the present invention includes: A data acquisition module configured to acquire the aviation safety accident report text to be processed and perform pre-processing; A model building module configured to build a semantic fusion enhanced topic modeling engine; The data extraction module is configured to extract the topic structure and keywords of each aviation safety accident report text after preprocessing based on a preset semantic fusion enhanced topic modeling engine; The data analysis module is configured to construct an aviation safety accident word co-occurrence network based on the extracted topic structure and keywords based on the dynamic cosine similarity threshold and perform network analysis, thereby determining key risk factors and outputting them as analysis results; The method for extracting the subject structure and keywords of each aviation safety accident report text after preprocessing is as follows: Determining keywords, topic distribution vectors, and sentence embedding vectors of the aviation safety accident report text, and combining them with an adaptive weighted fusion strategy to obtain a semantically fused enhanced feature vector; An improved deep embedding clustering model is used in combination with a contrastive learning method to perform feature learning and cluster analysis on the semantic fusion enhanced feature vector to obtain the optimal topic-word distribution result and then determine the topic structure.

[0059] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working process and related instructions of the system described above can refer to the corresponding process in the aforementioned method embodiment and will not be repeated here.

[0060] It should be noted that the aviation safety accident report analysis system based on topic modeling and word co-occurrence networks provided in the above embodiment is merely an example of the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of the present invention can be further decomposed or combined. For example, the modules in the above embodiment can be combined into one module or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the present invention are merely for the purpose of distinguishing the modules or steps and are not to be considered as improper limitations of the present invention.

[0061] An electronic device according to a third embodiment of the present invention includes: at least one processor; and a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the above-mentioned aviation safety accident report analysis method based on topic modeling and word co-occurrence network.

[0062] A fourth embodiment of the present invention provides a computer-readable storage medium storing computer instructions, wherein the computer instructions are configured to be executed by a computer to implement the aforementioned aviation safety accident report analysis method based on topic modeling and word co-occurrence network.

[0063] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes and related instructions of the electronic device and computer-readable storage medium described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0064] Those skilled in the art should be able to appreciate that, in conjunction with the modules and method steps of each example described in the embodiments disclosed herein, it is possible to implement them with electronic hardware, computer software, or a combination of the two, and the programs corresponding to the software modules and method steps can be placed in random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art. In order to clearly illustrate the interchangeability of electronic hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0065] Computer program code for performing the operations of the present application may be written in one or more programming languages, or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., through the Internet using an Internet service provider).

[0066] The flow charts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present application. In this regard, each box in the flow chart or block diagram can represent a module, program segment or a part of code, and the module, program segment or a part of code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order than that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart can be implemented by a dedicated hardware-based system that performs the specified function or operation, or can be implemented by a combination of dedicated hardware and computer instructions.

[0067] The terms "first", "second", etc. are used to distinguish similar objects, rather than to describe or indicate a particular order or sequence.

[0068] The term "comprise" or any other similar term is intended to cover non-exclusive inclusion such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0069] Thus far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art may make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will fall within the scope of protection of the present invention.

Claims

1. A method for analyzing aviation safety accident reports based on topic modeling and word co-occurrence networks is characterized by: The method comprises the following steps: S100, obtaining the aviation safety accident report text to be processed and performing pre-processing; S200, based on a preset semantic fusion enhanced topic modeling engine, extracting the topic structure and keywords of each pre-processed aviation safety accident report text, the method is as follows: S210: Determine keywords, topic distribution vectors, and sentence embedding vectors of the aviation safety accident report text, and combine them with an adaptive weighted fusion strategy to obtain a semantically fused enhanced feature vector; S220, using an improved deep embedding clustering model in combination with a contrastive learning method to perform feature learning and cluster analysis on the semantic fusion enhanced feature vector to obtain an optimal topic-word distribution result, thereby determining a topic structure; S300: Construct an aviation safety accident word co-occurrence network based on the dynamic cosine similarity threshold for the extracted topic structure and keywords and perform network analysis. Measure the importance of each risk node through the node centrality index, determine the key risk factors, and output them as analysis results.

2. The aviation safety accident report analysis method based on topic modeling and word co-occurrence network according to claim 1 is characterized in that: The pretreatment method is as follows: Based on a pre-built corpus, a standardized process for preprocessing text in the aviation field is constructed, and the aviation safety accident report text is preprocessed according to the standardized process for preprocessing text in the aviation field; wherein the corpus is constructed based on the NTSB aviation safety incident report.

3. The aviation safety accident report analysis method based on topic modeling and word co-occurrence network according to claim 2 is characterized in that: The aviation safety accident report text is preprocessed according to the standardized process of text preprocessing in the aviation field, and the method is as follows: S110, performing text cleaning on the aviation safety accident report text, extracting the NTSB aviation accident comprehensive data table from the corpus, and performing association based on the primary key accident number; S120. Based on the combined results, restore the abbreviations of commonly used aviation terms to their full forms; S130. Based on the results of abbreviation restoration, convert English phrases commonly used in the aviation field into multi-word expressions, and use a word segmenter to segment the aviation safety accident text; S140, restoring the word to basic word units based on the word segmentation result; S150. After the restoration is completed, remove the words that appear in all documents in the corpus, including general stop words and stop words in the aviation field.

4. The aviation safety accident report analysis method based on topic modeling and word co-occurrence network according to claim 1 is characterized in that: The semantic fusion enhanced feature representation is obtained by: S211, respectively determining the word frequency - inverse document frequency of each of the pre-processed aviation safety accident report texts, and then determining keywords; S212: Input the number of topics K. For a given d-th document containing K topics, based on the pre-set topic popularity covariate X of aviation safety accident reports, i , using the logistic normal model to generate the document-topic prior distribution vector, calculate the probability distribution of each topic in the text, and obtain the probability topic distribution vector; S213. For a given d-th document, generate a sentence embedding vector using the BERT pre-trained model; S214. Based on text features, an adaptive weighted fusion strategy is adopted to dynamically fuse the probabilistic topic distribution vector and the sentence embedding vector to generate a semantically fused enhanced feature vector.

5. The aviation safety accident report analysis method based on topic modeling and word co-occurrence network according to claim 4 is characterized in that: Generate semantic fusion enhanced feature vector, the method is: Calculate the topic specificity H of text features i : Among them, p i,k is the probability of document i on topic k, K is the number of topics; Calculate the content complexity L of text features i : Among them, len(text i ) is the number of word segments in document i, and N is the total number of documents in the corpus; Calculate the semantic clarity C of text features i : in, is the BERT encoding vector of document i, cos is the cosine similarity calculation; The H i , L i 、C i After standardization, weighted fusion is performed to obtain the dynamic fusion weight s i : in w1>w2>w3; The s i Dynamic fusion coefficient α mapped to 0-1 i : The probability topic distribution vector and the sentence embedding vector are weightedly concatenated based on the dynamic fusion coefficient to obtain a semantic fusion enhanced feature vector. Where γ is the STM basic magnification factor, θ d ∈R K is the STM probability topic distribution vector, α d ∈R D BERT sentence embedding vector.

6. The aviation safety accident report analysis method based on topic modeling and word co-occurrence network according to claim 1 is characterized in that: Identify potential topic structures by: S221. Constructing a candidate interval of the number of topics K according to a pre-set density constraint of the word co-occurrence network; S222. For each candidate topic number in the candidate interval, calculate the corresponding topic model evaluation index and determine the topic-word distribution result; S223. Based on the model evaluation index of each candidate topic number, a three-axis line comparison chart of topic number-topic model evaluation index is drawn to determine the optimal topic number K. ＊ , the optimal number of topics K ＊ The corresponding topic-word distribution results are used as the potential topic structure.

7. The aviation safety accident report analysis method based on topic modeling and word co-occurrence network according to claim 6 is characterized in that: Based on the number of candidate topics, the corresponding topic model evaluation index is calculated and the topic-word distribution result is determined as follows: B1. Determine the semantic fusion enhanced feature representation under the number of candidate topics and learn the feature vector based on the autoencoder Constructing a low-dimensional latent space representation of the semantically fused enhanced feature vector; B2. Use the K-means clustering algorithm to cluster the feature vectors after dimensionality reduction, determine the initial cluster center, calculate the soft assignment, and sharpen the soft assignment probability; B3. Use a pseudo-label strategy to mark the positive and negative sample pairs after sharpening the clustering results, construct a loss function for the joint optimization objective, introduce a contrastive learning loss term to improve the loss function of the deep embedding model, train the deep embedding clustering model, and optimize the clustering results; B4. Use word frequency-inverse document frequency to extract the topic words of each cluster and obtain the topic-word distribution result; B5. Calculate topic model evaluation indicators based on the topic-word distribution results; the topic model evaluation indicators include topic consistency, exclusivity index and silhouette coefficient.

8. The aviation safety accident report analysis method based on topic modeling and word co-occurrence network according to claim 7 is characterized in that: The loss function of the joint optimization objective is constructed as follows: The training loss functions include reconstruction loss, clustering loss and contrast loss. The weight combination is adjusted through random grid search to determine the weight parameters of each loss function. The parameter configuration that maximizes the silhouette coefficient and clustering purity of the clustering result is selected as the final weight parameter.

9. The aviation safety accident report analysis method based on topic modeling and word co-occurrence network according to claim 1 is characterized in that: Identify key risk factors in the field of aviation safety by: S301. Construct a word-document co-occurrence matrix based on the topic structure; wherein the rows of the matrix represent documents and the columns represent risk factors; S302, calculating the cosine similarity between different risk factors based on the word-document co-occurrence matrix; S303, determining a dynamic cosine similarity threshold, obtaining an adjacency matrix, and then constructing a word co-occurrence network with risk factors as nodes; S304, configuring attributes for each node in the generated word co-occurrence network; S305. Analyze the constructed risk factor network using a network analysis method, calculate node centrality indicators, and then determine the key factors leading to aviation safety accidents; the node centrality indicators include degree centrality, eigenvector centrality, and betweenness centrality of the node.

10. An aviation safety accident report analysis system based on topic modeling and word co-occurrence network, characterized by: A data acquisition module configured to acquire the aviation safety accident report text to be processed and perform pre-processing; A model building module configured to build a semantic fusion enhanced topic modeling engine; The data extraction module is configured to extract the topic structure and keywords of each aviation safety accident report text after preprocessing based on a preset semantic fusion enhanced topic modeling engine; The data analysis module is configured to construct an aviation safety accident word co-occurrence network based on the extracted topic structure and keywords based on the dynamic cosine similarity threshold and perform network analysis, thereby determining key risk factors and outputting them as analysis results; The method for extracting the subject structure and keywords of each aviation safety accident report text after preprocessing is as follows: Determining keywords, topic distribution vectors, and sentence embedding vectors of the aviation safety accident report text, and combining them with an adaptive weighted fusion strategy to obtain a semantically fused enhanced feature vector; An improved deep embedding clustering model is used in combination with a contrastive learning method to perform feature learning and cluster analysis on the semantic fusion enhanced feature vector to obtain the optimal topic-word distribution result and then determine the topic structure.

Images