File Compliance Detection Using a Word2Vec-Based Semantic Similarity Framework

This paper explores the application of the Word2Vec model in document compliance detection, and evaluates the performance of Word2Vec in calculating compliance similarity between documents by comparing it with the traditional text analysis method TFIDF, the topic modeling method LDA, and the advanced deep learning model BERT. During the research, we collected and preprocessed a large amount of archival data from multiple sources, generated document vectors using Word2Vec, TFIDF, LDA, and BERT, and comprehensively evaluated the models through indicators such as cosine similarity, precision, recall, F1 score, and AUC. The experimental results show that the Word2Vec model performs well in capturing the semantic similarity of documents, especially when distinguishing between compliant and non-compliant document pairs. Specifically, on the legal document dataset, Word2Vec achieved an F1 score of 0.84, which is 12% higher than TFIDF. In addition, the AUC of Word2Vec on the internal audit report dataset reached 0.92, which is 5 percentage points higher than LDA. However, compared with BERT, Word2Vec is slightly inferior in processing complex semantics and technical terms; for example, in the financial report dataset, BERT's F1 score is 0.78, while Word2Vec is 0.75, a gap of 3%. Word2Vec has obvious advantages in efficiency and simplicity, and is suitable for application scenarios that require fast deployment and low computing resources. At the same time, its performance in specific fields also proves its effectiveness as a compliance detection tool. 翻译搜索复制