An accident report extraction method and device based on word embedding template matching

By combining word embedding template matching with regular expressions and machine learning techniques, the problem of insufficient recall and precision in accident report extraction is solved, achieving a combination of high precision and high recall, which is suitable for extracting key attributes from accident reports.

CN118228727BActive Publication Date: 2026-07-21MINBO TECH (WUHAN) CO LTD +1
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
MINBO TECH (WUHAN) CO LTD
Filing Date
2024-04-08
Publication Date
2026-07-21

AI Technical Summary

Technical Problem

Existing accident report extraction methods suffer from low recall and insufficient precision when dealing with complex phrases, template matching methods cannot handle patterns that do not appear in the corpus, and machine learning methods require constant adjustments and have poor stability.

Method used

We employ a word embedding template matching method, which uses a pre-trained word vector generation model combined with regular expressions and machine learning techniques to first extract simple phrases, and then use cosine similarity and machine learning models to process complex phrases, achieving a combination of high precision and high recall.

Benefits of technology

While ensuring the accuracy of accident report extraction, it improves the recall rate, can handle category changes that template matching cannot handle, and reduces the randomness of model tuning.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118228727B_ABST
    Figure CN118228727B_ABST
Patent Text Reader

Abstract

The application discloses an accident report extraction method based on word embedding template matching, and comprises the following specific execution steps: S1. data acquisition and common sense preprocessing; S2. extracting simple phrases in the obtained data content; S3. extracting complex phrase analysis; a1. designing a candidate complex phrase table; a2. after cutting off irrelevant sentences, performing Chinese word segmentation processing to obtain a candidate word set; a3. obtaining a candidate word dictionary and embedding words into a vector space; a4. performing vector similarity calculation on word embedding of the dictionary words and word embedding of the candidate words, and reserving candidate words which are greater than a specified threshold value and have the highest K values; a5. processing the first K candidate word embeddings by using a machine learning method; and a6. finally discriminating the result after the candidate processing. The application has the advantages of high accuracy of template matching results when performing accident report extraction, and simultaneously makes up for the shortcomings of low recall rate of template matching by using the similarity of word vectors.
Need to check novelty before this filing date? Find Prior Art