Entity perception of Two-Step-Matching framework for public opinions

Area of Science:

• Natural Language Processing
• Information Retrieval
• Computational Linguistics

Background:

• Accurately identifying target entities in ambiguous user comments is crucial for analyzing public opinions.
• Existing methods struggle with multiple entities mentioned in public discourse, hindering precise target identification.

Purpose of the Study:

• To propose and evaluate a novel Two-Step-Matching method for precise target entity identification from public comments.
• To enhance the accuracy of identifying specific entities within large volumes of public opinion data.

Main Methods:

• Utilized BiLSTM-CRF model for potential entity extraction and TF-IDF for characteristic word extraction.
• Implemented a two-step matching process involving Jaro-Winkler distance against a business directory and an industry-characteristic dictionary.
• Defined associated rate and accuracy rate as key performance indicators for matching accuracy.

Main Results:

• The Two-Step-Matching method achieved high associated rates (up to 0.93) and accuracy rates (up to 0.95) on three public health event datasets.
• Demonstrated an average enhancement of 32% in associated rate and 30% in accuracy rate compared to using only the initial matching step.
• The industry-characteristic dictionary significantly improved the precision of entity identification.

Conclusions:

• The proposed Two-Step-Matching framework effectively identifies precise target entities from ambiguous public comments.
• This method offers a robust solution for extracting specific information from large-scale public opinion data.
• The framework has significant implications for target identification in fields like public health surveillance and market research.