Keyword Pool Generation for Web Text Collecting: A Framework Integrating Sample and Semantic Information

Abstract

Keyword pools are used as search queries to collect web texts, largely determining the size and coverage of the samples and provide a data base for subsequent text mining. However, how to generate a refined keyword pool with high similarity and some expandability is a challenge. Currently, keyword pools for search queries aimed at collecting web texts either lack an objective generation method and evaluation system, or have a low utilization rate of sample semantic information. Therefore, this paper proposed a keyword generation framework that integrates sample and semantic information to construct a complete and objective keyword pool generation and evaluation system. The framework includes a data phase and a modeling phase, and its core is in the modeling phase, where both feature ranking and model performance are considered. A regression model about a topic vector and word vectors is constructed for the first time based on word embedding, and keyword pools are generated from the perspective of model performance. In addition, two keyword generation methods, Recursive Feature Introduction (RFI) and Recursive Feature Introduction and Elimination (RFIE), are also proposed in this paper. Different feature ranking algorithms, keyword generation methods and regression models are compared in the experiments. The results show that: (1) When using RFI to generate keywords, the regression model using ranked features has better prediction performance than the baseline model, and the number of generated keywords is refiner, and the prediction performance of the regression model using tree-based ranked features is significantly better than that of the one using SHAP-based ranked features. (2) The prediction performance of the regression model using RFI with tree-based ranked features is significantly better than that using Recursive Feature Elimination (RFE) with tree-based one. (3) All four regression models using RFI/RFE with SHAP- based/tree-based ranked features have significantly higher average similarity scores and cumulative advantages than the baseline model (the model using RFI with unranked features). (4) Light Gradient Boosting Machine (LGBM) using RFI with SHAP-based ranked features has significantly better prediction performance, higher average similarity scores, and cumulative advantages. In conclusion, our framework can generate a keyword pool that is more similar to the topic, and more refined and expandable, which provides certain research ideas for expanding the research sample size while ensuring the coverage of topics in web text collecting.