Comparative Analysis of Word Embedding Techniques for Proposition Extraction in Concept Map Mining

Abstract

A vital part of knowledge representation, concept map mining allows for the extraction of structured data from unstructured text. Among the most difficult tasks in this field is proposition extraction, which entails finding relevant connections between ideas. By converting words into high-dimensional vector representations that contain syntactic and semantic links, word embedding approaches have greatly enhanced natural language processing (NLP) applications. Various fields, including AI, computer science, biology, and engineering, are represented in the collection, which also includes instructional materials, Wikipedia pages, and scholarly books. In order to ensure that the dataset was consistent and of high quality, it was preprocessed before training began. We looked examined Word2Vec, GloVe, FastText, and BERT, four word embedding models, to see how well they could extract relevant propositions. Accuracy, recall, and F1-score were used to evaluate performance. First place went to BERT with the greatest F1-score, followed by FastText, GloVe, and Word2Ve, according to the findings. Training BERT took 600 seconds, but GloVe only needed 90 seconds, indicating that BERT needed much more computing resources. Finding the right method for proposition extraction in knowledge representation tasks may be challenging; this work sheds light on the trade-offs between accuracy and computing efficiency in embedding models.