Performance Evaluation of Keyword Extraction Techniques and Stop Word Lists on Speech-To-Text Corpus

Abstract

The dawn of conversational user interfaces, through which humans communicate with computers through voice audio, has been reached. Therefore, Natural Language Processing (NLP) techniques are required to focus not only on text but also on audio speeches. Keyword Extraction is a technique to extract key phrases out of a document which can provide summaries of the document and be used in text classification. Existing keyword extraction techniques have commonly been used on only text/typed datasets. With the advent of text data from speech recognition engines which are less accurate than typed texts, the suitability of keyword extraction is questionable. This paper evaluates the suitability of conventional keyword extraction methods on a speech-to-text corpus. A new audio dataset for keyword extraction is collected using the World Wide Web (WWW) corpus. The performances of Rapid Automatic Keyword Extraction (RAKE) and TextRank are evaluated with different Stoplists on both the originally typed corpus and the corresponding Speech-To-Text (STT) corpus from the audio. Metrics of precision, recall, and F1 score was considered for the evaluation. From the obtained results, TextRank with the FOX Stoplist showed the highest performance on both the text and audio corpus, with F1 scores of 16.59% and 14.22%, respectively. Despite lagging behind text corpus, the recorded F1 score of the TextRank technique with audio corpus is significant enough for its adoption in audio conversation without much concern. However, the absence of punctuation during the STT affected the F1 score in all the techniques.