Resolving Ambiguity in Hedge Detection by Automatic Generation of Linguistic Rules

An understanding of natural language is key in order to robustly extract the linguistic features indicative of deceptive speech. Hedging is a key indicator of deceptive speech as it can indicate a speaker's lack of commitment in a conversation. Hedging is characterised by words and phrases that display a sense of vagueness or that lack precision, such as suppose, about. The identification of hedging terms in speech is a challenging task, due to the ambiguity of natural language, as a phrase can have multiple meanings. This paper proposes to automate the process of generating rules for hedge detection in transcripts produced by an automatic speech recognition system using explainable decision tree models trained on syntactic features. We have extracted syntactic features through dependency parsing to capture the grammatical relationship between hedging terms and their surrounding words based on meaning and context. We tested the effectiveness of our model on a dataset of conversational speech, for 75 different hedging terms, and achieved an F1 score of 0.88. The result of our automated process is comparable to existing solutions for hedge detection.