Automatic Detection of Irony

4Three Models for Automatic Irony Detection

4.1. Introduction

In the previous chapter, we analyzed a 2,000 tweet subset of the FrIC, making the following observations:

• – Irony in a tweet is indicated by the presence of two propositions P₁ and P₂ that contradict one another. We identified two types of contradiction: explicit, in cases where both P₁ and P₂ are present in lexical form in the tweet, and implicit, in cases where P₁ is present but P₂ must be inferred from external context. We showed that in 76.42% of cases in our corpus, irony took the form of an implicit contradiction, with only 23.58% of cases containing explicit contradictions.
• – The most common irony categories were oxymoron/paradox in the case of explicit contradictions, and false assertion in the case of implicit contradictions, with frequencies of 66% and 56%, respectively.
• – The most common linguistic cues found in tweets in our corpus included named entities, punctuation, opinion expressions, negation markers, personal pronouns and URLs. Furthermore, negation markers (no, never, won’t, etc.) are one of the most common cues found in both ironic and non-ironic tweets, with respective frequencies of 35% and 58%.

Based on these observations, we decided to develop a model for automatic detection of textual irony in tweets in cases of both explicit and implicit contradictions. Our model notably has the capacity to detect irony expressed through false assertions. Given the high numbers of negations in our corpus, ...