Description de quelques procédés linguistiques de l’ironie, par le biais des tweets sur les transports en commun en français et en polonais

Aude Grezka, Małgorzata Niziołek, Davide BuscaldiStudia 2019 · DOI

Abstract

Irony is a complex phenomenon. Even though the numerous theories of irony differ from one another, they all seem to agree on the fact that irony implies certain contradiction / discrepancy / inconsistency between what is said and the reality. The purpose of this paper is to describe the language criteria of irony. The studies are contrastive character (French and Polish corpora of tweets about public transport). In the first phase of the study based on contrastive linguistic analysis, we will answer the following questions: What are the language criteria of irony? Are they identical for both languages? The research will help to expand the knowledge of automatic processing of natural language and, more precisely, in the analysis of text data and searching in texts.

Introduction

Irony represents a major "stumbling block" for sentiment analysis and Natural Language Processing (NLP). Because irony relies on a fundamental contradiction between literal meaning and reality—such as exclaiming "I love it!" when a train is an hour late—standard algorithms often misclassify biting criticism as glowing praise. This paper addresses two main challenges: defining the specific linguistic criteria that signal irony in the informal environment of Twitter and determining if these markers are consistent across French and Polish to aid in building multi-lingual detection tools.

Methodology

The researchers employed a contrastive linguistic analysis, performing a deep dive into how users construct ironic messages rather than relying solely on black-box machine learning. They categorized irony into two primary types:

  • Explicit Opposition: Contradictions contained within the text itself (e.g., "I love it when there are no more trains").
  • Implicit Opposition: Contradictions between the text and the external context (e.g., calling a notoriously dirty station "spotless").

Datasets and Corpus Construction

The study utilized two primary corpora focused on public transport, a domain where user frustration frequently manifests as irony:

  • French Corpus: Derived from the DEFT 2018 challenge, containing 76,732 tweets regarding transport in the Île-de-France region.
  • Polish Corpus: A custom-built dataset of 9,465 tweets collected between September 2017 and June 2018. It was gathered using keywords related to Polish transport authorities and the specific term for "delay."
Distribution of queries in the Polish transport corpus.

As shown in the data distribution above, the keyword "delay" (opóźnienie) was the most fertile ground for irony in the Polish dataset, appearing in 5,359 tweets. Interestingly, the authors initially considered a dataset focused on political figures like Vladimir Putin but found it lacked irony; users tended to be strictly polarized, leaving little room for the playful mockery found in transit complaints.

Linguistic Markers of Irony

The research identified several key "markers" used to signal irony across both languages:

  • Intensity Markers: One of the most productive ways to signal irony is through extreme intensity. Adverbs like trop (French) or bardziej (Polish) are frequently used to over-praise a negative situation, making the irony easier to detect.
  • One-Way Semantic Shift: Irony almost always moves from positive to negative. Users use "beautiful" to mean "ugly" or "great" to mean "terrible," but rarely use negative terms to ironically mean something positive.
  • Universal Prosody and Typography: Despite structural differences between French and Polish, markers like hashtags (#ironie, #sarkazm), emojis, and punctuation (irony quotes) are remarkably similar. Using ALL CAPS to simulate emotional exacerbation is also a common cross-linguistic trait.
  • The Context Trap: A significant portion of irony is "implicit." Without knowing that a specific train line is currently on strike or seeing an attached photo, even advanced text-only algorithms struggle to decode the message.

Conclusion and Future Work

The paper successfully identifies a shared toolkit of ironic markers in both languages. However, several hurdles remain for fully automated detection, including non-standard orthography (slang and abbreviations) and the need for pragmatic context. Future work will focus on developing dedicated polarity dictionaries for the Polish language and refining algorithms that can accurately weight these linguistic markers to detect irony in real-time.

Summary updated

Abstract

Irony is a complex phenomenon. Even though the numerous theories of irony differ from one another, they all seem to agree on the fact that irony implies certain contradiction / discrepancy / inconsistency between what is said and the reality. The purpose of this paper is to describe the language criteria of irony. The studies are contrastive character (French and Polish corpora of tweets about public transport). In the first phase of the study based on contrastive linguistic analysis, we will answer the following questions: What are the language criteria of irony? Are they identical for both languages? The research will help to expand the knowledge of automatic processing of natural language and, more precisely, in the analysis of text data and searching in texts.
BibTeX
@inproceedings{grezka2019description,
  author = {Aude Grezka and Małgorzata Niziołek and Davide Buscaldi},
  title = {{Description de quelques procédés linguistiques de l’ironie, par le biais des tweets sur les transports en commun en français et en polonais}},
  booktitle = {Studia},
  year = {2019},
  doi = {10.14746/strop.2019.461.003},
}
Figure