TL;DR
This paper presents a preliminary corpus-based investigation into oxymorons in written Italian, identifying 376 such expressions and analyzing their syntactic structures. The work aims to bridge a gap in NLP research on figurative language by providing a foundational dataset and methodology for detecting these complex, context-dependent linguistic phenomena.
What Problem Does This Paper Address?
While NLP has made strides in understanding more common figurative language like metaphors and idioms, less attention has been paid to figures of speech like oxymorons. Oxymorons, defined as a juxtaposition of contradictory terms for emphasis (e.g., "deafening silence"), are challenging for computational analysis due to their reliance on semantic opposition and often context-specific interpretation. This paper addresses the lack of systematic investigation into oxymorons in natural language, particularly in Italian, and proposes methods for their automatic identification. The authors aim to build a resource that can aid in the detection of humor, irony, and sarcasm, which often employ oxymoronic structures.
Methods
The authors employed a multi-step methodology to identify oxymorons in Italian corpora.
- Antonymous Pair Extraction: They started with a list of canonical English antonymous pairs, translated into Italian, and filtered for noun pairs. This yielded 17 core antonymous pairs.
- Lexical Expansion: To broaden the search, they considered not only direct antonyms but also related lexemes (synonyms, hyponyms) and different grammatical forms (adjectival, verbal) of the antonymous pairs. This was inspired by the concept of indirect oxymorons.
- Corpus Search: The expanded lists of lexemes were searched in two large contemporary written Italian corpora: Italian Web 2016 (itTenTen16) and CORIS.
- Syntactic Structure Analysis: The identified oxymorons were then classified based on their syntactic structure (e.g., Noun-Adjective, Adjective-Noun, Sentence, Noun-Preposition-Noun).
- Word Sketch Tool: The Sketch Engine's Word Sketch tool was used to analyze the collocational behavior of words, helping to identify novel oxymoronic constructions, particularly those involving sentential patterns.
- Complementary Method (Metalinguistic Search): A secondary method involved searching for the word "ossimoro" (oxymoron) itself in the corpus, as speakers sometimes comment on oxymorons metalinguistically. This helped uncover highly creative and context-dependent examples.
Datasets
The study utilized two primary corpora of contemporary written Italian:
- Italian Web 2016 (itTenTen16): A very large corpus (4.9 billion words).
- CORIS: A smaller corpus (150 million words).
The paper also introduces a newly compiled dataset of 376 Italian oxymorons, which is made available as an Excel file through the University of Bologna Institutional Research Repository (AMSActa).
Key Insights and Ideas
- Oxymorons are surprisingly common: The study identified 376 oxymorons, suggesting they are more prevalent in everyday language than previously assumed.
- Syntactic flexibility: Oxymorons exhibit a low degree of fixedness, with many combinations appearing in both noun-adjective orders (NA and AN), though with different frequencies.
- Antonymous pair distribution is uneven: Some antonymous pairs (like "hot-cold" or "silence-noise") generate a significantly higher number of oxymorons than others (like "bravery-fear" or "war-peace"). This might be influenced by semantic abstractness, sensorial concepts, and the entrenchment of common examples.
- Context is king: Many oxymorons, especially those identified through metalinguistic search, are highly context-dependent and difficult to predict solely from lexical or syntactic patterns.
- NLP challenges: The inherent semantic opposition and contextual nature of oxymorons pose significant challenges for automatic detection and extraction in NLP.
Key Results
The preliminary investigation resulted in a list of 376 Italian oxymorons.
- The vast majority (almost 80%) were found in the larger itTenTen16 corpus.
- Syntactic Structures: The most frequent structures were Noun-Adjective (NA) and Adjective-Noun (AN), accounting for a large portion of the identified oxymorons. Sentence-level oxymorons were also a significant category.
- Antonymous Pair Performance: The "hot-cold" pair yielded the most oxymorons (87), followed closely by "silence-noise" (85) and "happiness-unhappiness" (50).
- The study also identified less common syntactic structures, including Noun-Preposition-Noun and various verb-based constructions.
Limitations and Future Work
- Preliminary Nature: The authors acknowledge that the list of 376 oxymorons is preliminary and not exhaustive.
- Manual Effort: The current approach heavily relies on manual checking, highlighting the need for greater automation.
- Broader Antonymy: The study could be enriched by including more antonymous pairs.
- Beyond Antonymy: A significant portion of oxymorons might not be directly traceable to antonymous pairs and may involve synonymous or even identical elements, requiring different detection methods.
- Contextual and Creative Oxymorons: The paper suggests that future work should focus on developing methods for detecting highly creative and context-dependent oxymorons, particularly those found in social media.
- Jocular Oxymorons: The "jocular" nature of many oxymorons (e.g., "military intelligence") suggests further investigation is warranted.
- NLP Applications: Advancing oxymoron detection could have clear benefits for other NLP tasks like humor, irony, and sarcasm detection.
Summary updated