📄 Recognition of Sarcasms in Tweets Based on Concept Level Sentiment Analysis and Supervised Learning Approaches

Summary

The paper tackles automatic detection of sarcasm in English tweets. Sarcasm is hard to detect because the literal wording and the intended meaning point in opposite directions. The authors collected 50,000 tweets with Twitter4J. Tweets tagged #sarcasm serve as positive examples, and tweets retrieved with random WordNet keywords serve as normal examples. Mentions, URLs and hashtags are removed, and the text is lemmatized. The core signal is a clash of positive and negative polarity inside one tweet, scored with SentiStrength and SenticNet. Words missing from those lexicons are expanded into related common-sense concepts, so that "raining" can contribute a negative "bad weather" reading. A polarity clash only counts as evidence when the sentences are coherent, which is checked with simple coreference-style heuristics. Punctuation and special symbols are also used as features. Two SVMs are trained, one on N-grams and one on these handcrafted features. Their outputs are merged by choosing the more confident verdict. The combination reaches 79.43% accuracy, ahead of both baselines. This shows that the semantic features catch cases that N-grams alone miss.

Key points

• Data: 50,000 tweets in a balanced set (sarcastic vs. normal), collected via Twitter4J. Labels come from the #sarcasm hashtag, so the authors admit the data is noisy. Preprocessing uses the Stanford Lemmatizer.
• Baseline 1 is a rule based purely on polarity contradiction, reaching 57.14% accuracy (F 0.56). Baseline 2 is an SVM on N-grams: unigrams give 73.81%, and uni+bi+trigrams give 76.40%.
• The proposed features alone reach only 63.42%. Unigrams combined with all proposed features give 76.35%, and uni/bi/trigrams combined with all proposed features give 79.43%, the best result.
• Ablations with unigrams show small effects from single components:
  • removing coherence gives 75.72%;
  • removing concept expansion gives 75.48%;
  • adding contradiction alone gives 72.83%;
  • adding sentiment score alone gives 75.64%.
• Each word missing from SentiStrength or SenticNet is expanded into five concepts without any sense selection. Ambiguity here, such as rain not always being bad, is a main error source. Future work names word sense disambiguation, better coherence modelling and evaluation on unbalanced, realistic data.

Interesting ideas

• I1. A polarity clash is treated as a sign of sarcasm only when the clashing parts are linked by coherence, for example a shared referent. Otherwise unrelated statements such as "he likes dogs, she hates cats" would be misread as ironic.
• I2. Words with no lexicon polarity inherit polarity from related common-sense concepts, as when a weather verb maps to "bad weather". The paper also shows that doing this blindly, without sense selection, introduces errors.
• I3. Two classifiers built on very different feature views (surface N-grams vs. semantic and pragmatic cues) are combined by trusting whichever one is more confident on each input. This beats either view alone.
• I4. Each component is evaluated by adding or removing it on top of a fixed base. The gains turn out modest, which honestly shows where the method really helps.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (frames): The irony detector in prag could require that the positive evaluation and the negative event attach to the same entity or event chain via coreference. This would make each irony verdict explainable as a specific linked pair.
• I2 (world): Compiled knowledge (WordNet/FrameNet/Wikidata) could supply default evaluative values for situations. These defaults should be marked defeasible and filtered by the sense chosen in context, not applied to every candidate concept.
• I3 (learn): Disagreement between witness parsers or teacher tiers could be resolved by per-instance confidence margins instead of fixed priority. The cases where only one view is right could then be logged as targets for repair rules.
• I4 (learn): Every new pragmatic or world-level module could ship with a "base vs. base+module vs. base−module" ablation report. This works as a negative control against modules that only look useful.