📄 An LSTM-Based Deep Learning Approach for Detecting Self-Deprecating Sarcasm in Textual Data

Summary

The paper addresses a narrow kind of sarcasm in which authors mock themselves, often with exaggerated claims of their own worthlessness. Brands increasingly use this style in marketing. The authors build a two-stage pipeline for English tweets. First, a filter keeps only tweets in which writers refer to themselves. Hand-written regular expressions catch explicit cases. Those explicit tweets are then clustered by shared tri-grams using Jaccard similarity, and each cluster's most frequent tri-gram becomes a pattern for finding implicit self-referential tweets. Second, the surviving tweets are mapped to pre-trained GloVe vectors and classified by an LSTM. Evaluation covers two benchmark corpora, Ptácek et al. (2014) and SemEval-2015, which were re-crawled from their tweet IDs, plus a new hashtag-based set called Twitter-280. Each corpus is tested in balanced and unbalanced versions. F-scores range from about 0.83 to 0.93. The LSTM generally beats the authors' earlier rule-plus-ML system. The authors present this as the first deep-learning treatment of self-deprecating sarcasm.

Key points

• Data: Ptácek et al. (2014) and SemEval-2015 were rebuilt from released tweet IDs, and some tweets could no longer be retrieved. Twitter-280 was collected from 1 April to 19 May 2019 via "#sarcasm", with non-sarcastic tweets gathered through #not and #hate.
• The explicit self-reference filter has two kinds of regex patterns:
  • specific ones, such as "(i | my) … love" or "when (my | our)";
  • generic ones, triggered by first-person singular or plural pronouns.
• Classifier: GloVe embeddings feed an LSTM with dropout of 0.4 after the LSTM layer, trained for 40 epochs.
• Results (P / R / F1):
  • Ptácek balanced: 0.93 / 0.94 / 0.93.
  • SemEval-2015 unbalanced: 0.93 / 0.75 / 0.83.
  • Twitter-280: F1 of 0.93 on the balanced version and 0.88 on the unbalanced one.
• Comparison with Abulaish and Kamal (2018): the LSTM wins on most settings, but the older method keeps a slight edge in precision and F-score on Ptácek.

Interesting ideas

• I1. A cheap pre-filter narrows the input to a relevant subpopulation before the expensive classifier runs, so the model only sees texts where the target phenomenon can plausibly occur.
• I2. Patterns are bootstrapped from high-confidence seeds: explicit matches are clustered, a recurring n-gram is mined from each cluster, and that n-gram is reused to catch implicit cases.
• I3. Self-reference is treated as a precondition of the phenomenon, so the question "who is the target of the remark" is answered separately from "is it sarcastic".
• I4. Every corpus is reported in both a balanced and an unbalanced version, which shows how recall drops when the target class is rare (SemEval recall falls to 0.75).

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 first require an experiencer or agent that corefers with the speaker, and only then test for evaluative mismatch.
• I2 (learn): Self-induced repair rules could be grown the same way, by clustering confident witness-agreed cases and extracting a shared pattern to apply to uncertain sentences.
• I3 (frames): Speaker-targeted irony could be modelled by checking whether the semantic-role holder of a positive emotion verb ("love", "enjoy") is the speaker, rather than by first-person regexes.
• I4 (learn): Evaluation of rare pragmatic labels such as irony or implicature could routinely include a skewed-distribution split, so that inflated balanced-set scores are exposed.