An Attention Based Multi-view Model for Sarcasm Cause Detection (Student Abstract)

Hejing Liu, Qiudan Li, Zaichuan Tang, Jie BaiAAAI 2021 · DOI

Abstract

While artificial intelligence has made significant strides in detecting sarcasm, identifying the specific triggers or "causes" behind sarcastic statements remains a complex challenge. This paper introduces a multi-view attention-based model designed for Sarcasm Cause Detection. By combining sentence-level semantic relevance with pair-level causal relationship learning, the model identifies which part of a conversation context prompted a sarcastic response. Evaluation on the SARC dataset shows that the proposed model, particularly when utilizing an attention-based merge of multiple semantic views, outperforms standard text classification baselines in both F1 and AUC metrics.

Introduction

Sarcasm is frequently used on social media to express implicit discontent. For sentiment analysis and customer feedback monitoring, simply identifying a comment as sarcastic is insufficient; understanding the root cause of the frustration is essential for actionable insights. The authors identify two primary challenges in this domain: topic specificity, where core triggers must be distinguished from general context, and deep semantics, which requires capturing complex causal logic between a sarcastic post and its preceding context.

Methodology

The proposed model addresses these challenges through a two-tiered architecture:

Sentence-Level Semantic Relevance

The model evaluates the relationship between a sarcastic post ($S$) and a context clause ($c_i$) using four distinct mathematical "views" derived from attentive Bi-LSTM vector representations:

  • Sub-abs ($|S - c|$): Measures absolute distance.
  • Sub-square ($(S - c)^2$): Accentuates larger differences to highlight semantic conflicts.
  • Hadamard product ($S \odot c$): Captures fine-grained word-to-word interactions.
  • Average: Provides a general combined semantic representation.

An attention mechanism is employed to dynamically weight these four measures, merging them into a single relevance representation ($R$).

Pair-Level Causal Relationship

To capture the broader logical flow, the model concatenates the sarcasm and context clause embeddings. This combined input is processed through another attentive Bi-LSTM to create a pair-level representation ($P$), treating the two sentences as a single unit of meaning. The final classification integrates individual sentence vectors, the multi-view relevance feature, and the pair-level representation.

Experimental Setup

The study utilized a refined version of the SARC (Self-Annotated Reddit Corpus) dataset. The authors extracted sarcasm responses and contexts from seven subreddits and employed three annotators to identify specific "cause" clauses. The resulting dataset contains 4,987 clause pairs, with a Fleiss’ kappa of 0.76, indicating high annotation reliability.

Results

The proposed multi-view model was compared against several baselines, including TF-IDF with SVM, CNN, and LSTM architectures.

Performance comparison of various methods on the SARC dataset.

As shown in the results table, the proposed model consistently outperforms the baselines. Key observations include:

  • Multi-View Efficacy: Merging the four semantic views via attention (F1: 0.7048) yielded better results than any single measure alone.
  • Conflict Weighting: The "Sub-square" measure outperformed simple subtraction, likely because it emphasizes semantic conflicts typical of sarcasm.
  • Contextual Integration: The full model, incorporating both sentence-level and pair-level causal learning, achieved the highest performance (F1: 0.7114, AUC: 0.6292), confirming that sarcasm causes are often embedded in the logical flow between speakers.

Conclusion

This research demonstrates that a multi-view approach effectively captures the nuanced triggers of sarcasm. While the model shows strong performance on Reddit data, future work could explore its generalizability to other platforms like Twitter or Facebook. Additionally, addressing implicit causes—where the trigger relies on external world knowledge not present in the text—remains an open challenge for the field.

Summary updated

BibTeX
@inproceedings{liu2021attention,
  author = {Hejing Liu and Qiudan Li and Zaichuan Tang and Jie Bai},
  title = {{An Attention Based Multi-view Model for Sarcasm Cause Detection (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i18.17913},
}
Figure