The Role of Conversation Context for Sarcasm Detection in Online Interactions
Summary
Most sarcasm detectors judge a message on its own, but a speaker's ironic intent often becomes clear only from what came before. This paper asks two questions. First, does modeling the preceding turn improve sarcasm classification? Second, can we find which part of that context provoked the sarcastic reply?
The authors use two sources:
- discussion-forum replies from the Sarcasm Corpus V2, labeled by crowd workers;
- Twitter conversations, labeled by the authors' own hashtags such as #sarcasm.
They compare classification on the reply alone with classification on context plus reply. The baseline is an SVM with n-gram, lexicon and sarcasm-marker features, set against several LSTM variants. The conditional LSTM and LSTMs with sentence-level attention over context and reply beat an LSTM that reads only the reply.
The authors also compare attention weights with crowd workers' judgments. The models tend to focus on context sentences that are semantically coherent or incongruous with the reply, and on explicit markers such as emoticons. Earlier Twitter studies had found little gain from conversation context, so this result shows a clearer benefit. Data and code are publicly released.
Key points
- Forum data: drawn from a portion of the Internet Argument Corpus known as the Sarcasm Corpus V2. Of its 9,400 posts, only 4,692 (balanced between classes) are publicly available.
- Twitter data: over 200K tweets were collected, and only about 13% were replies to another tweet. The final set has 25,991 context–reply instances (12,215 sarcastic, 13,776 non-sarcastic), and 30% have more than one tweet of context.
- Baseline (SVM bl): a linear-kernel libSVM model using BoW n-grams, LIWC, MPQA and Opinion Lexicon sentiment features, and sarcasm indicators from Burgers et al. (2012).
- Best models: the conditional LSTM (Rocktäschel et al., 2015) and attention LSTMs at sentence level. Sentence-level attention was chosen for interpretability and because the forum corpus is fairly small.
- Attention analysis: compared against Turkers' selections of triggering sentences. The authors caution that attention-based interpretations should be treated carefully.
Cited by
- Detecting Sarcasm in Conversation Context Using Transformer-Based Models
- Affective and Contextual Embedding for Sarcasm Detection
- Context-Aware Sarcasm Detection Using BERT
- R^3: Reverse, Retrieve, and Rank for Sarcasm Generation with Commonsense Knowledge
- Transformer-based Context-aware Sarcasm Detection in Conversation Threads from Social Media
- Sarcasm Analysis Using Conversation Context
- Interpreting Verbal Irony: Linguistic Strategies and the Connection to the Type of Semantic Incongruity
- A Report on the 2020 Sarcasm Detection Shared Task
- Neural Sarcasm Detection using Conversation Context
- Multi-source Semantic Graph-based Multimodal Sarcasm Explanation Generation
- Comparative Study of Multilingual Idioms and Similes in Large Language Models
- When did you become so smart, oh wise one?! Sarcasm Explanation in Multi-modal Multi-party Dialogues
Cite
@inproceedings{ghosh2017role,
title = {{The Role of Conversation Context for Sarcasm Detection in Online Interactions}},
author = {Ghosh, Debanjan and Fabbri, Alexander Richard and Muresan, Smaranda},
booktitle = {Proceedings of the 18th Annual SIGdial Meeting on Discourse and Dialogue},
year = {2017},
doi = {10.18653/v1/w17-5523},
eprint = {1707.06226},
archivePrefix = {arXiv},
url = {https://doi.org/10.18653/v1/w17-5523}
}