Magnets for Sarcasm: Making Sarcasm Detection Timely, Contextual and Very Personal

Abstract

Sarcasm is a pervasive phenomenon in social media, permitting the concise communication of meaning, affect and attitude. Concision requires wit to produce and wit to understand, which demands from each party knowledge of norms, context and a speaker's mindset. Insight into a speaker's psychological profile at the time of production is a valuable source of context for sarcasm detection. Using a neural architecture, we show significant gains in detection accuracy when knowledge of the speaker's mood at the time of production can be inferred. Our focus is on sarcasm detection on Twitter, and show that the mood exhibited by a speaker over tweets leading up to a new post is as useful a cue for sarcasm as the topical context of the post itself. The work opens the door to an empirical exploration not just of sarcasm in text but of the sarcastic state of mind.

Sarcasm Analysis Using Conversation Context

“…Last but not least, we empirically show that explicitly modeling the turns helps and provides better results than just concatenating the current turn and prior turn (and/or succeeding turn). This experimental result supports the conceptual claim that both we and Ghosh and Veale (2017) make that it is important to keep the C_TURN and the P_TURN (S_TURN) separate (e.g., modeled by different LSTMs), as the model is designed to recognize a possible inherent incongruity between them. This incongruity might become diffuse if the inputs are combined too soon (i.e., using one LSTM on combined current turn and context).…”

“…Independently, Ghosh and Veale (2017) have proposed a similar architecture based on Bi-LSTMs to detect sarcasm in Twitter. Unlike Ghosh and Veale (2017), our prior work used attention-based LSTMs that allowed us to investigate whether we can identify what part of the conversation context triggered the sarcastic reply, and showed results both on discussion forum data and Twitter. This paper substantially extends our prior work introduced in Ghosh, Fabbri, and Muresan (2017).…”

“…Parikh et al (2016) uses attention to decompose the RTE problem into subproblems that can be solved separately and (Sha et al 2016) presented an altered version ("re-read LSTM") of LSTM that is similar to word attention models of Rocktäschel et al (2016). Likewise, recently LSTMs are used in sarcasm detection research (Oraby et al 2017;Huang, Huang, and Chen 2017;Ghosh and Veale 2017). Oraby et al (2017) used LSTM models to identify sarcastic utterances (tweets and posts from the IAC v2 that are structured as rhetorical questions), Huang, Huang, and Chen (2017) applied LSTM for sense-disambiguation research (on the same dataset proposed by Ghosh, Guo, and Muresan (2015)), and Ghosh and Veale (2017) used bi-directional LSTMs to identify sarcastic tweets.…”

Reasoning with Sarcasm by Reading In-Between

“…), as such we omit comparisons for the sake of brevity and focus comparisons with recent neural models instead. Moreover, since our work focuses only on document-level sarcasm detection, we do not compare against models that use external information such as user profiles, context, personality information (Ghosh and Veale, 2017) or emoji-based distant supervision (Felbo et al, 2017). For our model, we report results on both multi-dimensional and single-dimensional intraattention.…”

“…(Ghosh and Veale, 2016) proposed a convolutional long-short-term memory network (CNN-LSTM-DNN) that achieves state-of-the-art performance. While our work focuses on document-only sarcasm detection, several notable works have proposed models that exploit personality information (Ghosh and Veale, 2017) and user context (Amir et al, 2016). Novel methods for sarcasm detection such as gaze / cognitive features (Mishra et al, 2016(Mishra et al, , 2017 have also been explored.…”