[Paper Review] EmoBERTa: Speaker-Aware Emotion Recognition in Conversation with RoBERTa

Taewoon Kim, Piek Vossen|arXiv (Cornell University)|Aug 26, 2021

Topic ModelingComputer Science28 references48 citations

TL;DR

EmoBERTa augments RoBERTa with speaker tokens to model intra- and inter-speaker context for emotion recognition in conversation, achieving new state-of-the-art on MELD and IEMOCAP without changing RoBERTa architecture.

ABSTRACT

We present EmoBERTa: Speaker-Aware Emotion Recognition in Conversation with RoBERTa, a simple yet expressive scheme of solving the ERC (emotion recognition in conversation) task. By simply prepending speaker names to utterances and inserting separation tokens between the utterances in a dialogue, EmoBERTa can learn intra- and inter- speaker states and context to predict the emotion of a current speaker, in an end-to-end manner. Our experiments show that we reach a new state of the art on the two popular ERC datasets using a basic and straight-forward approach. We've open sourced our code and models at https://github.com/tae898/erc.

Motivation & Objective

• Motivate and address emotion recognition in conversation (ERC) using text data only.
• Leverage speaker identity to model dialogue dynamics without architecturally complex pipelines.
• Evaluate a simple, end-to-end RoBERTa-based approach on standard ERC benchmarks.
• Provide open-source code and pretrained models for reproducibility.

Proposed method

• Start from pretrained RoBERTa-large and add a linear classifier on the [CLS] token for sequence classification.
• Encode dialogue by constructing three segments: past utterances, current utterance, and future utterances, with each utterance prepended by the corresponding speaker name.
• Use RoBERTa to attend over the combined sequence and predict the emotion of the current utterance.
• Cap the input to RoBERTa's 512-token limit; iteratively prepend/append past/future utterances until token limit is reached.
• Train with cross-entropy loss plus L2 regularization; tune peak learning rate with Optuna and employ mixed-precision training.

Experimental results

Research questions

• RQ1 Can speaker-aware input construction improve ERC performance within a RoBERTa framework?
• RQ2 Does including past and/or future utterances benefit emotion prediction in ERC datasets?
• RQ3 How does explicit speaker information influence attention dynamics and final predictions?

Key findings

• EmoBERTa achieves state-of-the-art weighted F1 on MELD and IEMOCAP compared with prior methods.
• Ablation shows that including speaker names significantly boosts performance versus a RoBERTa baseline without speaker identifiers.
• Using only past utterances yields strong results on IEMOCAP; past+future generally helps more on MELD.
• Combining past and future utterances with speaker-aware input yields the best MELD result of 65.61 and IEMOCAP result of 67.42 when using both past and future (in one configuration).
• Qualitative analysis indicates the model's last-layer [CLS] aggregation tends to focus on the current speaker, validating the design.