Introduction
Emotion Cause Analysis (ECA) is a critical task in Natural Language Processing aimed at identifying the stimulus behind an expressed emotion. Traditionally, this has been approached as a clause-level task (CECA), where entire sentences or clauses are identified as the cause. However, clauses often contain irrelevant information, leading to "noise" that can hinder downstream tasks like sentiment-aware dialogue systems. This paper proposes a shift toward Span-level Emotion Cause Analysis (SECA), focusing on identifying the precise boundaries of the cause within the text.
Methodology
The authors leverage BERT as a contextual encoder to generate token-level representations. The input is structured by concatenating the context and the emotion expression: [CLS] Context [SEP] Emotion [SEP]. Two distinct architectural heads are explored for boundary detection:
- Sequence Labeling: Tokens are classified using a B-I-O (Begin, Inside, Outside) tagging scheme. Three variants are tested: a simple Softmax layer, a Conditional Random Field (CRF) for global label dependencies, and a Gated Recurrent Unit (GRU) to capture sequential patterns.
- Position Identification (Pointer Network): This approach uses an attention mechanism to directly predict the start and end indices of the cause span, marking the first application of pointer networks to the ECA task.
Experimental Results
The models were evaluated on two benchmarks: the Chinese CHI dataset (multi-span) and the English ENG dataset (single-span).
Span-level Performance
The BERT-based models significantly outperformed previous baselines. On the ENG dataset, the BERT+Pointer model achieved an F1 score of 0.901, nearly 30 points higher than the previous state-of-the-art. On the more complex CHI dataset, the BERT+CRF model reached an F1 of 0.566, a massive leap over the 0.184 baseline.
Clause-level Performance
A key finding is that training on finer-grained span-level data actually improves performance on traditional clause-level benchmarks. On the CHI dataset, the proposed SECA models achieved a clause-level F1 of 0.867, surpassing dedicated CECA models (which hovered around 0.77–0.79). This suggests that pinpointing the "heart" of the cause makes the model more robust overall.
Key Insights and Limitations
- Finer Supervision is Better: Forcing the model to find the exact cause span improves its ability to identify the correct clause.
- The Pointer Network Advantage: The pointer network proved highly effective for single-span extraction in English, though its fixed-span configuration slightly limited its performance on multi-span Chinese data.
- Contextual Sensitivity: Performance was higher on English data, likely due to shorter context lengths reducing the search space for cause spans.
Conclusion
This study demonstrates that BERT-based boundary detection is highly effective for SECA. The pointer network shows particular promise for single-span extraction, while the CRF remains robust for multi-span scenarios. Future work will focus on improving multi-span detection in pointer networks and addressing the challenges of long-context boundary detection in complex datasets.
Summary updated