Introduction
Emotion Cause Identification (ECI) is a critical task in sentiment analysis that aims to identify the underlying reasons for an expressed emotion within a text. For instance, in the sentence "The man was happy because the thief was caught," the task is to link the emotion "happy" to the cause "the thief was caught." Traditionally, ECI has been treated as a binary classification problem where each clause in a document is evaluated independently against a target emotion clause.
However, this independent prediction approach suffers from two significant limitations. First, it often ignores the spatial proximity between the cause and the emotion, despite strong empirical evidence that causes are typically located near the emotion expression. Second, by predicting labels in isolation, traditional models can produce logically inconsistent results, such as identifying multiple causes in a short document where only one is likely to exist. This paper presents PAE-DGL, a novel neural architecture designed to address these issues by incorporating relative position and global label dependencies.
Methodology
The proposed PAE-DGL model enhances standard Bi-LSTM and Attention frameworks through two primary innovations: Relative Position Augmented Embedding (PAE) and Reordered Prediction with Dynamic Global Labels (DGL).
Relative Position Augmented Embedding (PAE)
The PAE module moves beyond simple word embeddings by explicitly modeling the distance between a candidate clause and the emotion clause. Using an encode-decode structure, the model is supervised to reconstruct the relative position of a clause. This auxiliary task forces the neural network to learn clause representations that inherently contain spatial information, which is crucial for identifying causes that are contextually and spatially linked to emotions.
Reordered Prediction with Dynamic Global Labels (DGL)
The DGL mechanism shifts the prediction strategy from a sequential (start-to-finish) order to a proximity-based order. The model first evaluates clauses closest to the emotion clause, which statistically have the highest probability of being the cause.
As shown in the figure above, the model maintains a "Dynamic Global Label" vector that is updated after each prediction. If a clause is identified as a cause with high confidence, this information is propagated to subsequent predictions, effectively acting as a soft constraint to prevent the over-identification of causes in a single document.
Data Analysis and Rationale
The design of PAE-DGL is grounded in a thorough analysis of the benchmark Chinese emotion cause corpus. Two key observations drove the development:
- The Proximity Rule: Analysis shows that 54.45% of causes occur immediately before the emotion clause (position -1), and 23.58% are contained within the emotion clause itself (position 0). By reordering predictions to prioritize these positions, the model leverages the natural structure of causal expressions.
- Global Constraints: Approximately 97.20% of documents in the dataset contain exactly one cause. The DGL mechanism incorporates this global knowledge, allowing the model to "realize" when a cause has likely already been found, thereby improving precision.
Experimental Results
The PAE-DGL model was evaluated against several baselines and state-of-the-art methods. The results demonstrate a significant performance leap:
- F1-Score: The model achieved a state-of-the-art F1-score of 0.7242, a dramatic improvement over the standard Bi-LSTM baseline (0.2529).
- Precision: The model reached a precision of 0.7619, outperforming the previous best model (ConvMS-Memnet) by 5.43%. This high precision validates the effectiveness of the DGL mechanism in reducing false positives.
- Ablation Study: Removing the reordering and DGL components resulted in a nearly 5% drop in F1-score, confirming that the order of evidence processing is vital for ECI.
Conclusion
By transitioning from independent to reordered prediction, the PAE-DGL model successfully integrates spatial and global context into the Emotion Cause Identification task. The use of relative position as a supervisory signal and the implementation of dynamic global labels allow the model to respect the inherent constraints of human language. Future work will explore joint emotion and cause extraction and the application of these techniques to multi-lingual datasets.
Summary updated