End-to-End Emotion Semantic Parsing

Xiaotong Jiang, Zhongqing Wang, Guodong ZhouACL 2024 · DOI

Introduction

Traditional emotion detection in natural language processing (NLP) has long been treated as a categorical classification task. Models are typically trained to assign a single label—such as "Joy," "Anger," or "Sadness"—to a given text snippet. However, this approach overlooks the rich contextual nuances inherent in human emotion. To truly understand an emotional expression, one must identify not just the emotion itself, but also its holder (the person experiencing the emotion), its target (the entity the emotion is directed toward), and its trigger (the event or reason causing the emotion).

This paper introduces the task of emotion semantic parsing, which seeks to map these constituents into a structured representation. By moving from flat labels to hierarchical trees, the authors provide a framework for capturing the "who, what, and why" of emotional language.

The Abstract Semantic Tree (AST)

To represent the complex relationships between emotional components, the authors propose the Abstract Semantic Tree (AST). In this structure, the Trigger serves as the root node, reflecting its central role in the emotional event. The tree then branches into two primary categories:

  1. Feelings: This branch contains nodes for the specific Emotion label and the broader Sentiment (positive/negative).
  2. Entities: This branch identifies the Holder and the Target, extracting the relevant spans directly from the input text.
The Abstract Semantic Tree (AST) structure representing the emotional context of a sentence.

By organizing these elements hierarchically, the model can better resolve ambiguities, such as which entity is the source versus the object of an emotion in complex sentences.

Methodology: Generative Parsing and Task Decomposition

The proposed framework utilizes an end-to-end generative approach based on the Flan-T5 backbone. The model is trained to generate a linearized string representation of the AST, using brackets and tags to denote the tree structure.

Task Decomposition Strategy

A significant challenge in semantic parsing is the high complexity of the output space. Generating a full, multi-node tree in a single step is difficult for even advanced models. To address this, the authors developed a task decomposition strategy, a form of curriculum learning that breaks the parsing process into manageable sub-tasks:

  • Mono Parsing: The model first learns to identify individual nodes (e.g., extracting just the trigger or just the holder).
  • Pair Parsing: The model then learns to identify relationships between pairs of nodes (e.g., Trigger + Target or Emotion + Sentiment).
  • Re-composition: Finally, the model is fine-tuned on the complete AST generation task, leveraging the knowledge gained from the simpler sub-tasks.
The multi-stage task decomposition strategy used to train the generative model.

This strategy ensures that the model "understands the parts" before attempting to "build the whole," leading to more robust performance on complex sentence structures.

Experimental Evaluation

The researchers evaluated their approach on a unified benchmark aggregating three diverse datasets: electoral tweets (Mohammad et al., 2015), literature (Kim and Klinger, 2018), and news headlines (Bostan et al., 2020). The final dataset comprised 7,234 samples.

Key Results

The proposed model achieved a state-of-the-art Micro F1 score of 46.7, significantly outperforming several strong baselines:

  • Standard T5: 38.2 F1
  • BERT-CRF (Sequence Labeling): 33.1 F1

Ablation studies confirmed that the tree structure itself is a vital signal; removing the AST hierarchy and treating the task as simple sequence generation caused performance to drop to 40.7 F1. Furthermore, the "Mono Parsing" sub-task was identified as the most critical component of the decomposition strategy, providing the largest boost to final accuracy.

Discussion and Insights

The study yields several important insights for the field of affective computing:

  1. Structure Matters: Treating emotion constituents as a tree provides a "cognitive map" that helps models disambiguate roles in complex text.
  2. Generative Superiority: Natural language generation (NLG) frameworks are more flexible and effective for structural parsing than traditional discriminative sequence labeling.
  3. Fine-Tuning is Essential: Zero-shot evaluations of massive models like ChatGPT and LLaMA showed poor performance on this specific task, indicating that specialized structural training is still necessary for niche semantic parsing.

Conclusion

Emotion semantic parsing represents a significant step toward deeper machine understanding of human sentiment. By combining a structured AST representation with a generative model and a clever task decomposition strategy, this work sets a new standard for extracting the full context of emotional expressions. Future research will look toward cross-lingual applications and more efficient fine-tuning methods for large-scale deployment.

Summary updated

BibTeX
@inproceedings{jiang2024end,
  author = {Xiaotong Jiang and Zhongqing Wang and Guodong Zhou},
  title = {{End-to-End Emotion Semantic Parsing}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2024},
  doi = {10.18653/v1/2024.findings-acl.4},
}
Figure