Understanding Emotion Valence is a Joint Deep Learning Task

Gabriel Roccabruna, Seyed Mahed Mousavi, Giuseppe RiccardiACL 2023

TL;DR

This paper explores the natural inter-dependency between emotion valence and "emotion carriers" (people, objects, actions that convey emotion) in personal narratives. By framing this as a joint deep learning task, they demonstrate that multi-task learning significantly improves prediction accuracy for both valence and emotion carriers, especially for discriminative models.

What Problem Does This Paper Address?

Understanding the emotional state of individuals is crucial, particularly in contexts like mental health monitoring through personal narratives. While valence analysis (pleasantness/unpleasantness) captures the intensity of emotions, "emotion carriers" (ECs) provide the context and source of these emotions. For instance, "the office" might be an EC for a negative valence, while "vacation" might be for a positive one. The challenge lies in effectively modeling the inherent relationship between valence and ECs, as they are not independent but rather co-occur and inform each other. Previous work often treated these as separate tasks, potentially missing out on valuable inter-task information.

Methods

The authors investigate two multi-task learning (MTL) approaches: joint prediction (predicting valence and ECs simultaneously) and two-step prediction (using the output of one task as input for the other). They experiment with both discriminative (AlBERTo, a BERT-based model) and generative (GePpeTto, a GPT-2 based model) pre-trained language models (PLMs).

  • Discriminative Model (AlBERTo): For valence prediction, it's treated as a classification task. For EC prediction, it identifies EC candidate spans within a functional unit (FU) and classifies them as ECs or not. The joint setting uses a shared PLM backbone, while the two-step setting uses the output of the first task to condition the second.
  • Generative Model (GePpeTto): This model frames both tasks as causal language modeling. Specific prompts are designed to guide the model for valence prediction, EC prediction, and the two-step approach. Teacher forcing is used to stabilize training in the two-step setting.

Datasets

The study utilizes a corpus of personal narratives (PNs) in Italian, annotated for valence and emotion carriers at the functional unit (FU) level. This dataset, previously collected and annotated by Mousavi et al. (2022), consists of 481 narratives from 45 subjects. The authors use the official train (80%), validation (10%), and test (10%) splits provided by the original dataset creators.

Key Insights and Ideas

  • Inter-dependency is Key: The core insight is that valence and ECs are not independent. The presence of an EC often implies a non-neutral valence, and vice-versa. This inter-dependency can be leveraged for better prediction.
  • MTL Boosts Performance: Multi-task learning, whether joint or two-step, consistently improves performance over single-task baselines, demonstrating the benefit of exploiting the relationship between valence and EC prediction.
  • Discriminative Models Shine in Joint Settings: For the discriminative AlBERTo model, the joint prediction setting offers the best trade-off, achieving strong results on both tasks.
  • Generative Models Benefit from Two-Step: The generative GePpeTto model sees more significant improvements in the two-step prediction setting, particularly when domain adaptation is applied.
  • EC Prediction is Harder: EC prediction is generally more challenging due to unbalanced data distribution and the personal, context-specific nature of ECs. This explains why improvements in EC prediction often don't directly translate to valence prediction in the same way.

Key Results

  • Overall Improvement: Both joint and two-step MTL settings outperform single-task baselines for both valence and EC prediction.
  • Best Trade-off: The discriminative AlBERTo model in the joint prediction setting achieves the best overall trade-off between valence and EC prediction performance.
  • Highest Valence Prediction: The generative GePpeTto model, when fine-tuned with the single-task setting, achieves the best performance on valence prediction alone.
  • Highest EC Prediction: AlBERTo in the joint prediction setting achieves the highest macro F1-score on EC prediction.
  • Upper Bound: The "upper bound" experiment (using ground truth for the first step in the two-step setting) highlights the significant potential of the inter-dependency, showing substantial performance gains.

Limitations and Future Work

  • Dataset Limitations: The study is limited by the Italian language of the dataset and PLMs, potentially influencing language-specific properties. The sparsity of ECs and their specificity to narrators also pose a challenge, particularly for the EC → Val. two-step setting.
  • Future Directions:
    • Investigate bidirectional inter-dependency in the two-step setting (Val. ↔ EC).
  • Use a larger dataset with more narratives per narrator to better model the EC-Valence relationship.
  • Explore domain adaptation more extensively, especially for the generative model.

Summary updated

BibTeX
@inproceedings{roccabruna2023understanding,
  author = {Gabriel Roccabruna and Seyed Mahed Mousavi and Giuseppe Riccardi},
  title = {{Understanding Emotion Valence is a Joint Deep Learning Task}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2023},
}
Figure