Task Selection and Assignment for Multi-Modal Multi-Task Dialogue Act Classification with Non-Stationary Multi-Armed Bandits

Xiangheng He, Junjie Chen, Björn W. Schuller

TL;DR

This paper introduces a dynamic training strategy for multi-task learning (MTL) using non-stationary multi-armed bandits (MAB) to solve Dialogue Act Classification (DAC). By using discounted Thompson Sampling to decide which auxiliary task to train on at any given moment, the model avoids "negative transfer" and significantly boosts performance on minority classes, outperforming both standard multi-task baselines and current state-of-the-art models.

What Problem Does This Paper Address?

Multi-task learning (MTL) is designed to help a "primary task" (like identifying if a speaker is asking a question or making a statement) by training it alongside "auxiliary tasks" (like recognizing the speaker's emotion). However, MTL has a "randomness" problem: most systems pick auxiliary tasks at random during training.

The authors point out two major hurdles:

  1. Task Selection: Not all auxiliary tasks are helpful. Some can actually confuse the model, a phenomenon known as negative transfer.
  2. Task Assignment: A task that is helpful at the start of training might become useless or harmful later on.

The paper specifically tackles these issues within Multi-modal Dialogue Act Classification (DAC), where the system must process both text and audio to understand a speaker's intention.

Methods

The authors propose a transformer-based architecture guided by an MAB Controller.

  • The Model: A standard Transformer encoder-decoder receives features from two modalities: text (via BERT) and audio (via VGGish). These are fused using self-attention.
  • The Bandit Strategy: They model task selection as a Multi-Armed Bandit problem. Each "arm" is a task. To handle the fact that a task's utility changes over time, they use Non-stationary MAB with discounted Thompson Sampling (TS) and Gaussian priors.
  • The Reward Mechanism: The "reward" for the bandit is defined as the improvement in validation performance after training on a specific task. If a task helps the model reach a new "best" validation score, it receives a high reward; if it doesn't, the reward is zero or negative.
  • Exploration vs. Exploitation: The algorithm uses a "discounted" factor ($\gamma$) to prioritize recent rewards over old ones, allowing the controller to "forget" old task utilities and adapt as the model matures.

Datasets

The study uses an extended version of the IEMOCAP (Interactive Emotional Dyadic Motion Capture) dataset.

  • Primary Task: Dialogue Act Classification (8-class and 12-class versions).
  • Auxiliary Tasks: Emotion classification (5-class), speaker classification (10-class), and three regression tasks for sentiment (Arousal, Valence, and Dominance).
  • Data Split: 70% training, 20% validation, and 10% testing.

Key Insights and Ideas

  • Task Utility is Dynamic: The "aha moment" is the visualization of task selection over time (Fig. 3). In the first stage, the model focuses almost entirely on the primary task. In the second stage, it explores everything. By the third and fourth stages, it settles on specific auxiliary tasks (like emotion classification and arousal/valence regression) while almost completely ignoring others (like speaker classification).
  • Random MTL Can Be Worse Than Single-Task: Surprisingly, the standard multi-task baseline (MT) performed significantly worse than the single-task baseline (ST). This proves that blindly adding tasks without a selection strategy can be actively harmful.
  • Stability for Minority Classes: One of the most important findings is that the MAB approach provides much higher stability for rare dialogue acts (the "minority classes"). While standard models often failed completely on these classes (F1 score of 0), the proposed method maintained consistent, decent performance.

Key Results

The proposed method outperformed both single-task (ST) and traditional multi-task (MT) baselines across all metrics:

Metric Single-Task (ST) Multi-Task (Random) Proposed (MAB)
UAR 0.618 0.522 0.667
F1 Score 0.604 0.499 0.655
Accuracy 0.675 0.608 0.683
  • SOTA Comparison: On the more difficult 12-class dataset, the MAB method achieved an F1 score of 0.6858, surpassing the previous state-of-the-art (0.6786).
  • Significance: The improvements were statistically significant with p-values < 0.05.

Limitations and Future Work

  • Generalizability: The authors acknowledge that while this works for DAC, they haven't yet tested if this MAB strategy holds up for primary tasks in completely different domains (e.g., computer vision or robotics).
  • Computational Overhead: While not explicitly detailed as a failure, MAB-based selection requires frequent validation checks to calculate rewards, which can increase total training time compared to simple random selection.
  • Future Directions: The authors suggest investigating how this strategy scales when the number of auxiliary tasks becomes very large.

Summary updated

BibTeX
@inproceedings{he2024task,
  author = {Xiangheng He and Junjie Chen and Björn W. Schuller},
  title = {{Task Selection and Assignment for Multi-Modal Multi-Task Dialogue Act Classification with Non-Stationary Multi-Armed Bandits}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2024},
  doi = {10.1109/icassp48485.2024.10447862},
}
Figure