Abstract
Dialogue Act (DA) classification is a fundamental task in dialogue systems, identifying speaker intentions such as greetings, statements, or questions. This paper introduces the Adversarial Learning Dynamic Memory Network (ALDMN), a framework that reformulates DA classification as a Question-Answering (QA) task. The model utilizes a hierarchical pyramidal utterance encoder to efficiently capture contextual features and employs adversarial training to enhance robustness. Evaluated on the Switchboard Dialogue Act (SwDA) and MapTask corpora, ALDMN achieves state-of-the-art accuracy, demonstrating the effectiveness of memory-based reasoning and adversarial regularization in natural language understanding.
Introduction
Dialogue Act (DA) classification is critical for building intelligent conversational agents. However, two primary challenges persist: contextual dependency and model fragility. The meaning of an utterance is often contingent on the preceding conversation, requiring models to maintain long-term memory. Furthermore, deep learning models are frequently vulnerable to adversarial perturbations—minor input changes that can lead to misclassification. This is particularly problematic in spontaneous speech, which is often ungrammatical and noisy.
To address these issues, we propose the Adversarial Learning Dynamic Memory Network (ALDMN). By treating DA classification as a reasoning task over a memory of "facts" (previous utterances), the model can better disambiguate intent.
Methodology
The ALDMN framework consists of three primary components designed to improve reasoning and robustness.
QA Formulation and Dynamic Memory
Rather than treating DA classification as a standard multi-class problem, we cast it as a QA task: "What is the label of this utterance?" This allows the model to leverage the iterative reasoning capabilities of Dynamic Memory Networks (DMNs). The DMN processes a sequence of facts and a query to produce an answer, effectively modeling the flow of a conversation.
Hierarchical Pyramidal Utterance Encoder
To represent utterances, we implement a pyramidal Bi-directional Gated Recurrent Unit (BiGRU). Unlike standard RNNs, the pyramidal structure reduces the number of hidden states by half in each subsequent layer. This architecture allows the model to extract high-level semantic features more efficiently while significantly reducing the computational overhead for the subsequent attention mechanism.
Adversarial Training
To improve generalization and robustness, we incorporate adversarial training. During the training phase, small perturbations are added to the word embeddings. By training the model to correctly classify these "noisy" examples, we encourage it to learn more robust decision boundaries. This process acts as a powerful regularizer, helping the model distinguish between semantically similar words that might appear in similar contexts but carry different intent.
Experimental Results
The ALDMN model was evaluated on two benchmark datasets: the Switchboard Dialogue Act Corpus (SwDA), containing over 1,100 conversations with 42 labels, and the MapTask Corpus, a smaller dataset focused on directional instructions.
Performance and Robustness
ALDMN established new state-of-the-art results on both datasets, achieving 81.5% accuracy on SwDA and 68.5% on MapTask. These results surpass previous benchmarks such as UCI (79.9% on SwDA) and PDI (65.9% on MapTask).
A key finding was the significant impact of adversarial training. As shown in the accuracy curves, the inclusion of adversarial perturbations ("with-adv") consistently led to higher accuracy and faster convergence compared to standard training ("without-adv").
Insights on Model Architecture
Ablation studies revealed that a 2-layer pyramidal encoder provides the optimal balance between feature abstraction and overfitting. Additionally, the model demonstrated peak performance on utterances of approximately 40 words, suggesting it effectively balances the high information content of longer sentences with the noise inherent in very lengthy dialogue turns.
Conclusion
The ALDMN framework demonstrates that combining memory-based reasoning with adversarial training significantly improves Dialogue Act classification. While the model excels at capturing context, it still faces challenges with highly subjective categories, such as distinguishing between "Statement-Opinion" and "Statement-Non-Opinion." Future research will focus on refining the model's ability to handle such subjective nuances and scaling the framework to larger, more diverse conversational datasets.
Summary updated