Multimodal Contextual Dialogue Breakdown Detection for Conversational AI Models

Md Messal Monem Miah, Ulie Schnaithmann, Arushi Raghuvanshi, Youngseo SonNAACL 2024 · DOI

TL;DR

This paper introduces MultConDB, a novel multimodal model that leverages both audio and text to detect dialogue breakdowns in real-time conversational AI, achieving a significant F1 score of 69.27. This advancement is crucial for improving the robustness and user experience of AI systems in demanding industry settings like healthcare, where breakdowns can have serious consequences.

What Problem Does This Paper Address?

Conversational AI systems, especially in critical domains like healthcare, need to accurately detect "dialogue breakdowns" – moments where the conversation goes off track and requires intervention. These breakdowns can stem from various issues: noisy audio leading to speech recognition errors, unexpected user behavior, or complex conversational flows. Existing methods often struggle with the nuances of real-world scenarios, particularly in phone calls where latency is critical and explicit cues of frustration are rare. The paper highlights the challenge of detecting subtle breakdowns that arise from factors like tone, cadence, or even the AI's failure to follow specific protocols, which are difficult to capture with text-only analysis.

Methods

The core of the proposed solution is the MultConDB model, a multimodal architecture designed to integrate acoustic and textual information. It features:

  • Unimodal Encoder Branches: Separate pathways process audio (using Wav2Vec2) and text (using RoBERTa) to generate distinct utterance embeddings. These branches involve temporal convolutions and LSTM with attention to capture local and contextual information.
  • Multimodal Encoder Branch: This branch combines the unimodal embeddings using Bi-LSTMs and attention to create a fused representation.
  • Cross-Modal Transformer: A key innovation is the use of cross-modal attention blocks where audio and text features query each other, enabling a deeper exchange of information. This is followed by self-attention layers.
  • Fusion and Classification: The outputs from the multimodal transformer are pooled, concatenated, and passed through linear layers for final dialogue breakdown classification.

The model is designed for real-time processing, with a focus on balancing performance and inference speed.

Datasets

The study utilized a dataset of 1,689 phone call conversations collected from August 2023 between conversational AI agents and users (insurance company employees) for verifying insurance benefits. These calls were annotated with binary labels indicating dialogue breakdowns (requiring human intervention) or coherent conversation flows. An additional 94 calls from September 2023 were used to test the generalizability of the model.

Key Insights and Ideas

  • Multimodality is Key: The research strongly suggests that combining audio and text signals is significantly more effective for detecting dialogue breakdowns in phone calls than text-only approaches. The subtle cues in voice (noise, intonation, pauses) are critical.
  • Context is Crucial: The model's ability to incorporate contextual information from previous turns, both acoustically and textually, is vital for accurately identifying breakdowns.
  • Industry-Specific Challenges: The paper emphasizes that industry settings, particularly healthcare, present unique challenges due to the lack of explicit frustration cues and the high cost of errors, making accurate breakdown detection paramount.
  • Subtle Breakdown Causes: The MultConDB model demonstrates an ability to implicitly categorize breakdown causes (e.g., AI going silent, interrupting users, skipping actions) without explicit training labels, highlighting its nuanced understanding.

Key Results

  • State-of-the-Art Performance: MultConDB achieved an F1 score of 69.27, outperforming other models, including the multimodal Transformer (MulT A+T) by 15%.
  • Improved Detection Accuracy: Multimodal models generally showed higher F1 scores than text-only models.
  • False Positive Analysis: Multimodal models tended to predict breakdowns closer to the actual event, which is beneficial for timely human intervention.
  • Generalizability: The model demonstrated strong performance on a separate dataset from a different month and with updated AI agent models, indicating good generalizability.
  • Inference Speed: The model achieves an inference time of 0.06s per utterance, meeting real-time requirements.

Limitations and Future Work

  • Dataset Availability: Due to Protected Health Information (PHI) concerns, the dataset cannot be made publicly available.
  • Local Hosting: The models were explored in a locally hosted setup rather than API calls.
  • Context Window Size: While a context window of 5 was chosen to balance performance and inference time, further exploration of larger context windows might yield better results at the cost of speed.
  • Future Work: The authors hope their findings will contribute to the development of more reliable conversational AI models. The paper implicitly suggests further research into optimizing multimodal fusion techniques and exploring more sophisticated contextual modeling.

Summary updated

BibTeX
@inproceedings{miah2024multimodal,
  author = {Md Messal Monem Miah and Ulie Schnaithmann and Arushi Raghuvanshi and Youngseo Son},
  title = {{Multimodal Contextual Dialogue Breakdown Detection for Conversational AI Models}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
  doi = {10.18653/v1/2024.naacl-industry.25},
}
Figure