[Paper Review] Neural model robustness for skill routing in large-scale conversational AI systems: A design choice exploration

Han Li, Sunghyun Park+6 more|arXiv (Cornell University)|Mar 4, 2021

Topic Modeling17 references9 citations

TL;DR

This paper investigates model robustness in skill routing for large-scale conversational AI systems, focusing on dynamic changes in skill subscriptions post-deployment. It demonstrates that data augmentation significantly improves robustness against hypothesis list changes—especially insertions—while Bi-LSTM architectures degrade performance under such shifts, even when mitigated by augmentation.

ABSTRACT

Current state-of-the-art large-scale conversational AI or intelligent digital assistant systems in industry comprises a set of components such as Automatic Speech Recognition (ASR) and Natural Language Understanding (NLU). For some of these systems that leverage a shared NLU ontology (e.g., a centralized intent/slot schema), there exists a separate skill routing component to correctly route a request to an appropriate skill, which is either a first-party or third-party application that actually executes on a user request. The skill routing component is needed as there are thousands of skills that can either subscribe to the same intent and/or subscribe to an intent under specific contextual conditions (e.g., device has a screen). Ensuring model robustness or resilience in the skill routing component is an important problem since skills may dynamically change their subscription in the ontology after the skill routing model has been deployed to production. We show how different modeling design choices impact the model robustness in the context of skill routing on a state-of-the-art commercial conversational AI system, specifically on the choices around data augmentation, model architecture, and optimization method. We show that applying data augmentation can be a very effective and practical way to drastically improve model robustness.

Motivation & Objective

• Address the challenge of maintaining model performance in skill routing systems when skills dynamically change their subscriptions to intents after model deployment.
• Investigate how different modeling design choices—data augmentation, model architecture, and optimization method—affect robustness under dynamic hypothesis list changes.
• Evaluate the impact of hypothesis removal and insertion on model performance to understand real-world robustness in production-scale conversational AI systems.
• Identify optimal design choices that ensure consistent model performance despite evolving skill ontologies in large-scale digital assistant platforms.

Proposed method

• Apply data augmentation techniques, including noise injection and hypothesis list perturbations, during training to simulate dynamic changes in skill subscriptions.
• Compare three model architectures: Bi-LSTM, Transformer-based attention, and a simple aggregation-based model, for encoding cross-hypothesis context in list-wise ranking.
• Evaluate two optimization methods: binary cross-entropy (BCE) and margin-based cross-entropy (MCE), using offline and online evaluation protocols.
• Conduct controlled micro-benchmarks by varying hypothesis removal and insertion ratios to isolate the impact of list dynamics on model robustness.
• Use offline evaluation on a held-out test set and online evaluation under simulated production conditions with changing hypothesis lists.
• Measure model accuracy under dynamic conditions to assess robustness, comparing performance degradation across different design choices.

Experimental results

Research questions

• RQ1 How does data augmentation affect model robustness in skill routing when the list of candidate skills (hypotheses) changes at inference time?
• RQ2 How do different model architectures—Bi-LSTM, attention-based, and aggregation-based—perform under dynamic hypothesis list changes?
• RQ3 Which optimization method—BCE or MCE—yields better robustness in the presence of dynamic skill subscriptions?
• RQ4 What is the relative impact of hypothesis removal versus insertion on model performance in online evaluation settings?
• RQ5 Can data augmentation effectively mitigate performance degradation caused by dynamic changes in skill routing candidate lists?

Key findings

• Data augmentation reduces online accuracy degradation to zero across all model variants, maintaining performance even under 10 inserted hypotheses.
• Bi-LSTM models suffer a 56.74% accuracy drop when 10 new hypotheses are inserted, indicating high sensitivity to list ordering and length changes.
• The attention-based model achieves the best online performance, outperforming Bi-LSTM and aggregation-based models under dynamic conditions.
• Binary cross-entropy (BCE) loss outperforms margin-based cross-entropy (MCE) in 4 out of 6 evaluation cases, suggesting better generalization for this task.
• Even with similar offline accuracy, Bi-LSTM models show significantly worse online robustness than attention-based models, highlighting architectural limitations.
• After applying data augmentation, all models achieve online accuracy comparable to offline performance, indicating improved generalization to dynamic input changes.