Introduction
Current Conversational Question Answering (CQA) systems frequently top leaderboards with impressive scores. However, research presented at ACL 2021 suggests these metrics may be artificially inflated. The authors argue that standard evaluation protocols rely on a "Teacher Forcing" mechanism that does not exist in real-world applications, leading to models that are brittle when deployed autonomously.
The Problem: Teacher Forcing and Filtering Bias
In CQA tasks, such as those in the QuAC dataset, systems must answer a sequence of questions based on a passage. Because questions often contain coreferences (e.g., "Where was he born?"), models must utilize the context of previous turns.
The standard academic evaluation protocol provides models with the ground-truth answers of all previous turns to help answer the current question. This is known as "Teacher Forcing." In a real-world chatbot, however, only the system's own previous (and potentially incorrect) predictions are available.
The authors identify a "filtering bias" resulting from this protocol: models may not be learning to understand context so much as learning to exclude text already identified as an answer in previous turns. When this "teacher" is removed, performance typically collapses.
Methodology
To address these issues, the authors propose a shift in both evaluation and training.
Evaluation without Teacher Forcing (w/o TF)
The authors introduce a "standalone" evaluation protocol where the model processes a conversation sequentially, relying solely on its own previous predictions as history markers.
Robust Training Protocols
To prepare models for standalone deployment, three training heuristics were developed:
- Robust: The model uses its own previous predictions throughout training.
- Robust-CF (Coin Flip): The model randomly selects between its own prediction and the ground truth for each turn.
- Robust-P (Progressive): The model starts with ground truth and gradually transitions to using its own predictions as training progresses.
Additionally, a baseline called BERT-AH (Appended History) was tested, which avoids answer history entirely by only appending previous questions to the input.
Key Results and Insights
The experimental results highlight a significant performance gap between standard evaluation and realistic deployment scenarios.
The Standalone Performance Gap
When evaluated without ground-truth history, standard models like BERT-HAE saw their F1 scores drop from 63.4 to 53.5. This 10-point drop indicates that leaderboard rankings may not accurately reflect real-world utility.
The Filtering Trap
In an "Adversarial" experiment, the authors marked the correct answer as part of the history. For models like BERT-HAE, performance plummeted to 41.7 F1. This confirms that the models were using history markers primarily as a "negative filter" (i.e., "don't pick this text again") rather than for contextual understanding.
Success of Robust Training
The BERT-PHAE Robust-P model achieved an F1 of 58.1 in the standalone scenario, significantly narrowing the gap compared to the standard version (54.2). Furthermore, robust training mitigated the filtering bias, as evidenced by the Adversarial F1 score rising to 51.6.
Stability of Question History
The BERT-AH model, which uses only question history, achieved a respectable 58.3 F1. This suggests that question history is a more stable signal than answer history, as the system always knows exactly what the user asked, regardless of its own previous accuracy.
Conclusion
This research exposes critical flaws in current CQA evaluation. By removing Teacher Forcing and implementing Robust Training, we can develop models that are better prepared for the messiness of real-world interaction. Future work should focus on dynamic datasets that account for conversation branching and exploring backpropagation through time to further refine these autonomous systems.
Summary updated