Abstract
Interactive sequence labeling (ISL) presents a path toward efficient, privacy-preserving NLP by allowing models to learn directly from user interactions. However, ISL systems are often plagued by "context bias"—where models rely on spurious correlations—and "feedback bias"—where models fail to explore rare entity types due to limited user feedback. We introduce CITISL, a framework that leverages causal inference to mitigate these biases. By implementing a deconfounding layer and entropy-based modulation within a Structural Causal Model (SCM), CITISL significantly improves performance across POS tagging and NER tasks while reducing human labeling effort.
Introduction
Sequence labeling tasks, such as Named Entity Recognition (NER), typically require massive, manually annotated datasets. Crowdsourcing these labels is expensive and often incompatible with private data. Interactive Sequence Labeling offers a solution: the model predicts tags, and the user provides real-time feedback.
Despite its potential, ISL faces two critical challenges:
- Context Bias: Models often learn spurious correlations from initial training data. For example, if "Service" frequently follows "Government," the model may incorrectly label "Service" as a government entity even in a software context (e.g., "Service Pack 1").
- Feedback Bias: Users only interact with the top $K$ predictions shown by the model. This creates a feedback loop where the model ignores rare classes, never receiving the supervision needed to learn them.
The CITISL Framework
To address these issues, the authors propose CITISL (Context-aware Information-Theoretic Interactive Sequence Labeling). The framework treats bias as a causal problem, modeled through a Structural Causal Model (SCM).
Deconfounding Context Bias
CITISL identifies frequent context words as "confounders" that create non-causal paths to predictions. A deconfounding layer applies backdoor adjustment to mathematically "block" these spurious paths. This ensures the model focuses on the intrinsic features of the token rather than just its neighbors.
Alleviating Feedback Bias
To prevent the model from getting stuck in a "certainty trap," CITISL uses entropy-based modulation. By calculating the mutual information between context and tokens, the system identifies when a prediction is overly reliant on context. It then injects controlled noise (acting as an instrumental variable) to encourage the display of a more diverse set of candidate labels, facilitating exploration of the "long tail" of entity types.
Handling Negative Feedback
When a user selects "none of the above," the model receives negative feedback but no ground truth. CITISL employs Doubly Robust Estimation, using an imputation model trained on a small "gold" dataset to provide a proxy label, ensuring the model continues to learn from its mistakes.
Experimental Evaluation
The researchers evaluated CITISL on four benchmarks: CoNLL-2003-POS, UD-ENG, Few-NERD, and OntoNote. Starting with as little as 3% initial data, the system learned from simulated user clicks.
Key Results
- Performance: On CoNLL-2003-POS with 10% initial data, CITISL achieved an F1 score of 85.90, vastly outperforming a standard BERT baseline (66.71).
- Efficiency: CITISL performed comparably to models with access to full ground truth labels, even when only showing the user 3 candidates.
- Human Impact: In real-world trials, the framework reduced the average time to label a sentence from 23.96 seconds to 16.99 seconds, demonstrating significant gains in human productivity.
Conclusion and Future Work
CITISL demonstrates that causal inference can effectively break the bias loops inherent in interactive learning. While the current token-level approach is highly effective, future work could integrate sequence-level dependencies (e.g., via CRFs) and optimize computational overhead to further improve real-time responsiveness.
Summary updated