Abstract: Mo Yu, Shiyu Chang, Yang Zhang, Tommi Jaakkola. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP).
2019.
Introduction
In the field of interpretable Natural Language Processing (NLP), selective rationalization has emerged as a prominent technique for explaining model decisions. The framework typically involves a two-player cooperative game: a generator selects a subset of the input text (the rationale), and a predictor uses only that subset to make a final classification. The objective is to find the most concise rationale that maintains high predictive accuracy. However, standard cooperative frameworks suffer from two significant limitations that undermine their reliability and interpretability.
The Limitations of Cooperative Games
The authors identify two primary failure modes in existing rationalization models:
- The Blind Generator: In traditional setups, the generator must select words without knowing the final prediction. This lack of foresight often results in the extraction of suboptimal rationales, as the generator is essentially guessing which features the predictor will find most useful.
- Degeneration (The "Cheating" Problem): Because there is no constraint on the information left out of the rationale, models often develop "degenerate" communication schemes. For instance, a model might learn to include a specific punctuation mark (like a period) to signal a positive sentiment and another (like a comma) for a negative one. While these rationales are compact and sufficient for the predictor, they offer no semantic value to human readers.
Methodology: Introspection and Complement Control
To address these flaws, the researchers propose a new "three-player game" framework that introduces two key architectural innovations:
The Introspective Generator
Instead of selecting words blindly, the generator first performs an internal classification on the entire input text to predict a label. This predicted label is then used as an additional input, allowing the generator to "introspect" on the likely outcome and select words that specifically justify that prediction.
The Three-Player Game
The framework introduces a third player, the Complement Predictor, to enforce comprehensiveness. The game is structured as follows:
- The Generator: Aims to select a rationale that enables the Predictor to succeed while simultaneously causing the Complement Predictor to fail.
- The Predictor: Attempts to predict the label using only the selected rationale.
- The Complement Predictor: Attempts to predict the label using only the discarded text.
By making the game adversarial, the generator is forced to move all task-relevant information into the rationale, leaving only uninformative noise for the complement predictor. This effectively eliminates the possibility of "punctuation cheats."
Experimental Results
The model was evaluated on several benchmarks, including the Multi-aspect Beer Reviews dataset and SemEval 2010 Task 8. The results demonstrated that the proposed framework significantly improves both predictive performance and rationale quality.
Quantitative Performance
On the beer review task, the full "Intros+minimax" model achieved an accuracy of 86.16% using only 10% of the text as a rationale. This represents a substantial improvement over the standard cooperative baseline (Lei et al., 2016), which achieved 82.05% accuracy under the same constraints.
Human Alignment and Information Leakage
Beyond raw accuracy, the rationales generated by the three-player model showed much higher precision and recall when compared to human-annotated ground truths. Furthermore, in human evaluations where participants were asked to guess the sentiment based only on the discarded text, they chose "Unknown" 54% of the time for the proposed model, compared to only 43.5% for the baseline. This confirms that the adversarial framework successfully stripped meaningful information from the non-rationale text.
Conclusion
By rethinking the cooperative game as an introspective, three-player adversarial process, this work provides a robust solution to the problem of degenerate rationales. The framework ensures that extracted explanations are not only sufficient for prediction but are also comprehensive and semantically meaningful to humans, marking a significant step forward for interpretable AI.
Summary updated