Rethinking Cooperative Rationalization: Introspective Extraction and Complement Control

Mo Yu; Shiyu Chang; Yang Zhang; Tommi S. Jaakkola

Source-linked AI summary

Rethinking Cooperative Rationalization: Introspective Extraction and Complement Control

Mo Yu, Shiyu Chang, Yang Zhang, Tommi S. Jaakkola

arXiv:1910.13294v2Oct 29, 2019cs.CLcs.LG

TL;DR

Selective rationalization can fail when generators lack direct outcome information and when useful information remains outside selected text. This paper adds an introspective generator and adversarial complement control, with experiments showing improved accuracy and more comprehensive rationales.

Problem
Cooperative rationale selection may yield poor performance and incomplete rationales because the generator lacks direct outcome access and unselected information is not explicitly controlled.
Method
The paper combines an introspective generator conditioned on the full-text predicted label with a three-player game whose complement predictor controls information in unselected words.
Results
The combined framework maintains high predictive accuracy and produces non-degenerate, comprehensive rationales across the reported experiments.
Takeaways & Limitations
Explicitly controlling rationale complements alongside introspective selection improves both predictive accuracy and rationale quality in cooperative rationalization.
Takeaways & Limitations
The introspective generator can aggravate degeneration because it may encode its predicted label into the rationale, making the three-player game critical for the introspective model.

Abstract

from arXiv · show

Selective rationalization has become a common mechanism to ensure that predictive models reveal how they use any available features. The selection may be soft or hard, and identifies a subset of input features relevant for prediction. The setup can be viewed as a co-operate game between the selector (aka rationale generator) and the predictor making use of only the selected features. The co-operative setting may, however, be compromised for two reasons. First, the generator typically has no direct access to the outcome it aims to justify, resulting in poor performance. Second, there's typically no control exerted on the information left outside the selection. We revise the overall co-operative framework to address these challenges. We introduce an introspective model which explicitly predicts and incorporates the outcome into the selection process. Moreover, we explicitly control the rationale complement via an adversary so as not to leave any useful information out of the selection. We show that the two complementary mechanisms maintain both high predictive accuracy and lead to comprehensive rationales.

1 Introduction

Selective rationalization frames rationale generation and prediction as a cooperative game, but existing approaches can suffer from poor performance and uncontrolled information outside the rationale. The paper addresses these problems with introspective generation and adversarial complement control, improving accuracy and rationale comprehensiveness.

Selective rationalization exposes only the input portion considered relevant for prediction through a separately trained rationale generator.The generator and predictor cooperate to achieve high predictive accuracy under concise rationale constraints.
Existing cooperative selection can perform poorly because the generator typically lacks direct access to the outcome it must justify.The generator may fail to disentangle label information and may learn communication codes that overfit training data.
The introspective generator receives the label predicted from the full text as an additional input, helping it generate rationales for the intended outcome.This mechanism is designed to maintain downstream predictive performance.
Existing methods also leave the information outside the rationale uncontrolled, allowing rationales to omit context or encode labels through trivial word patterns.Examples include punctuation-based communication schemes that distinguish positive and negative predictions.
The three-player framework adds a complement predictor that uses unselected words, while the generator tries to fool it and preserve predictive accuracy.Controlling the rationale complement discourages useful target information from remaining outside the selected rationale.
Combining introspection with the three-player game yields high predictive accuracy and non-degenerate, more comprehensive rationales.The paper reports improvements in both predictive accuracy and rationale quality for cooperative frameworks.

2 Problem Formulation

The paper formalizes rationales as masked subsets of text and characterizes them by sufficiency, comprehensiveness, and compactness. It identifies degeneration as a failure of cooperative selection in which trivial signals encode the label while relevant information remains outside the rationale.

2.1 Problem Formulation of Rationalization: The rationalization task selects words from an input sequence X to produce a rationale r(X) that accounts for label Y.The rationale is represented as a hard-masked version of the input, with a binary mask determining selected positions.
2.1 Problem Formulation of Rationalization: The complement of a rationale consists of the input words excluded by its binary selection mask.The paper explicitly defines this complement to assess information left outside the rationale.
2.2 Rationale Conditions: Ideal rationales should be sufficient for predicting Y, comprehensive against predictive information in the complement, and compact in sparsity and continuity.These conditions constrain both the selected text and the information excluded from it.
2.2 Rationale Conditions: Sufficiency requires the rationale to retain the predictive power of the full input, while compactness prevents selecting unnecessary words or the entire input.Compactness constrains both the number and continuity of selected segments.
2.3 Comprehensiveness and Degeneration: Comprehensiveness regulates information outside the rationale so that relevant and useful label information remains in the selected text.It is also the condition used to prevent the degeneration failure case.
2.3 Comprehensiveness and Degeneration: Degeneration occurs when a rationale encodes the label through trivial information such as punctuation or position instead of explanatory words.A rationale can satisfy sufficiency and compactness while still selecting different positions for different labels.
2.3 Comprehensiveness and Degeneration: Previous cooperative frameworks can degenerate when a capable generator guesses Y and exploits the absence of control over unselected words.Predictive power in the rationale complement can reveal whether relevant predictors were left unselected.

3 The Proposed Three-Player Models

The proposed framework adds a complement predictor to control information outside the rationale and introduces an introspective generator that uses a predicted label during selection. The resulting three-player game is theoretically tied to rationale properties and is designed to avoid degeneration.

3.1 The Basic Three-Player Model: The basic model uses a rationale generator, a predictor operating on R, and a complement predictor operating on Rc.The generator cooperates with the predictor while playing an adversarial game against the complement predictor.
3.1 The Basic Three-Player Model: The generator extracts R and Rc through a mask determined by a weighted combination of four losses, including the complement-control loss Lg.The model also imposes sparsity and continuity constraints and encourages a gap between predictor and complement-predictor losses.
3.1 The Basic Three-Player Model: The complement predictor regularizes selection by predicting Y from words left outside the rationale, while the generator reduces that predictive ability.This explicitly controls information in the rationale complement rather than evaluating only the selected text.
3.1 The Basic Three-Player Model: Theorem 1 states that any rationalization scheme satisfying the three rationale conditions is a global optimizer of the generator objective.The proof associates each objective term with one of the required rationale properties.
3.2 The Introspection Generator: The introspective generator first predicts a label from the full input and then uses that predicted label with the input to generate the rationale mask.This design preserves the three-player formulations and their theoretical guarantee while adding label information to selection.
3.2 The Introspection Generator: Accurate introspection can worsen degeneration in a cooperative game because the generator may encode only the predicted label into R.The complement-control game is therefore critical for the introspective model.

4 Experimental Settings

The experiments evaluate rationale selection across beer-review sentiment, relation classification, and text matching tasks. The setup varies task granularity and uses bidirectional LSTM-based generators and predictors.

4.1 Datasets: The evaluation includes multi-aspect and single-aspect beer-review sentiment tasks, SemEval 2010 Task 8 relation classification, and the AskUbuntu text-matching benchmark.The single-aspect beer task and relation task require fine-grained rationales.
4.1 Datasets: The multi-aspect beer task predicts appearance sentiment from multi-aspect reviews using a threshold of 0.6 to create balanced binary classifications.The task limits the beer review labels to the appearance aspect.
4.1 Datasets: The single-aspect beer task extracts appearance-specific sentences and predicts appearance sentiment from those sentences.The authors describe this setting as more challenging because it requires finer-grained rationale selection.
4.1 Datasets: SemEval relation classification predicts the directed relation type between two target entities, including None when no relation exists.The task is also treated as a fine-grained rationalization setting.
4.2 Implementation Details: The generators and predictors use bidirectional LSTMs with hidden dimension 400, while relation classification additionally uses relative-position embeddings.The introspection classifier uses the same bidirectional LSTM architecture.

5 Experiments

Experiments evaluate rationale quality through predictive accuracy, agreement with human highlights, and information remaining in unselected text. Across beer-review and relation-classification settings, the three-player and introspective mechanisms improve rationale comprehensiveness while preserving accuracy.

Multi-Aspect Beer Review: 82.05 versus 87.59 accuracy shows that the original cooperative model loses predictive performance when using only 10% of words.Adding the third player slightly improves accuracy, while introspection reaches 86.16 versus 82.05 with only a 0.47% precision drop.
Multi-Aspect Beer Review: More than 5% higher recall results from adding the mini-max module at the 20% highlighting level.The result supports controlling unselected words to make rationales more comprehensive.
Multi-Aspect Beer Review: The three-player framework consistently improves extracted-rationale quality for both the original and introspective models.Without complement control, introspection experiences serious degeneration; with it, predictive accuracy and rationale quality remain high.
Single-Aspect Beer Review: The three-player game improves evaluation performance and lowers complement accuracy under comparable rationale lengths.For the single-aspect beer task, this indicates less informative content remains outside the selected rationale; introspection alone suffers degeneration.
Single-Aspect Beer Review: Human evaluators are more often confused and perform worse at sentiment prediction when shown only unselected text from the proposed model.These outcomes are reported as evidence that the model leaves less informative text unattended.
Relation Classification: The proposed methods show consistent gains on the relation-classification task: introspection helps accuracy, while the three-player game regularizes rationale complements.Illustrative examples also show improved entity highlighting, though dataset-correlated words may be retained.

6 Related Work

The related-work discussion situates selective rationalization among interpretability methods, game-theoretic approaches, and concurrent three-player rationale-selection work. It distinguishes this paper's motivation as addressing degeneration in cooperative rationalization.

Model Interpretability: Selective rationalization is related to feature-importance methods, local interpretable models, and document summarization viewed as a special rationale.The discussion also notes a concurrent differentiable approach to optimizing cooperative rationalization.
Game-Theoretical Methods: The three-player framework shares the minimax-game perspective used in self-play, generative models, and multi-agent reinforcement learning.The cited related methods apply game-theoretic formulations to other machine-learning problems.
Version 2 Update: An independent contemporaneous effort also controls rationale complements, but this paper begins from the distinct motivation of cooperative-rationalization degeneration.The authors state that the two works use different motivations despite their similar three-player structure.

7 Conclusion

The paper presents a framework combining introspection with three-player complement control to improve selective rationalization. Automatic and subjective evaluations support improved predictive accuracy and rationale comprehensiveness.

Conclusion: The framework regularizes unselected words with a three-player game and augments the generator with introspection to maintain downstream performance.Experiments using automatic evaluation and subjective studies confirm advantages for the proposed framework.

A Proof of Theorem 1

The proof establishes a correspondence between rationale properties and the loss terms, showing that schemes satisfying the relevant equations globally minimize those losses.

The proof uses lemmas linking rationale properties in Eqs. (4) and (5) to the optimization losses Lp and Lg.
A scheme z(X) satisfying Eq. (4) is the global minimizer of Lp.
A scheme z(X) satisfying Eq. (5) is the global minimizer of Lg.
A scheme satisfying Eq. (6) makes Ls and Lc equal to zero and is the global minimizer of both losses.The proof states that Ls and Lc are zero if and only if Eq. (6) holds.
Combining Lemmas 1.2–1.4 completes the proof of Theorem 1.

B Experimental Setup of Examples in Table 1 and Degeneration Cases of (Lei et al., 2016)

The experiment tests whether cooperative rationalization degenerates into a communication code unrelated to meaningful evidence, using concise rationales in a single-beer review task.

The experiment tests degeneration in the original Lei et al. (2016) model using a single-beer review task.
Lei et al. (2016) often selected color words such as “dark-brown” and “yellow” as rationales, reflecting a learned label communication code.
76.4% versus 82.3% performance showed that the degenerated generator failed to generalize well.
The proposed three-player method achieved higher accuracy and more meaningful rationales by comparison.

C Data Construction of the Single-Aspect Beer Reviews

The single-aspect beer-review dataset was extracted from multi-aspect reviews using frequent aspect-marker patterns, producing a balanced classification task with performance close to the original task.

Data construction: The dataset construction extracts appearance-related sentences from multi-aspect beer reviews to form a single-aspect task.
Data construction: Frequent patterns of the form “X:” were retained as anchors when their frequency exceeded 400.
Data construction: Sentences between consecutive anchor patterns were assigned to the aspect named by the first anchor.
Data construction: The constructed data formed balanced single-review binary classification tasks for the experiments.
Validation: 87.1% development accuracy on the single-aspect task was close to 87.6% on the corresponding multi-aspect task.The comparison supports the claim that the construction method introduced insignificant noise.

E Experiment Designs for Human Study

The human study evaluates whether masked texts become unpredictable after rationale removal, combining forced label guesses with subsequent UNK judgments.

Study design: The study masks rationale words in original texts and mixes outputs from different methods before evaluation.
Study design: Evaluators first predict the sentiment label of each masked text, then report whether the remaining spans provide sufficient clues.
Interpretation: Around 50% label accuracy indicates that a rationalizing method included all informative pieces in the rationale.
Study design: The forced initial label prediction is designed to reduce bias toward assigning UNK labels immediately.
Agreement: 76% interannotator agreement on UNK labels was observed on a subset of 50 examples.

F Additional Experiments on AskUbuntu

On AskUbuntu, question retrieval is evaluated as pairwise classification using selected spans, with a three-step training pipeline and complement-aware results.

Method: AskUbuntu question retrieval is formulated as pairwise classification of query–candidate question pairs, using a siamese generator to extract rationales.The predictor and complement predictor operate on the pairwise selected spans.
Implementation Details: The training strategy pre-trains a full-text classifier, pre-trains rationale generators with the classifier fixed, then fine-tunes all modules end-to-end.The classifier is used for both the predictor and complement predictor in the three-player game.
Results: Table 9 reports testing MAP and complement MAP, with high MAP and low MAPc representing the desired rationalization outcome.MAPc is the MAP score of the complement predictor.
Results: Adding the minimax game improves prediction MAP while lowering complement MAP for both the original and introspection models.The original model otherwise fails to maintain performance relative to full-text training.
Results: AskUbuntu has relatively large complement MAPs because many informative words make it difficult for generators to include all useful information.The reported results also depend substantially on the three-step training strategy and possible premature generator training from early stopping.
Implementation Details: Training and inference use different rationale decisions: sampling according to π(·) during training but selecting the highest-probability action during inference.This mismatch can leave development MAP unchanged early, risking predictor overfitting before early stopping detects improvement.