Rethinking Cooperative Rationalization: Introspective Extraction and Complement Control

By Mo Yu, Shiyu Chang, Yang Zhang, Tommi S. Jaakkola. 2019

159 citations recorded by the source corpus.

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.

Subfields: Natural Language Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Natural Language Processing

Citations in the corpus

SELFEXPLAIN: A Self-Explaining Architecture for Neural Text Classifiers (2021, background): “Sufﬁciency aims to evaluate whether model explanations alone are highly indicative of the predicted label (Jacovi et al., 2018; Yu et al., 2019).”
ERASER: A Benchmark to Evaluate Rationalized NLP Models (2019, methodology): “In particular, following Yu et al. (2019) we deﬁne metrics intended to measure the comprehensiveness (were all features needed to make a prediction selected?) and sufﬁciency (do the extracted rationales contain enough signal to come to a disposition?) of rationales, respectively.”
Learning to Faithfully Rationalize by Construction (2020, methodology): “Prior efforts (Lei et al., 2016; Yu et al., 2019; Bastings et al., 2019) have proposed methods that produce faithful explanations via a two-model setup, deﬁning a generator network that imposes hard attention over inputs and then passes these to a second model for prediction.”
Teach Me to Explain: A Review of Datasets for Explainable Natural Language Processing (2021, background): “[141] introduce a third criterion, (iii) comprehensiveness, that all the evidence that supports the prediction is selected, not just a sufficient set.”
A Game Theoretic Approach to Class-wise Selective Rationalization (2019, background): “It has been pointed out in [32] that the existing generator-predictor framework in [12] and [19] can suffer from the problem of degeneration .”
An Information Bottleneck Approach for Controlling Conciseness in Rationale Extraction (2020, background): “Building on Lei et al. (2016), Chang et al. (2019) and Yu et al. (2019) consider beneﬁts like class-wise explanation extraction while Chang et al. (2020) explore invariance to domain shift.”
multiPRover: Generating Multiple Proofs for Improved Interpretability in Rule Reasoning (2021, background): “Explanations in NLP take three major forms – (1) extractive rationales or highlights (Zaidan et al., 2007; Lei et al., 2016; Yu et al., 2019; DeYoung et al., 2020) where a subset of the input”
Teach Me to Explain: A Review of Datasets for Explainable NLP (2021, background): “Yu et al. (2019) introduce a third property, (iii) comprehensiveness , that all the evidence that supports the prediction is selected, not just a sufﬁcient set—i.e., the tokens that are not selected do not predict the label.”
Rational LAMOL: A Rationale-based Lifelong Learning Framework (2021, methodology): “Existing rationalization techniques commonly use the maximum mutual information (MMI) criterion to select rationales, which is prone to choosing spurious correlation between input features and outputs as rationales (Lei et al., 2016; Yu et al., 2019).”
Paragraph-level Rationale Extraction through Regularization: A case study on European Court of Human Rights Cases (2021, methodology): “• We study the effect of previously proposed rationale constraints, i.e., sparsity , continuity (Lei et al., 2016), and comprehensiveness (Yu et al., 2019), formulated as regularizers.”
Learning to Give Checkable Answers with Prover-Verifier Games (2021, unknown): “A number of game theoretic selective rationalization methods have been proposed [20, 37, 4].”
Aligning Faithful Interpretations with their Social Attribution (2020, background): “Movies (Zaidan and Eisner, 2008) Movie reviews labeled for binary polarity.”
ExplaGraphs: An Explanation Graph Generation Task for Structured Commonsense Reasoning (2021, background): “6), and (3) unlike semi-structured templates (Ye et al., 2020; Mostafazadeh et al., 2020) or extractive rationales (Zaidan et al., 2007; Lei et al., 2016; Yu et al., 2019; DeYoung et al., 2020), they allow for more ﬂexi-bility and expressiveness.”
Evaluating and Characterizing Human Rationales (2020, methodology): “Popular automatic metrics for evaluating machine-generated rationales are sufﬁciency and comprehen-siveness , articulated by Yu et al. (2019) and employed in the ERASER benchmark (DeYoung et al., 2019).”