PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning to Rationalize for Nonmonotonic Reasoning with Distant Supervision

Faeze Brahman, Vered Shwartz, Rachel Rudinger, Yejin Choi

arXiv:2012.08012v1cs.CL

TL;DR

Neural rationale generation is often trained with costly, dataset-specific human explanations, motivating a more transferable distant-supervision approach. This paper builds rationale generators for defeasible inference using automatically sourced rationales and evaluates post-hoc and joint setups. Post-hoc models can produce relevant and correct rationales, but they frequently rely on trivial target-value mappings, while joint prediction remains substantially harder.

  • Problem

    Human-written rationale supervision is costly and may not generalize across datasets and domains, motivating rationale learning without additional annotation.

    Source: §1 · p. 1, §2 · p. 2

  • Method

    The paper constructs e-δ-NLI from rationales gathered from language models, knowledge bases, and related-task supervision, then trains post-hoc and joint generative models.

    Source: §1 · p. 1

  • Results

    Post-hoc rationales were mostly relevant and factually correct, with 40% considered explanatory on average, but models often learned trivial target-value mappings.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Distant supervision can support post-hoc rationale generation for defeasible inference, but realistic joint prediction and rationale generation remains challenging.

    Source: §7 · p. 7

    Post-hoc success may reflect access to update values or types, enabling trivial mappings rather than faithful explanations.

    Source: §7 · p. 7, §6.2 · p. 7

Abstract

from arXiv · show

The black-box nature of neural models has motivated a line of research that aims to generate natural language rationales to explain why a model made certain predictions. Such rationale generation models, to date, have been trained on dataset-specific crowdsourced rationales, but this approach is costly and is not generalizable to new tasks and domains. In this paper, we investigate the extent to which neural models can reason about natural language rationales that explain model predictions, relying only on distant supervision with no additional annotation cost for human-written rationales. We investigate multiple ways to automatically generate rationales using pre-trained language models, neural knowledge models, and distant supervision from related tasks, and train generative models capable of composing explanatory rationales for unseen instances. We demonstrate our approach on the defeasible inference task, a nonmonotonic reasoning task in which an inference may be strengthened or weakened when new information (an update) is introduced. Our model shows promises at generating post-hoc rationales explaining why an inference is more or less likely given the additional information, however, it mostly generates trivial rationales reflecting the fundamental limitations of neural language models. Conversely, the more realistic setup of jointly predicting the update or its type and generating rationale is more challenging, suggesting an important future direction.

1 Introduction

The paper studies distant-supervised rationale generation for defeasible inference, aiming to explain how updates strengthen or weaken hypotheses without costly human rationale annotation. Post-hoc rationales show promise but often exploit trivial target-value patterns, while joint prediction is more difficult.

Task

  • The authors introduce defeasible-inference rationale generation to explain why an update makes a hypothesis more or less likely.The task covers both strengthener and weakener updates.Source: §1 · p. 1

Method

  • The approach creates e-δ-NLI by combining rationales from pre-trained language models, knowledge bases, and related-task supervision, then trains generative models.The models include post-hoc generation conditioned on target values and joint prediction with rationale generation.Source: §1 · pp. 1–2

Results

  • Post-hoc models produced rationales that were mostly relevant and factually correct, with 40% considered explanatory on average.Human and automatic evaluations were used.Source: §1 · p. 2

Analysis

  • Post-hoc models often learned trivial mappings from target values to label-specific rationale patterns, such as stating that the update implies the hypothesis.This limits how explanatory the generated rationales are despite their surface relevance and correctness.Source: §1 · p. 2

Limitations

  • The more realistic joint setup removes access to the target value, but its lower performance makes it a challenging direction for future research.The paper characterizes this setup as more realistic because the model must jointly predict the update or its type and generate the rationale.Source: §1 · p. 2

2 Background

The background frames the work around explainability without costly human rationales and situates it in NLI, defeasible inference, and joint or post-hoc rationale generation. It also describes distant supervision from existing rationale resources and related-task models.

Natural Language Inference

  • NLI classifies whether a hypothesis is entailed, contradicted, or neutral given a premise.The paper treats entailment as a judgment that the hypothesis is most likely true for a human reader.Source: §2 · p. 2

Distant Supervision

  • Because human explanations are costly and unavailable for many datasets, the paper proposes distant supervision using resources such as e-SNLI without additional supervision.The approach repurposes NLI rationale generation and maps related-task rationale resources to defeasible inference.Source: §2 · p. 2

Explainable NLI

  • Natural-language explanation methods can use ante-hoc, joint, or post-hoc rationalization, but their rationales are not guaranteed to reflect the model’s true decision process.The paper focuses on joint and post-hoc approaches while acknowledging that they are not constructed to be faithful.Source: §2 · p. 2

Defeasible Inference

  • Defeasible inference allows valid conclusions to change when new information arrives, with updates classified as weakeners or strengtheners.The δ-NLI tasks either classify an update type from premise, hypothesis, and update or generate an update of a specified type.Source: §2 · p. 2

Pre-trained Language Models

  • The study also draws on pre-trained language models that can generate or score textual knowledge for commonsense reasoning.Prior work used these models to complete knowledge bases, assess answer plausibility, and generate clarification information.Source: §2 · p. 3

3 e-δ-NLI Dataset

The e-δ-NLI dataset augments defeasible inference with distantly supervised rationales generated from language models, knowledge resources, and related-task supervision, then filters candidates using an NLI proxy classifier.

3 e-δ-NLI Dataset

  • e-δ-NLI augments defeasible inference instances with rationales explaining why updates strengthen or weaken hypotheses, avoiding new human rationale annotation.Rationales are gathered from multiple sources rather than elicited from humans.Source: §3 · p. 3

3.1 Collecting Rationales

  • Salient spans are extracted from the hypothesis and update using attention weights from a δ-NLI update-type classifier.The method retains the top 20% non-continuous spans by attention score.Source: §3.1 · p. 3
  • Candidate rationales combine vanilla and ConceptNet-enhanced language-model generations, COMeT commonsense completions, and NLI-derived explanations.Vanilla models generate definitions, purposes, and span relationships; COMeT supplies preconditions and postconditions.Source: §3.1 · pp. 3–4

3.2 Filtering Rationales

  • The filtering procedure approximates defeasible inference because it maps strengtheners to entailment and weakeners to contradiction, although that equivalence is not strict.The approximation achieves 64% update-type prediction accuracy, while defeasible updates need only change likelihood rather than determine likelihood.Source: §3.2 · p. 4
  • A binary RoBERTa classifier ranks candidates by confidence in entailment for strengtheners or contradiction for weakeners, retaining the top 10% per instance.The filtered dataset contains approximately 8 rationales per instance on average.Source: §3.2 · p. 5
  • The final dataset follows an 80%/10%/10% train/test/development split and contains 731,579 training, 15,781 test, and 15,527 development instances.Multiple retained rationales are added to each original δ-NLI instance.Source: §3.2 · p. 5

4 Rationale Generation Model

The rationale-generation models fine-tune pretrained transformers on e-δ-NLI under post-hoc or joint prediction setups. Post-hoc models are easier but can exploit target-value shortcuts, whereas joint prediction is more realistic and difficult.

4 Rationale Generation Model

  • The models treat each premise, hypothesis, update type, update, and rationale combination as a separate training instance for rationale generation.The dataset supplies multiple rationales for each original example.Source: §4 · p. 5

4.1 Architecture and Implementation Details

  • The study fine-tunes GPT2-XL and BART-L transformer language models on the e-δ-NLI dataset.Each model is trained for two epochs; preliminary T5 experiments showed no improvement.Source: §4.1 · p. 5

4 Rationale Generation Model

  • Evaluation generates test-set rationales with five-beam search and combines automatic metrics with human assessment.Table 3 reports automatic and human results, separating human evaluation for strengtheners and weakeners.Source: Table 3 · p. 5

4.2 Training Objective

  • Training minimizes conditional output log-likelihood, computed over the full concatenated sequence for GPT2 and only the output sequence for encoder-decoder BART.This reflects the different language-modeling objectives used by the two architectures.Source: §4.2 · p. 5
  • Post-hoc setups condition rationale generation on the target update or update type, while joint setups predict an update or its type together with an explanation.Multi-task post-hoc training additionally predicts update type and update alongside rationale generation.Source: §4.2 · p. 5

5 Results

The evaluation combines automatic overlap metrics with human judgments of rationale quality. Post-hoc models produce generally relevant and correct rationales, but joint prediction is substantially harder and generated explanations often remain trivial.

5.1 Automatic Evaluation

  • GPT2-based models achieved better automatic BLEU and ROUGE scores overall, while the multi-task setup added gains on both metrics.The automatic evaluation used BLEU-4 and ROUGE-L against distantly supervised rationales.Source: §5.1 · p. 6

5.2 Human Evaluation

  • Rationale BART-L produced 80% relevant rationales, over 55% correct rationales, and 33%-47% explanatory rationales, with better results for strengtheners than weakeners.Most rationales across models were grammatical or understandable, at 83%-99%.Source: §5.2 · p. 6, Table 4 · p. 6
  • The multi-task setup did not improve human-rated rationale generation, and BART-based post-hoc models substantially outperformed GPT2-based models.This contrasts with the automatic evaluation, where multi-task training showed additive gains.Source: §5.2 · p. 6, §5.1 · p. 6
  • Only 0.5%-8.5% of joint-model rationales were considered explanatory, whereas post-hoc models produced more explanatory rationales despite relying on trivial target-value shortcuts.Access to the update or update type improved explanatory quality in post-hoc settings, but the joint setup was extremely challenging.Source: §5.2 · p. 6
  • Among rationales judged explanatory, almost all followed a small number of patterns that could be generated trivially from the target value.The patterns motivate focusing future research on the joint setup.Source: §5.2 · p. 6, Table 5 · p. 6

6 Analysis

The analysis finds that distantly supervised training data can be lower quality than generated rationales, while generated explanations still exhibit trivial templates and reasoning errors. Ablations show that removing filtering or non-NLI sources improves relevance but harms correctness and explanatory quality.

6.1 Quality of the Distant Supervision

  • NLI-derived rationales were judged most explanatory, whereas LM-generated training rationales were often incorrect and non-explanatory; only 7.1% of training rationales were NLI-derived.The filtering step retained 58% NLI-derived rationales, helping explain why generated rationales could surpass the training data in quality.Source: §6.1 · p. 6

6.2 Quality of Generated Rationales

  • Among 160 human-labeled explanatory rationales, almost all followed trivial patterns that could be generated from the target update type.This finding motivated focusing future work on the joint setup rather than post-hoc rationalization.Source: §6.2 · p. 7
  • Generated rationales also included rephrasing errors and explanations of the premise instead of the hypothesis.These errors indicate that grammatical, relevant, and correct outputs may still fail to connect the update to the hypothesis properly.Source: §6.2 · p. 7
  • Many weakener errors were insufficient because the rationale omitted another required reasoning hop.Other errors included incorrect implications, wrong pre- or post-conditions, partial use of the update or hypothesis, repetitive statements, and wrong templates.Source: §6.2 · p. 7

6.3 Ablation Studies

  • Both ablations increased relevance but reduced factual correctness and produced less explanatory weakener rationales.Using only NLI-derived rationales likely encouraged copying input fragments into label-specific templates, yielding relevant but not necessarily correct or explanatory outputs.Source: §6.3 · p. 7, Table 8 · p. 7

7 Conclusion

The paper presents distant-supervised rationale generation for defeasible inference, comparing post-hoc and joint prediction-and-rationalization setups. Post-hoc models often produce explanatory rationales but can exploit trivial mappings from update types, whereas the more realistic joint setup is harder and remains a future direction.

7 Conclusion

  • The approach generates defeasible-inference rationales using distant supervision rather than collecting human explanations.It experiments with post-hoc rationalization and joint prediction with rationalization.Source: §7 · p. 7
  • Post-hoc rationalization is easier and often yields human-judged explanatory rationales, but access to the update type enables trivial template mappings.The joint setup removes that access, making it more challenging and potentially more faithful.Source: §7 · p. 7
  • The authors identify jointly predicting the label and generating the rationale as a realistic direction for obtaining less trivial and more faithful explanations.The conclusion frames this as future work rather than an established result.Source: §7 · p. 7

Related papers

Loading 2012.08012v1…