Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
41 tokens/sec
GPT-4o
59 tokens/sec
Gemini 2.5 Pro Pro
41 tokens/sec
o3 Pro
7 tokens/sec
GPT-4.1 Pro
50 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Prompting Contrastive Explanations for Commonsense Reasoning Tasks (2106.06823v1)

Published 12 Jun 2021 in cs.CL and cs.AI

Abstract: Many commonsense reasoning NLP tasks involve choosing between one or more possible answers to a question or prompt based on knowledge that is often implicit. Large pretrained LLMs (PLMs) can achieve near-human performance on such tasks, while providing little human-interpretable evidence of the underlying reasoning they use. In this work, we show how to use these same models to generate such evidence: inspired by the contrastive nature of human explanations, we use PLMs to complete explanation prompts which contrast alternatives according to the key attribute(s) required to justify the correct answer (for example, peanuts are usually salty while raisins are sweet). Conditioning model decisions on these explanations improves performance on two commonsense reasoning benchmarks, as compared to previous non-contrastive alternatives. These explanations are also judged by humans to be more relevant for solving the task, and facilitate a novel method to evaluate explanation faithfulfness.

Overview of "Prompting Contrastive Explanations for Commonsense Reasoning Tasks"

The paper "Prompting Contrastive Explanations for Commonsense Reasoning Tasks" addresses the limitations of pretrained LLMs (PLMs) in providing interpretable evidence for commonsense reasoning tasks. Although PLMs exhibit remarkable performance on commonsense reasoning without relying on external knowledge sources, they typically do not offer insights into their decision-making processes. This work introduces an unsupervised approach that enhances the interpretability of PLM outputs by generating contrastive explanations, which articulate the distinguishing attributes necessary to justify the correct choice among alternatives.

Key Contributions

This paper proposes a method utilizing contrastive prompts to elicit explanations from PLMs. The researchers hypothesize that commonsense reasoning tasks often require contrasting plausible alternatives based on specific attributes. For instance, distinguishing between peanuts and raisins involves contrasting their tastes—salty and sweet, respectively. The strategy involves using the PLMs, like T5 and BART, to generate explanations that contrast these alternatives based on contextually relevant attributes.

The paper demonstrates improved performance in commonsense reasoning benchmarks—specifically, the Winograd Schema Challenge and multiple-choice question answering related to physical commonsense—by conditioning PLMs on these contrastive explanations. Furthermore, human evaluations reveal that explanations generated through this method are perceived as more relevant and useful compared to non-contrastive or clarification question-based alternatives.

Quantitative Results

The approach marked significant quantitative enhancements, outperforming previous explainability methods by substantial margins. In zero-shot settings, the technique achieved superior performance compared to reported results for publicly available PLMs, with improvements up to 11% over previous attempts. Importantly, the method demonstrated its effectiveness in generating explanations deemed useful by human judgment, confirming the relevance of the insightful attribute-based contrasts.

Implications and Future Directions

The paper's implications span both practical and theoretical realms within the field of AI and NLP. Practically, the approach provides a framework for enhancing the transparency of model predictions, which is paramount in deploying AI systems in sensitive domains. Theoretically, it provides insights into how human-like explanations might be integrated into machine reasoning processes, aligning closer with cognitive models of explanation that emphasize contrastive reasoning.

Looking forward, the approach invites exploration into the broader applicability of contrastive prompts across diverse reasoning tasks beyond commonsense benchmarks. Future research may investigate the integration of this method with more extensive PLMs, such as GPT-3, potentially overcoming current limitations within zero-shot tasks. Additionally, finding optimal ways to marry explanation generation with other forms of knowledge representation could lead to even more robust and accountable AI systems.

In summary, this paper provides a compelling methodology for enriching the interpretability of PLM outputs through contrastive explanations, pushing the boundaries of what is achievable in commonsense reasoning tasks using NLP models. The insights gleaned from this research pave the way for more human-aligned AI decision-making models.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (5)
  1. Bhargavi Paranjape (18 papers)
  2. Julian Michael (28 papers)
  3. Marjan Ghazvininejad (33 papers)
  4. Luke Zettlemoyer (225 papers)
  5. Hannaneh Hajishirzi (176 papers)
Citations (61)
Youtube Logo Streamline Icon: https://streamlinehq.com