Evaluating Explanations: How Much Do Explanations from the Teacher Aid Students?

Author:

Pruthi Danish1,Bansal Rachit2,Dhingra Bhuwan3,Soares Livio Baldini4,Collins Michael5,Lipton Zachary C.6,Neubig Graham7,Cohen William W.8

Affiliation:

1. Carnegie Mellon University, USA. ddanish@cs.cmu.edu

2. Delhi Technological University, India. racbansa@gmail.com

3. Google Research, USA. bdhingra@google.com

4. Google Research, USA. liviobs@google.com

5. Google Research, USA. mjcollins@google.com

6. Carnegie Mellon University, USA. zlipton@cs.cmu.edu

7. Carnegie Mellon University, USA. gneubig@cs.cmu.edu

8. Google Research, USA. wcohen@google.com

Abstract

Abstract While many methods purport to explain predictions by highlighting salient features, what aims these explanations serve and how they ought to be evaluated often go unstated. In this work, we introduce a framework to quantify the value of explanations via the accuracy gains that they confer on a student model trained to simulate a teacher model. Crucially, the explanations are available to the student during training, but are not available at test time. Compared with prior proposals, our approach is less easily gamed, enabling principled, automatic, model-agnostic evaluation of attributions. Using our framework, we compare numerous attribution methods for text classification and question answering, and observe quantitative differences that are consistent (to a moderate to high degree) across different student model architectures and learning strategies.1

Publisher

MIT Press - Journals

Subject

Artificial Intelligence,Computer Science Applications,Linguistics and Language,Human-Computer Interaction,Communication

Reference33 articles.

1. Neural machine translation by jointly learning to align and translate;Bahdanau;3rd International Conference on Learning Representations, ICLR,2015

2. Deriving machine attention from human rationales;Bao,2018

3. Do explanations make VQA models more predictable to a human?;Chandrasekaran,2018

4. Reading tea leaves: How humans interpret topic models;Chang,2009

5. Invariant rationalization;Chang,2020

Cited by 6 articles. 订阅此论文施引文献 订阅此论文施引文献,注册后可以免费订阅5篇论文的施引文献,订阅后可以查看论文全部施引文献

1. How rationals boost textual entailment modeling: Insights from large language models;Computers and Electrical Engineering;2024-10

2. Explanation based Bias Decoupling Regularization for Natural Language Inference;2024 International Joint Conference on Neural Networks (IJCNN);2024-06-30

3. Human-annotated rationales and explainable text classification: a survey;Frontiers in Artificial Intelligence;2024-05-24

4. Towards Faithful Model Explanation in NLP: A Survey;Computational Linguistics;2024

5. Towards a Unified Multidimensional Explainability Metric: Evaluating Trustworthiness in AI Models;2023 19th International Conference on Distributed Computing in Smart Systems and the Internet of Things (DCOSS-IoT);2023-06

同舟云学术

1.学者识别学者识别

2.学术分析学术分析

3.人才评估人才评估

"同舟云学术"是以全球学者为主线,采集、加工和组织学术论文而形成的新型学术文献查询和分析系统,可以对全球学者进行文献检索和人才价值评估。用户可以通过关注某些学科领域的顶尖人物而持续追踪该领域的学科进展和研究前沿。经过近期的数据扩容,当前同舟云学术共收录了国内外主流学术期刊6万余种,收集的期刊论文及会议论文总量共计约1.5亿篇,并以每天添加12000余篇中外论文的速度递增。我们也可以为用户提供个性化、定制化的学者数据。欢迎来电咨询!咨询电话:010-8811{复制后删除}0370

www.globalauthorid.com

TOP

Copyright © 2019-2024 北京同舟云网络信息技术有限公司
京公网安备11010802033243号  京ICP备18003416号-3