Reframing Human-AI Collaboration for Generating Free-Text Explanations
By Sarah Wiegreffe, Jack Hessel, Swabha Swayamdipta, Mark O. Riedl, Yejin Choi. 2021
176 citations recorded by the source corpus.
Large language models are increasingly capable of generating fluent-appearing text with relatively little task-specific supervision. But can these models accurately explain classification decisions? We consider the task of generating free-text explanations using human-written examples in a few-shot manner. We find that (1) authoring higher quality prompts results in higher quality generations; and (2) surprisingly, in a head-to-head comparison, crowdworkers often prefer explanations generated by GPT-3 to crowdsourced explanations in existing datasets. Our human studies also show, however, that while models often produce factual, grammatical, and sufficient explanations, they have room to improve along axes such as providing novel information and supporting the label. We create a pipeline that combines GPT-3 with a supervised filter that incorporates binary acceptability judgments from humans in the loop. Despite the intrinsic subjectivity of acceptability judgments, we demonstrate that acceptability is partially correlated with various fine-grained attributes of explanations. Our approach is able to consistently filter GPT-3-generated explanations deemed acceptable by humans.
Subfields: Artificial Intelligence, Information Systems and Management, Natural Language Processing, Alignment & RLHF
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Artificial Intelligence
- Mohit Bansal
- Yejin Choi
- Percy Liang
- Luke S. Zettlemoyer
- Jianfeng Gao
- Noah A. Smith
- Quoc V. Le
- Graham Neubig
In Information Systems and Management
- Omar Shaikh
- Xiang Ren
- Mohit Bansal
- Yejin Choi
- James L. McClelland
- Greg Durrett
- Quoc V. Le
- Ana Marasovic
In Natural Language Processing
- Mohit Bansal
- Yejin Choi
- Greg Durrett
- Luke S. Zettlemoyer
- Kyunghyun Cho
- Noah A. Smith
- Hannaneh Hajishirzi
- Jianfeng Gao
In Alignment & RLHF
Citations in the corpus
- Chain of Thought Prompting Elicits Reasoning in Large Language Models (2022, background): “, 2021), and produces explanations either simultaneously to or after the final prediction (Narang et al., 2020; Majumder et al., 2021; Wiegreffe et al., 2021).”
- Repairing the Cracked Foundation: A Survey of Obstacles in Evaluation Practices for Generated Text (2022, background): “As a result, reference-based evaluation in NLG settings works even worse in these settings where model output is often rated as preferrable to the (noisy) references (Goyal et al., 2022; Wiegreffe et al., 2022).”
- Can language models learn from explanations in context? (2022, background): “It could also potentially help with user interpretability, though with the caveat that explanations might not be entirely reliable, at least without further refinement (Marasović et al., 2021; Wiegreffe et al., 2021).”
- Few-Shot Self-Rationalization with Natural Language Prompts (2021, methodology): “Wiegreffe et al. (2022) analyze explanations obtained by prompting GPT-3 multiple times to get multiple explanation candidates, and then filter these candidates using a model trained to predict acceptability of explanations.”
- Rationale-Augmented Ensembles in Language Models (2022, methodology): “• Input-rationale ensemble: Leveraging the ability of large language models to generate high-quality explanations (Wiegreffe et al., 2022), model-generated rationales can replace human-written rationales in the input prompts (e.g., via the process described in Section 2.1), which can then be used…”
- Large Language Model as Attributed Training Data Generator: A Tale of Diversity and Bias (2023, methodology): “Then, we adopt the human-ai collaboration scheme [26, 54, 61] to interactively select the attribute dimensions of the highest quality that best suit the dataset.”
- REV: Information-Theoretic Evaluation of Free-Text Rationales (2022, methodology): “We introduce R EV , an information-theoretic measure to evaluate the amount of new, label-relevant information in free-text rationales, beyond the information contained in the input.”
- Reasoning with Language Model Prompting: A Survey (2022, background): “, 2022b), Human-AI (Wiegreffe et al., 2022), MT-CoT (Li et al.”
- Towards Reasoning in Large Language Models: A Survey (2022, methodology): “…method being used, e.g., deductive reasoning (Cobbe et al., 2021; Creswell et al., 2022; Han et al., 2022b, inter alia ), inductive reasoning (Yang et al., 2022; Misra et al., 2022, inter alia ) or abductive reasoning (Wiegreffe et al., 2022; Lampinen et al., 2022; Jung et al., 2022, inter alia ).”
- Are Hard Examples also Harder to Explain? A Study with Human and Model-Generated Explanations (2022, methodology): “erate free-text explanations using in-context learning with GPT-3 (Marasović et al., 2022; Wiegreffe et al., 2022).”
- ToolQA: A Dataset for LLM Question Answering with External Tools (2023, methodology): “To address these challenges, we propose a human-guided LLM generation approach that uses question templates to bridge human guidance and automatic LLM generation [67, 79].”
- The Unreliability of Explanations in Few-shot Prompting for Textual Reasoning (2022, background): “The utility of explanations for few-shot in-context learning has also been discussed concurrently (Nye et al., 2021; Wei et al., 2022; Marasović et al., 2022; Chowdhery et al., 2022; Lampinen et al., 2022; Wiegreffe et al., 2022), especially in symbolic reasoning tasks.”
- Can Large Language Models Transform Computational Social Science? (2023, methodology): “Dataset construction through human curation of LLM generations has already emerged in recent NLP works on decision explanation (Wiegreffe et al., 2022), model error identification (Ribeiro and Lundberg, 2022), and even to build the figurative language benchmark used in this work (Chakrabarty et…”
- VisIT-Bench: A Benchmark for Vision-Language Instruction Following Inspired by Real-World Use (2023, background): “The rationale for using GPT-4 verification instead is derived from prior results that show promising humanmachine collaboration of this form [37].”