AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Diplomat: A Dialogue Dataset for Situated PragMATic Reasoning

Hengli Li, Song-Chun Zhu, Zilong Zheng

NeurIPS 2023 · Large language models › Language, safety and interpretability · approach: Faithfulness of LLM explanations

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
provide human-annotated rationales as input (ablation)BART-base · DiPlomat CQAaccuracy: 20.2 -> 29.6Improved
provide human-annotated rationales as input (ablation)T5-small · DiPlomat CQAaccuracy: 24.9 -> 34.1Improved
provide human-annotated rationales as input (ablation)mT5-small · DiPlomat CQAaccuracy: 19.7 -> 29.8Improved
provide human-annotated rationales as input (ablation)UnifiedQA-base · DiPlomat CQAaccuracy: 28.9 -> 38.8Improved
provide human-annotated rationales as input (ablation)UnifiedQA-large · DiPlomat CQAaccuracy: 32.8 -> 42.4Improved
provide human-annotated rationales as input (ablation)ChatGPT (pm) · DiPlomat CQAaccuracy: 40.6 -> 45.1Improved
provide human-annotated rationales as input (ablation)ChatGPT (em) · DiPlomat CQAaccuracy: 1.0 -> 1.5No clear effect
add chain-of-thought prompt 'Let's think step by step' (ablation)ChatGPT · DiPlomat zero-shot NLIaccuracy: 85.7 -> 63.8Hurt
give the full dialogue context (ChatGPT) instead of only the pragmatic turn (DeBERTa/T5)NLI models · DiPlomat zero-shot NLIaccuracy: 45.3 -> 85.7Improved
require model to identify pragmatic turn then reason (C->PR) vs reasoning given the turn (CP->R)BERT-base · DiPlomataccuracy: 91.3 -> 50.2Hurt
scale RoBERTa from base to largeRoBERTa · DiPlomat CP->Raccuracy: 92.0 -> 60.8Hurt
scale RoBERTa from base to largeRoBERTa · DiPlomat C->PRaccuracy: 50.0 -> 0.0Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy