[Paper Review] A Multitask, Multilingual, Multimodal Evaluation of ChatGPT on Reasoning, Hallucination, and Interactivity

Yejin Bang, Samuel Cahyawijaya+11 more|arXiv (Cornell University)|Feb 8, 2023

Topic Modeling352 citations

TL;DR

The paper presents a zero-shot, multitask, multilingual, and multimodal benchmark framework to evaluate ChatGPT across 23 datasets (8 tasks), revealing strengths in zero-shot multitasking and limitations in low-resource languages, reasoning reliability, hallucinations, and the benefits of interactivity.

ABSTRACT

This paper proposes a framework for quantitatively evaluating interactive LLMs such as ChatGPT using publicly available data sets. We carry out an extensive technical evaluation of ChatGPT using 23 data sets covering 8 different common NLP application tasks. We evaluate the multitask, multilingual and multi-modal aspects of ChatGPT based on these data sets and a newly designed multimodal dataset. We find that ChatGPT outperforms LLMs with zero-shot learning on most tasks and even outperforms fine-tuned models on some tasks. We find that it is better at understanding non-Latin script languages than generating them. It is able to generate multimodal content from textual prompts, via an intermediate code generation step. Moreover, we find that ChatGPT is 63.41% accurate on average in 10 different reasoning categories under logical reasoning, non-textual reasoning, and commonsense reasoning, hence making it an unreliable reasoner. It is, for example, better at deductive than inductive reasoning. ChatGPT suffers from hallucination problems like other LLMs and it generates more extrinsic hallucinations from its parametric memory as it does not have access to an external knowledge base. Finally, the interactive feature of ChatGPT enables human collaboration with the underlying LLM to improve its performance, i.e, 8% ROUGE-1 on summarization and 2% ChrF++ on machine translation, in a multi-turn "prompt engineering" fashion. We also release codebase for evaluation set extraction.

Motivation & Objective

• Assess ChatGPT's zero-shot performance across diverse NLP tasks (summarization, translation, sentiment analysis, QA, dialogue, misinformation) and multilingual settings.
• Evaluate ChatGPT's multimodal capabilities using code as an intermediate medium to bridge vision and language.
• Quantify ChatGPT's reasoning, factuality/hallucination, and interactivity effects in multi-turn prompts.
• Investigate GPT-4 and compare its performance to ChatGPT within the proposed framework.
• Provide a publicly reproducible evaluation protocol and dataset access for researchers.

Proposed method

• Compile results from 23 public datasets covering 8 NLP tasks under zero-shot settings.
• Assess multilinguality by grouping languages into high/medium/low/extremely low-resource categories.
• Explore multimodal capabilities via a code-based SVG drawing task to simulate vision-language interaction.
• Evaluate reasoning across 10 categories using curated QA datasets (including deductive, inductive, abductive, temporal, spatial, mathematical, causal, and commonsense).
• Measure factuality and hallucination using COVID-related datasets and TruthfulQA benchmarks.
• Examine interactivity through multi-turn prompting in summarization, translation, and multimodal generation tasks.

Experimental results

Research questions

• RQ1 How does ChatGPT perform in zero-shot multitask evaluation across 23 datasets and 8 tasks?
• RQ2 What are ChatGPT's strengths and weaknesses in multilingual understanding and generation, especially for low-resource languages?
• RQ3 Can ChatGPT's intermediate-code representation enable multimodal capabilities between text and vision?
• RQ4 How reliable is ChatGPT's reasoning across various reasoning categories and datasets?
• RQ5 To what extent does interactive prompting (multi-turn) improve ChatGPT's performance in summarization, translation, and multimodal tasks?

Key findings

• ChatGPT outperforms previous zero-shot models on 9/13 datasets and can surpass some fine-tuned models on certain tasks.
• Performance degrades for extremely low-resource languages and non-Latin scripts, especially in translation.
• ChatGPT shows weaknesses in inductive, multi-hop, spatial, and mathematical reasoning, while commonsense reasoning is relatively strong.
• Hallucinations persist, with extrinsic inaccuracies evident across tasks; some verification refusals observed on societal claims.
• Interactive, multi-turn prompting yields measurable gains (e.g., 8% ROUGE-1 in summarization, 2% ChrF++ in MT) and enables prompt engineering-like improvements.
• GPT-4 generally matches or exceeds ChatGPT across many tasks, with notable gains in inductive, mathematical, multi-hop, temporal, and spatial reasoning.