PaperlayerView exact revision on arXiv

Source-linked AI summary

Large Language Models on Wikipedia-Style Survey Generation: an Evaluation in NLP Concepts

Fan Gao, Hang Jiang, Rui Yang, Qingcheng Zeng, Jinghui Lu, Moritz Blum, Dairui Liu, Tianwei She, Yuang Jiang, Irene Li

arXiv:2308.10410v4cs.CL

TL;DR

LLMs' effectiveness and limitations in scientific education remain underexplored, despite the effort and cost involved in creating and updating specialized survey materials. This study evaluates LLM-generated NLP concept surveys across models and prompting conditions, finding strong GPT-4 performance alongside limitations in content and evaluation.

  • Problem

    LLMs' effectiveness and limitations in generating scientific education surveys remain underexplored, despite the expert effort and cost of creating and updating them.

    Source: §1 · p. 1

  • Method

    The authors compare four language models across prompting settings for NLP surveys and assess outputs with automatic benchmarks, human experts, and GPT evaluators.

    Source: §1 · pp. 1–2, §2.2 · p. 2

  • Results

    GPT-4 outperformed other baselines by around 2% to 20% when prompts were enhanced; its surveys rivaled human-expert quality, but GPT-4 evaluation showed bias toward machine-generated texts.

    Source: §2.3 · p. 3, §4 · p. 5, §3 · p. 5

  • Takeaways & Limitations

    The findings suggest generative LLMs could help structure domain-specific knowledge for general learners in education.

    Source: §4 · p. 5

    GPT-4 surveys can omit detail or contain factual inaccuracies, requiring extra verification, and GPT evaluation bias can skew quality judgments.

    Source: Limitations and Ethical Considerations · p. 5

Abstract

from arXiv · show

Educational materials such as survey articles in specialized fields like computer science traditionally require tremendous expert inputs and are therefore expensive to create and update. Recently, Large Language Models (LLMs) have achieved significant success across various general tasks. However, their effectiveness and limitations in the education domain are yet to be fully explored. In this work, we examine the proficiency of LLMs in generating succinct survey articles specific to the niche field of NLP in computer science, focusing on a curated list of 99 topics. Automated benchmarks reveal that GPT-4 surpasses its predecessors, inluding GPT-3.5, PaLM2, and LLaMa2 by margins ranging from 2% to 20% in comparison to the established ground truth. We compare both human and GPT-based evaluation scores and provide in-depth analysis. While our findings suggest that GPT-created surveys are more contemporary and accessible than human-authored ones, certain limitations were observed. Notably, GPT-4, despite often delivering outstanding content, occasionally exhibited lapses like missing details or factual errors. At last, we compared the rating behavior between humans and GPT-4 and found systematic bias in using GPT evaluation.

1 Introduction

LLMs perform strongly on general NLP tasks, but their effectiveness and limitations in domain-specific scientific education remain underexplored. This study evaluates LLM-generated NLP concept surveys and their alignment with human judgment.

1 Introduction

  • The study presents itself as the first evaluation of LLMs generating educational surveys in scientific NLP, addressing whether they can explain concepts in a structured manner (Li et al., 2022).The motivation is that LLM-generated text can contain false information or hallucinations, while survey generation aims to create structured overviews.Source: §1 · p. 1
  • The paper asks how proficient LLMs are at generating NLP surveys, whether they can emulate human judgment, and whether their evaluations favor machine-generated over human-written text.These questions cover survey quality, evaluation alignment, and possible bias in rating generated texts.Source: §1 · p. 1
  • Experiments compare LLaMa2, PaLM2, GPT-3.5, and GPT-4 across four settings, with human experts adding a qualitative perspective.The authors report releasing the generated surveys.Source: §1 · pp. 1–2

2 Method

The study uses Wikipedia-style NLP concept surveys to compare language models and prompting conditions, then evaluates the generated surveys with automatic metrics and human and GPT-based judgments.

2.1 Dataset

  • Surfer100 contains 100 manually written Wikipedia-style NLP concept articles, each organized into five sections of 50–150 words.The sections are Introduction, History, Key Ideas, Uses/Applications, and Variations.Source: §2.1 · p. 2

2.2 Experiment Setup

  • The experiments compare zero-shot, one-shot, description, and combined one-shot-plus-description prompts, using a Word2Vec article as the one-shot example.GPT-4 OS+IR additionally links to Wikipedia and web data; the primary analyses focus on settings without external data.Source: §2.2 · pp. 2–3

2.3 Evaluation Metrics

  • GPT-4 consistently outperformed other baselines by around 2% to 20% when prompts were enhanced, with GPT-4 OSP ranking highest in most situations.Prompt enrichment was not uniformly beneficial: LLaMa2 did better with one-shot or description prompts than OSP, while PaLM2 results were similar across prompt types.Source: §2.3 · p. 3
  • Two NLP experts and GPT evaluators rated GPT-4 OSP surveys on five-point scales, and the paper reports strong consistency with human judgment under specific guidance.Completeness received marginally the lowest scores, and GPT-4 and G-Eval showed more rating variability across sessions than human experts.Source: §2.3 · p. 3

3 Analysis

The analysis finds that GPT-4 OSP surveys most often omit information, while GPT-4’s entity novelty is not clearly associated with performance. GPT-4 also favors its generated surveys when evaluating them.

Error Types

  • Most GPT-4 OSP errors were missing information, followed by verbosity and factual inaccuracies; History and Introduction had most errors, while Application performed best.The analysis classified errors as Verbose, Wrong Fact, Missing Information, or No Error.Source: §3 · p. 3

Novel Entity Mention

  • GPT-4 showed no marked novelty in entity mentions despite strong evaluations, and new-entity tendency had no significant correlation with performance.PaLM2 showed the least variation in entity mentions, while LLaMa2-13b showed the most.Source: §3 · pp. 3–4

LLM and Human Preference

  • GPT-4 favored its generated surveys, rated ground-truth surveys marginally lower across all three human-preference groups, and preferred human-disliked over human-liked ground truth.In a blind side-by-side preference comparison, the two experts reached a Cohen’s Kappa score of 0.68.Source: §3 · p. 5

4 Discussion and Conclusion

The paper finds that LLMs, particularly GPT-4, can produce NLP surveys that rival human-expert quality, although they may omit information. It also identifies bias in GPT-4’s evaluation of generated and human-written texts.

  • LLMs, particularly GPT-4, can write guideline-following NLP surveys that rival human-expert quality, though some contain incomplete information.Source: §4 · p. 5
  • GPT-4 may not replace human judgment when evaluating human-composed text, and it shows bias when rating machine-generated texts.Source: §4 · p. 5
  • The authors conclude that generative LLMs could help structure domain-specific knowledge for general learners and support more interactive, personalized learning.Source: §4 · p. 5

Limitations and Ethical Considerations

GPT-4 surveys can be contemporary and accessible but may sacrifice depth and detail, with information gaps and occasional factual errors; GPT-based evaluations also show systematic bias. The study reports no harmful content or private information in the data used or produced.

A Human Evaluation Guidance

The human evaluation guidance scores generated surveys on readability, relevance, concision, hallucination, completeness, and factual accuracy. Before evaluating all concepts, the researchers selected two judges whose scores aligned more closely.

  • The guidance contrasts bad and good generated text on readability, relevance, concision, hallucination, completeness, and factual accuracy, describing low-quality outputs as difficult, irrelevant, repetitive, misleading, incomplete, or inaccurate.Source: A Human Evaluation Guidance · p. 8
  • Four NLP specialists first assessed GPT surveys on 20 handpicked topics, but score variability led the researchers to select two more-aligned judges for the full evaluation.Source: A Human Evaluation Guidance · p. 8

B Comparisons with External Knowledge

The section compares GPT-4 with and without external knowledge from Wikipedia links or web search. Both external-knowledge variants improve ROUGE, while web search also improves MoverScore and UniEval.

B Comparisons with External Knowledge

  • For the Wikipedia-linked setting, the researchers used Embedchain and crawled 87 effective links for topics in the Surfer 100 dataset.GPT-4 received the corresponding links and a sample survey as references.Source: B Comparisons with External Knowledge · p. 8
  • The web-retrieval setting prompted GPT-4 to search the web for helpful information using web search APIs.Source: B Comparisons with External Knowledge · p. 8
  • Table 4 lists 20 selected concepts for pre-selection, while Table 5 reports human evaluation scores on 20 topics from four experts.Source: Table 4 · p. 9, Table 5 · p. 9
  • Both Wikipedia links and information retrieval significantly improve GPT-4’s ROUGE scores, and web search improves MoverScore and UniEval.The authors conclude that external knowledge yields higher-quality surveys than internal knowledge alone.Source: B Comparisons with External Knowledge · p. 9

C.1 Understanding of “Survey”

The analysis finds that models sometimes interpret “survey” as a questionnaire or produce content that is not a typical survey article. It also identifies incomplete historical accounts that can mislead readers.

C.1 Understanding of “Survey”

  • Models sometimes interpret “survey” as a questionnaire instead of generating the requested survey article.Source: §C.1 · p. 9
  • Even when the format resembles an article, GPT-3.5 zero-shot outputs may include inconsequential filler, such as repeated references to sections and participants.Source: §C.1 · p. 9

C.2 Incomplete Information

  • GPT models occasionally omit historical developments, potentially misleading readers; examples include Knowledge Graph’s long history and landmark events in Decision Trees.Source: §C.2 · p. 9

C.3 Nebulous Sentence Structure

GPT models often use vague, reusable sentence structures, particularly in the Application section. The examples include broad claims that remain applicable when the topic name is replaced by another NLP topic.

C.3 Nebulous Sentence Structure

  • In the Application section especially, GPT models use vague sentences that lack specificity and can fit different NLP topics.Examples include claims that a topic has a wide spectrum of applications or plays a vital role in Natural Language Processing and Understanding.Source: §C.3 · p. 9
  • The examples include broad application claims that retain meaning when “The Topic” is replaced with any NLP topic.Source: §C.3 · p. 9
  • Figure 5 depicts a generated questionnaire presented as a survey part.Source: Figure 5 · p. 10

C.4 High-quality Survey

The paper presents an LSTM survey as an accessible, detailed example, then compares GPT-4 output with ground truth and outlines risks of relying on LLM-generated educational content.

C.4 High-quality Survey

  • The LSTM survey is described as readable, detailed, well-structured, accurate, and concise, with a summary followed by section-level explanations.Its Applications discussion is noted for domain specificity.Source: §C.4 · p. 10

C.5 Going Beyond the Ground Truth

  • In two examples, GPT-4 gives a more informative mBERT explanation and a more precise, complete HMM account than the ground truth.The HMM response follows an algorithm-category-to-features-to-applications-to-motivation flow; the ground truth spends more space on category terminology.Source: §C.5 · p. 10

D Potential Risks

  • The authors warn that unvetted GPT-4 factual errors could propagate misinformation in education, while reliance may reduce human experts’ role and experience-based insights.They also caution that systematic bias in GPT evaluations could lead to incorrect judgments about content quality.Source: D Potential Risks · p. 10

E Experimental Details

The experiments used paid GPT-4 generation, automatic metrics, and human-evaluation tools, without fine-tuning.

E Experimental Details

  • The experiments mainly used paid GPT-4 and cost around 230 USD in total.Source: E Experimental Details · p. 11
  • Automatic evaluation used ROUGE and BERTScore APIs, while MoverScore, UniEval, and BARTScore were run from downloaded source code.Source: E Experimental Details · p. 11
  • Experiments ran on four A100 40GB GPUs; without fine-tuning, each setting finished in a few hours, and authorized APIs supplied human-evaluation agreement scores.The reported agreement measures were Krippendorff’s and Kendall’s scores.Source: E Experimental Details · p. 11
Loading 2308.10410v4…