PaperlayerView exact revision on arXiv

Source-linked AI summary

Investigating Bias in LLM-Based Bias Detection: Disparities between LLMs and Human Perception

Luyang Lin, Lingzhi Wang, Jinsong Guo, Kam-Fai Wong

arXiv:2403.14896v2cs.CY

TL;DR

Media-bias detection matters amid social-media misinformation, while the use of LLMs for prediction motivates examining bias within the systems themselves. The paper studies political bias in prediction and continuation, topic variation, and debiasing, and reports varied bias leanings across tested models without an apparent performance-bias correlation.

  • Problem

    LLMs are increasingly used for media-bias prediction, but it remains important to examine whether bias within those systems affects prediction, including when weak results may reflect capability or inherent bias.

    Source: §1 · pp. 1–2

  • Method

    The study examines system bias through political article prediction and text continuation, explores bias across topics and models, and investigates prompt engineering and fine-tuning as debiasing approaches.

    Source: §1 · p. 2

  • Results

    Tested LLMs show varying bias leanings, bias is present across them, and performance does not appear to correlate with degree of bias.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The authors report inherent LLM biases with significant implications for media-bias detection and highlight the need for debiasing strategies.

    Source: §7 · p. 9

    The analysis relies on two political-bias datasets with three-way labels and assumes human labels are unbiased, although annotations can be subjective.

    Source: Limitations · p. 9

Abstract

from arXiv · show

The pervasive spread of misinformation and disinformation in social media underscores the critical importance of detecting media bias. While robust Large Language Models (LLMs) have emerged as foundational tools for bias prediction, concerns about inherent biases within these models persist. In this work, we investigate the presence and nature of bias within LLMs and its consequential impact on media bias detection. Departing from conventional approaches that focus solely on bias detection in media content, we delve into biases within the LLM systems themselves. Through meticulous examination, we probe whether LLMs exhibit biases, particularly in political bias prediction and text continuation tasks. Additionally, we explore bias across diverse topics, aiming to uncover nuanced variations in bias expression within the LLM framework. Importantly, we propose debiasing strategies, including prompt engineering and model fine-tuning. Extensive analysis of bias tendencies across different LLMs sheds light on the broader landscape of bias propagation in language models. This study advances our understanding of LLM bias, offering critical insights into its implications for bias detection tasks and paving the way for more robust and equitable AI systems

1 Introduction

Because misinformation and disinformation spread on social media, media-bias detection matters; as LLMs are increasingly used for prediction, this section motivates examining bias within the systems themselves. It outlines investigations of political bias, topic variation, debiasing, and differences across models.

1 Introduction

  • Vanilla GPT-3.5 scores 26.2% F1 on FlipBias, raising whether weak political-leaning prediction reflects limited capability or inherent model bias.Source: §1 · p. 2
  • The study examines political bias through article-prediction and text-continuation tasks, addressing model judgments of input articles and the leaning of generated content.Source: §1 · p. 2
  • Unlike work organized around predefined demographic dimensions, the study assesses bias across predefined and latent topics to examine whether it is consistent.The authors also describe topic-specific cases and evaluation metrics.Source: §1 · p. 2
  • The proposed debiasing investigations use prompt engineering to isolate inherent bias and fine-tuning to adjust a model’s leaning.Source: §1 · p. 2
  • Different tested LLMs show varying political-bias leanings, bias is present, and model performance does not appear to correlate with degree of bias.The investigation also frames the human-labeled ground truth as a basis for examining disparities between LLMs and human perception.Source: §1 · p. 2
  • The paper examines open- and closed-source LLMs and reports that tested models vary in bias leaning, while bias is present across them.Source: §1 · p. 2

2 Related Work

Related work describes LM bias as context-sensitive, with manifestations in outputs and system performance. It also surveys mitigation approaches spanning input changes, inference-time interventions, training changes, and output processing.

2 Related Work

  • LM bias is complex and subjective, shaped by contextual and cultural factors, and studied through manifestations such as negative representation, disparate performance, and norm reinforcement.Source: §2 · p. 2
  • Prior work addresses LM bias broadly, while this paper focuses on political-standing bias and discrepancies between LM cognition and human perceptions.Source: §2 · p. 2
  • Bias mitigation spans input preprocessing, inference-time intraprocessing, training interventions, and post-processing of model outputs.Examples include prompt modification, decoding strategies, changes to optimization, and replacing biased tokens.Source: §2 · pp. 2–3

3 RQ1: Do LLMs exhibit political bias?

The tested LLM shows left-leaning tendencies in political-bias prediction and generated continuations, with continuation leanings varying by prefix length. The section examines prediction and continuation as complementary ways to assess political bias.

3.1 LLM-based Bias Prediction

  • Across FlipBias and ABP, the tested LLM shows left-leaning prediction bias, yet classifies right-grounded articles more accurately than left-grounded ones.The authors infer left-leaning bias from more Left-to-Center than Right-to-Center predictions; Right-to-Right predictions also substantially exceed Left-to-Left predictions.Source: §3.1 · p. 3

3.2 LLM-based Article Continuation

  • Continuation bias is assessed by matching generated suffixes to same-event article embeddings and by comparing left- and right-leaning vocabulary counts.Embedding matching assigns the label of the most similar same-event article instance; vocabulary matching uses separate word-frequency vocabularies for left- and right-leaning articles.Source: §3.2 · p. 4
  • Across 20–80-token prefixes, both matching methods yield more left-leaning continuations, but at 320 tokens both shift toward right-leaning suffixes.The authors suggest longer prefixes reveal more of the article's political leaning; their explanation also notes that right-leaning articles average 794 tokens versus 1,111 for left-leaning articles.Source: Figure 4 · p. 4, §3.2 · pp. 4–5
  • About 70% of generated articles are labeled Center by vanilla ChatGPT, versus around 15% matched as Center by the other methods.The authors note that ChatGPT may regard text it generated as objective and question whether a potentially biased classifier should measure bias.Source: §3.2 · p. 5

4 RQ2: Do LLMs demonstrate consistent bias across all topics?

The section finds that LLM political leaning varies across topics rather than remaining consistent, and compares topic-level patterns across FlipBias and ABP. Its analyses also examine how topic frequency and dataset context relate to those patterns.

  • FlipBias latent-topic and ABP predefined-topic joint plots show similar patterns, although the predefined-topic distributions are more focused than the latent-topic distributions.Figure 5 uses kernel density estimates; points farther from the origin indicate more incorrect judgments, and deviation from x=y reflects imbalance between right- and left-classified cases.Source: §4 · p. 5, Figure 5 · p. 5
  • Most topics show left-leaning positive BTI-1 values, but some show right-leaning negative values; BTI-2 is more evenly spread across positive and negative values.The overall left-leaning tendency therefore does not hold uniformly across topics.Source: Figure 6 · p. 6, §4 · p. 6
  • Across the datasets, extreme-value topics show similarities: positive values often concern security and terrorism, while negative values often concern Trump’s government and the US-China trade war.The authors infer that short-term hot topics such as coronavirus tend to be negative, whereas broader topics such as LGBT rights trend positively.Source: Figure 7 · p. 6, §4 · p. 7
  • FlipBias latent-topic clusters have a more even instance-frequency distribution than ABP topics, while widely discussed topics are reported to reflect human perspectives more closely.The section notes that data scale may influence bias trends.Source: §4 · p. 6
  • FlipBias case analyses associate positive topic values with subjects such as journalism citations, Obama’s policies, and immigration critique, and negative values with Trump-administration topics.The authors suggest training-data and structural biases may shape the former pattern and note a possible correlation between FlipBias’s 2013–2018 collection period and the latter.Source: §4 · pp. 6–7, Table 3 · p. 7

5 RQ3: How to debias LLMs and further improve performance?

The section tests prompt-based and fine-tuning approaches to debiasing, assessing overall bias, topic-level bias, and prediction performance. Results show trade-offs between reducing bias and improving performance.

5 RQ3: How to debias LLMs and further improve performance?

  • The debiasing exploration covers in-context strategies and fine-tuning, evaluating overall debiasing, topic-level bias elimination, and possible prediction-performance gains.Source: §5 · p. 7

5.1 Prompts to Isolate Bias

  • The prompt strategies include label explanations, few-shot headline examples, and a statement asking the model to avoid bias and stereotypes.The label explanations draw on Wikipedia, and the few-shot experiments use 3, 6, or 12 examples.Source: §5.1 · p. 7

5.2 Fine-Tuning to Debias

  • Fine-tuning tests use 300 labeled instances in three left/center/right label-proportion settings, motivated by a suspected weakness in recognizing left-leaning evaluation criteria.The settings are L-FT, LC-FT, and LCR-FT; LCR-FT is also applied to a BERT baseline.Source: §5.2 · pp. 7–8

5.3 Assessment of Debiasing Strategies

  • Prompt-engineering debiasing centers topic BTI values around 0.0, while LCR-FT shifts the distribution in a way that may accompany better performance but more severe bias.Figure 8 shows FlipBias topic BTI distributions after debiasing.Source: Figure 8 · p. 8, §5.3 · p. 8
  • Fine-tuning generally improves bias-prediction performance but increases BTI-1 or BTI-2, while prompt-based methods—especially DS—are reported as easy and effective.Table 4 reports the FlipBias debiasing results.Source: Table 4 · p. 8, §5.3 · p. 8
  • Bias Label Explanation and Debiasing Statements make the model more cautious in biased judgments at an accuracy cost, whereas few-shot and fine-tuning methods improve performance but mitigate bias less effectively.Source: §5.3 · p. 8

6 RQ4: Do various LLMs exhibit similar bias tendencies?

The comparison extends the analysis from GPT-3.5 to closed- and open-source LLMs. Although LLaMa2 and Mistral have better political-bias classification metrics than GPT-3.5 and GPT-4, their outputs raise important qualifications.

  • The comparison covers Llama-2-7B-Chat, Vicuna-7B-v1.5, Mistral-7B-v0.1, and gpt-4-0125-preview alongside the earlier GPT-3.5 analysis.Source: §6 · p. 8
  • LLaMa2 and Mistral outperform GPT-3.5 and GPT-4 on current political-bias classification metrics, but deny answering or produce unrelated content on about 20% of testing; almost all tested LLMs exhibit bias, with Mistral generally right-leaning.The topic-level distribution is reported as consistent with the overall bias results.Source: Figure 9 · p. 9, §6 · p. 9

7 Conclusion

The conclusion identifies inherent LLM bias as relevant to media-bias detection and emphasizes political-bias prediction as a focus of the investigation. It presents debiasing strategies as important to understanding bias propagation in language models.

  • The study identifies inherent biases within LLMs and their significant impact on media-bias detection.Source: §7 · p. 9
  • The investigation focuses particularly on bias in political-bias prediction within LLM systems.Source: §7 · p. 9
  • The findings highlight the need for debiasing strategies and offer insight into bias propagation in language models.Source: §7 · p. 9

Limitations

The paper notes limits from the availability of three-way political-leaning labels and from treating human annotations as unbiased ground truth. The appendices also describe vocabulary and latent-topic construction, plus prompt details.

Limitations

  • Bias analysis relies on two political-prediction datasets with left, center, and right labels; binary biased/non-biased labels are insufficient for this analysis.Human-labeled data is assumed to be unbiased ground truth, although annotators’ subjectivity and individual biases may affect evaluation validity.Source: Limitations · p. 9

A Left-Right Vocabulary Corpus Construction

  • The corpus includes tokens in its Left vocabulary only when their frequency normalized by total Left-article tokens exceeds twice their corresponding normalized frequency in Right articles.It selects the top 2,000 Left tokens and 1,295 Right tokens to match the Left tokens’ total frequency sum.Source: A Left-Right Vocabulary Corpus… · p. 11

B Latent Topic Construction

  • Latent topics are formed by clustering 19,671 article bias indicators into 152 clusters using hierarchical clustering with Euclidean distance and threshold 2.ChatGPT constructs fine-grained indicators from FlipBias; each cluster represents a latent topic.Source: B Latent Topic Construction · p. 11

C.1 Details of Prompts to Isolate Bias

  • The prompt-method appendix provides bias-label explanations and four randomly selected Left-Center-Right FlipBias triples as few-shot examples.The appendix introduces these as details of the debiasing methods and the Few-shot Instruction method, respectively.Source: §C.1 · p. 12

C.2 More Finetuning Implementation Details

The GPT finetuning implementation uses 300 randomly selected ABP instances and specifies three epochs with a batch size of 32.

C.2 More Finetuning Implementation Details

  • GPT-3.5-turbo was fine-tuned through the OpenAI API.Source: §C.2 · p. 12
  • The training set consisted of 300 instances randomly selected from ABP.Source: §C.2 · p. 12
  • Fine-tuning used 3 epochs and a batch size of 32.Source: §C.2 · p. 12

D Debiasing Results

The debiasing appendix reports topic-level BTI distributions after prompt debiasing, distinguishes the study’s systematic scope, and finds a predominant left-leaning continuation trend with T5-Base embeddings.

E.1 Distinguish from Related Works

  • The study emphasizes systematic bias detection across four research questions, intricate experiments, and a broader range of LLMs.Source: §E.1 · p. 13

D Debiasing Results

E.2 Exploration of Different Embeddings

  • T5-Base embeddings show a predominant left-leaning trend in continued articles, similar to GPT-3.5 embeddings, with slight prediction variations at prefix length 320.The appendix compares continuation results across embeddings; left and right percentages follow the Figure 4 methodology.Source: §E.2 · pp. 13–14

E.3 Result of Article Continuation based on classifier

For LLaMa2, finetuning leaves average BTI-1 scores without significant change but yields a more centralized topic-level BTI-1 distribution. The appendix also describes classifier settings and continuation examples.

E.3 Result of Article Continuation based on classifier

  • The GPT-3.5-turbo-0613 classifier experiment compares zero-shot labeling with few-shot labeling using a same-event Left-Center-Right article triple; Left/Right ratios are reported in Table 10.Source: §E.3 · p. 14, Table 10 · p. 15

E.4 Article Continuation Examples

  • GPT-3.5 article continuations have average suffix lengths of 490.1, 487.5, 479.0, 463.7, and 473.9 tokens for prefix lengths 20, 40, ..., 320, respectively.The paper says the generated suffixes are quite consistent with their prefixes and provides randomly selected examples.Source: §E.4 · p. 14, Table 11 · p. 15

E.5 Finetuning Other LLMs

  • LLaMa2 finetuning yields a more centralized topic-level BTI-1 distribution, although average BTI-1 scores do not change significantly.Figure 12 presents the topic distribution for LLaMa2 and LLaMa2 LCR-FT; Table 13 reports BTI-1 scores.Source: §E.5 · p. 14, Figure 12 · p. 16, Table 13 · p. 16

Related papers

Loading 2403.14896v2…