Detection and Interpretability Analysis of Quotation Errors by Large Language Models

Technical Summary: Detection and Interpretability Analysis of Quotation Errors by Large Language Models

Problem Statement
Quotation errors, defined as inconsistencies between cited information and the original source text, pose significant risks to academic integrity. These errors can lead to the misinterpretation of original research, distort the collective understanding of the academic community, and undermine the accuracy of citation-based evaluation systems. While manual verification is the traditional standard, it is labor-intensive and inefficient. Existing automated detection methods, including similarity-based approaches and prompt-based Large Language Model (LLM) techniques, often rely solely on abstract data or lack deep interpretability. Furthermore, the potential of fine-tuning LLMs specifically for this task, combined with the integration of full-text data and interpretability analysis, remains under-explored.

Methodology
The study proposes a framework for the automated detection of quotation errors that integrates three core components: fine-tuning, full-text data integration, and interpretability analysis.

1. Fine-Tuning Approach: The authors employ Low-Rank Adaptation (LoRA) to fine-tune large language models (specifically Qwen-2.5-7B and DeepSeek-R1-Distill-Llama-8B) for the task of quotation error detection. This is compared against baseline models including SCIBERT, T5, and various prompt-based LLM configurations (zero-shot, few-shot, and Retrieval-Augmented Generation).
2. Dataset Construction: Two open-source datasets, SCIFACT and NealSmith, were utilized. The study constructed both abstract-level and full-text-level datasets. For the full-text integration, three methods were compared:
   • Selecting full-text sentences based on semantic similarity to the target citation sentence.
   • Selecting full-text sentences based on semantic similarity to the cited paper's abstract.
   • Directly incorporating the full text (evaluated only on models supporting ultra-long contexts).
   • Labels were standardized into three categories: SUPPORT, CONTRADICT, and NULL.
3. Interpretability Analysis: To address the "black box" nature of model predictions, the study applies the TokenSHAP method. This tool quantifies the contribution of individual tokens in the input prompt to the model's output, visualizing which parts of the citation or source text drive the classification decision.
4. Error Analysis: A manual analysis of misclassified predictions by the fine-tuned models was conducted to identify specific failure modes, such as issues with prior knowledge, quantitative reasoning, and semantic nuance.

Key Results

• Performance of Fine-Tuning: Fine-tuning LLMs demonstrated superior performance compared to both traditional Pre-trained Language Models (PLMs) and prompt-based LLM approaches. On the SCIFACT abstract-level dataset, the fine-tuned Qwen-7B model achieved a macro-averaged F1 score of 87.37%, outperforming the best prompt-based model (Qwen-Max) by 2.43%. Similar improvements were observed on the NealSmith dataset.
• Impact of Full-Text Integration: The integration of full-text data improved detection performance, particularly on the NealSmith dataset which was annotated with full-text context. Among the integration methods, selecting full-text sentences based on similarity to the abstract yielded the best performance (Macro-F1 50.66% for Qwen-7B on NealSmith), outperforming direct full-text incorporation and citation-similarity-based selection.
• Error Patterns: Analysis revealed that pre-fine-tuned models often relied on shallow keyword matching, incorrectly predicting "SUPPORT" based on topic overlap even when evidence was absent. Fine-tuned models showed improved ability to evaluate evidence sufficiency and semantic consistency. Specific error types identified included confusion between prior knowledge and explicit text, difficulties with quantitative relationships (e.g., unit changes), and misinterpretations of "increment" versus "absolute quantity."
• Interpretability: TokenSHAP visualization confirmed that the fine-tuned model successfully identified key semantic contradictions (e.g., specific directional features of microtubules) that the prompt-based model missed, providing a transparent view of the model's reasoning process.

Significance and Claims
The paper claims three primary contributions to the field of automated quotation error detection:

1. Methodological Advancement: It is the first study to apply LLM fine-tuning specifically to the quotation error detection task, demonstrating that task-specific parameter optimization significantly enhances performance over prompt engineering and traditional PLMs.
2. Data Granularity: The study systematically evaluates the impact of incorporating full-text data, identifying that abstract-based sentence selection is the optimal strategy for integrating full-text information in this context.
3. Interpretability: By applying TokenSHAP, the study moves beyond quantitative metrics to provide a qualitative understanding of why models make specific predictions, thereby increasing trust in automated detection systems.

The authors maintain a modest stance, acknowledging limitations such as the relatively small size of the NealSmith dataset and the potential imprecision introduced by merging classification labels during dataset re-annotation. They conclude that while the fine-tuning approach shows promise, further research is needed to address specific reasoning limitations and expand dataset scales.