View Proposal

---

Proposer

Zi Hau Chin

Title

Automated Detection of Hallucinated and Unsupported Citations

Goal

Design and Evaluation of a PDF-Based System for Detecting Hallucinated and Incorrect Citations in Documents

Description

This project involves developing an automated system that accepts an academic paper in PDF format, extracts its in-text citations and reference list, and verifies whether the cited publications exist. The system will analyse bibliographic information such as article titles, authors, publication years, journals, page numbers, and Digital Object Identifiers. It will retrieve and compare publication metadata from scholarly sources such as Crossref, OpenAlex, arXiv, or Semantic Scholar. Citations may then be classified as verified, partially incorrect, potentially fabricated, possible combinations of unrelated publication details, or unverifiable. The project will also examine whether a valid publication supports the claim associated with its citation. The system should identify the sentence or paragraph containing the citation, retrieve available evidence from the publication’s title, abstract, or legally accessible full text, and apply semantic similarity or natural-language inference techniques. Each citation may be classified as supported, partially supported, contradicted, or containing insufficient evidence. The system must present the retrieved evidence and explain why a citation has been flagged rather than relying on an unsupported language-model judgement. The final prototype should provide an interactive PDF viewer in which citations are highlighted and colour-coded according to their verification status. Users should be able to select a citation, inspect the extracted reference, compare it with the matched publication, view metadata discrepancies, examine supporting evidence, and export a verification report. The student will evaluate citation extraction, reference linking, bibliographic verification, claim-support classification, processing time, uncertainty handling, and usability. The system should be presented as an academic verification assistant rather than an automatic authority for determining research misconduct. Corresponds to a real publication? -Has accurate authors, title, year, venue, volume, pages, and DOI? -Combines information from several unrelated publications? -Has been attached to a claim that the source does not support? The system should: -Process text-based academic PDFs -Detect the reference section -Extract at least two major citation styles -Link in-text citations to references -Verify citations through at least two scholarly sources -Flag major metadata inconsistencies -Highlight the citation location in the PDF -Export an explainable report -Outperform a simple DOI-only or exact-match baseline

Resources

VeraCite, Verify Citation, BibGuard, CheckIfExist

Background

Python programming and text processing, REST APIs and structured data formats such as JSON, Natural language processing and semantic similarity , PDF processing and information extraction

Url

Difficulty Level

High

Ethical Approval

Full

Number Of Students

1

Supervisor

Zi Hau Chin

Keywords

citation hallucination, reference verification, bibliographic entity matching, doi validation, semantic similarity, natural language inference

Degrees

Bachelor of Science in Computing Science