Daejeon, July 27–31
Influence studies spanning intertextuality, text reuse, and citation networks depend on reliable mappings between referenced passages and their source texts. However, large-scale dataset construction is hindered by "dirty" citations in scholarly PDFs and multilingual documents: inconsistent notations, abbreviations, and OCR-induced noise. This paper presents a human-in-the-loop workflow and a web-based tool that resolves unstructured citations against a federated corpus architecture integrating four heterogeneous text collections: 70,000+ Project Gutenberg works, a fully indexed KJV Bible, the Aozora Bunko Japanese literary corpus, and a curated Coptic text collection drawn from the Coptic SCRIPTORIUM. An optional LLM layer automates abbreviation extraction and contextual reference resolution, while human expert verification at each stage ensures dataset quality. The system exports verified influence pairs as structured datasets with concrete downstream applications for computational literary studies and generative AI.
Influence studies examining intertextuality, text reuse, and citation networks are foundational to literary scholarship (Kristeva, 1980; Fowler, 2000). While computational approaches to stylistic comparison (Piper and Algee-Hewitt, 2014; Long, 2021) and word embeddings (Li, 2024; Kondo, 2024) have advanced the field, the integration of generative AI remains constrained by a critical bottleneck: the scarcity of high-quality, large-scale datasets that map "influence pairs" with precision. Arnold and Jäschke (2023) demonstrate that automated extraction from screduces
holarly texts yields significant noise, while Karjus (2025) identifies dataset quality as a primary constraint on machine-assisted humanities research. Therefore, human-judged high-quality datasets are critically needed.
Existing frameworks such as Canonical Text Services (CTS) offer structured precision but lack scalability, covering only limited, curated corpora. General-purpose annotation tools such as INCEpTION and BRAT support named entity and relation labeling but do not address the specific challenge of resolving physical citation locations (page, line) to logical source structures (book, chapter, verse), nor do they handle the dirty citation problem inherent in humanities PDFs. To address this gap, we propose a human-in-the-loop workflow and web-based tool that bridges unstructured PDFs and a federated multi-corpus backend, augmented by an optional LLM layer for document-specific citation intelligence.
Our tool integrates machine efficiency with human expert verification through a federated corpus architecture that provides unified citation resolution across four heterogeneous text collections: Project Gutenberg (70,000+ works, UTF-8 plain text), the KJV Bible (41,899 indexed verses), Aozora Bunko (Japanese literary texts, Shift-JIS encoded archives), and a curated Coptic corpus drawn from the Coptic SCRIPTORIUM. Despite their radically different underlying formats, access methods, and encoding schemes, each collection exposes a common candidate interface to the resolution layer, enabling ranked results across all corpora from a single query. The system is implemented as a web application with a JavaScript frontend built on PDF.js (v3.11.174). The workflow consists of four phases.
Figure 1: Resolution workflow from PDF citation input to verified structured output.
The user uploads a PDF; a Split View mode enables simultaneous display of body text and footnotes on separate panels, supporting efficient cross-referencing. The system extracts textual content and its physical location across scripts and languages — Latin, CJK, Coptic, and others — using PDF.js with CMap support. These languages were selected to stress-test the pipeline across maximally distinct scripts, resource levels, character encoding systems (including two-byte character setscharacters), and citation traditions, evaluating the generalizability of the parsing and resolution approach.
A central challenge in humanities PDFs is document-specific abbreviations (e.g., "PL" for Paradise Lost) and relative references ("Ibid.", "Op. cit."). The optional LLM layer addresses both: given an Anthropic API key, the system automatically detects the abbreviations page via regular expression, renders it as a high-resolution image, and submits it to a multimodal LLM (Claude, claude-sonnet-4) via the Vision API, returning a structured JSON list of abbreviation–expansion pairs that is cached and injected into the resolution pipeline. The same LLM layer resolves relative references by interpreting full citation context rather than relying on rule-based backtracking. This feature currently requires users to supply their own API key — a limitation we aim to address in future deployment. A rule-based fallback and manual abbreviation entry remain available for API-free use.
The resolution engine applies a cascaded search across the federated corpus. Exact matching via hash table lookup handles unambiguous references. For noisier input, Jaccard similarity with substring-based boosting ranks title variants against the Gutenberg catalog. Scripture references in standard formats (e.g., "Gen 1:1", "Matthew 5:3–5") are resolved against the indexed Bible corpus. Japanese literary references route to the Aozora Bunko backend, where server-side processing handles ZIP extraction, Shift-JIS to UTF-8 conversion, and removal of ruby and annotation markup before returning clean plain text. Coptic references are resolved against the Coptic SCRIPTORIUM collection, which provides TEI-encoded texts that are normalized to plain Unicode for display and alignment. Each corpus adapter is independently maintainable and new collections can be integrated by implementing the shared candidate interface, making the architecture extensible to further languages and repositories such as HathiTrust.
Machine-retrieved source text is presented in a full-text viewer supporting up to 50,000 lines with Shift+Click range selection. The annotator confirms the match or adjusts the referenced range. Each confirmed pair requires a five-level influence rating — Similar Author, Shared Theme, Moderate Influence, Strong Influence, or Direct Quotation — before saving, ensuring metadata quality throughout. A step-count analysis illustrates the efficiency gain: traditional manual resolution requires six operations per pair (locating the passage, decoding abbreviations, corpus search, transcription, verification, and formatting), whereas our tool reduces this to two user actions (highlighting and confirmation), with intervening steps handled by the resolution engine. This reduces research labor and opens opportunities for more non-technical researchers. The optional LLM layer further reduces the preliminary cost of abbreviation setup from manual list construction to a single API call. Formal timing evaluation across annotators is planned as future work.
Beyond generative AI, the verified influence pairs serve multiple scholarly purposes: they constitute reusable ground-truth datasets for reproducible computational literary studies (contributing also to disciplinary knowledge inheritance), can directly enrich library metadata and knowledge graphs such as Wikidata and VIAF, and support citation network analysis without any machine learning component. The federated architecture is particularly significant for diversity: by integrating Japanese and Coptic corpora alongside Western literary collections, the system moves beyond the high-resource, Anglophone bias that characterizes most existing tools, and provides a practical framework for low-resource and non-Western literary traditions.
The tool exports verified pairs as structured datasets (JSON/CSV), enabling: (1) intertextuality detection via BERT- or Siamese-style encoders trained on verified pairs; (2) citation style transfer by fine-tuning sequence-to-sequence models; (3) cross-lingual citation alignment across languages and scripts; and (4) influence network analysis using graph-based models. Future development will expand corpus coverage and improve handling of low-resource languages.
This work bridges a persistent gap between unstructured scholarly texts and structured digital corpora. By combining a federated multi-corpus architecture, an optional LLM layer for document-specific citation intelligence, and expert verification, the system provides a scalable and extensible pathway to high-quality ground-truth datasets for influence studies. The unified resolution interface allows heterogeneous collections spanning Western, Japanese, and Coptic literary traditions to be queried and annotated within a single workflow, directly addressing the multilingual and cross-cultural demands of contemporary influence studies. Ultimately, the tool democratizes dataset creation by keeping computational workflows grounded in scholarly verification standards while expanding coverage well beyond Western, high-resource corpora.