Bibliographic Fingerprints: Authorship & Duplicate Detection

• Bibliographic fingerprints are compact, intentionally discriminative signatures derived from bibliographic fields to detect duplicates and support authorship attribution.
• The technique incorporates metadata-based methods using titles and author names alongside textual stylometric analysis to capture stable linguistic attributes.
• Empirical evaluations reveal high precision and recall, underscoring their effectiveness in managing heterogeneous bibliographic databases.

Bibliographic fingerprints are compact, intentionally discriminative signatures used to identify, compare, and authenticate documents in bibliographic systems. In one established sense, they are short keys derived from bibliographic fields such as title and author in order to detect duplicates across heterogeneous databases; in another, they are stylometric signatures derived from linguistic attributes that remain repeatably distributed for an author and can therefore support attribution, provenance checks, de-duplication, and corpus-integrity analysis. The common principle is reduction: instead of relying on a full record or a full text, a fingerprint encodes a small set of stable signals that remain useful under noise, formatting variation, or incomplete metadata (Turenne, 2015, 0802.2234).

1. Conceptual scope and bibliographic function

The term covers two closely related but methodologically distinct practices. The first is metadata-oriented duplicate detection, where a fingerprint is a compact key derived from bibliographic fields such as title, author, venue, and year. The second is content-oriented textual fingerprinting, where a fingerprint is a vector of numerical linguistic attributes characteristically distributed for a given writer. In library and digital-collection workflows, the latter can complement metadata-driven bibliographic fingerprints, especially when attribution is uncertain or paratext is unreliable (0802.2234).

This distinction is operational rather than absolute. Traditional bibliographic fingerprints rely on stable paratextual cues such as title pages and colophons, whereas textual fingerprints operate on the linguistic content itself. The content-oriented approach is used for author identification and disambiguation, edition or work-level provenance checks, de-duplication and clustering by author or style, and integrity checks such as detecting intrusions or later edits within a corpus. The metadata-oriented approach is optimized for efficient exact matching across heterogeneous sources and is designed to remain robust to punctuation, capitalization, hyphenation, token-boundary cuts, and similar indexing artifacts (Turenne, 2015, 0802.2234).

A recurring misconception is that stable identifiers such as DOI make fingerprinting unnecessary. The duplicate-detection study explicitly rejects that assumption: DOI is not universally present, particularly in older records, and core bibliographic fields often differ slightly across sources. Another misconception is that topical similarity suffices for textual attribution. The stylometric study instead emphasizes stylistic, largely content-agnostic properties such as function-word behavior, sentence and clause structure, and vocabulary-distribution statistics, because topical cues are volatile across themes whereas many stylistic habits are comparatively unconscious and stable (Turenne, 2015, 0802.2234).

2. Record-level fingerprints for duplicate detection

In duplicate detection, a bibliographic fingerprint is a short ordered signature derived from a record’s title and, in one method, the first author’s name. Records are preprocessed by lower-casing selected fields and tokenizing titles on delimiters such as spaces and punctuation. Stopword removal is not applied. The paper defines characters over an alphabet $A$, delimiters $D$, and record fields $F_k$ such as AU, TI, SO, and DP. The fingerprint then becomes a blocking key: records sharing the same key are treated as candidate duplicates, which reduces comparisons from $O(NM)$ to $O(N \log M)$ through hash lookups (Turenne, 2015).

The principal method is the Socio-Semantic Fingerprint (SSF), which combines lexical and minimal social information. Let $W=(W_1,\ldots,W_{|W|})$ be the title tokens, $A_1$ the first-author string, and $b(\cdot)$ the operator returning the first two characters of a token. With $N=8$, the fingerprint is

$$F_{\mathrm{SSF}}(r)=\big(b(W_1), b(W_2), \ldots, b(W_{\min(N,|W|)}), b(A_1)\big).$$

Matching is by exact key equality. The social component is deliberately minimal: only the first author’s bigram is appended, with no co-authorship graph, affiliation signal, or score fusion (Turenne, 2015).

A second method, the Monogram Fingerprint (MGF), operates only on the title. After lower-casing and removing delimiters, the title is scanned left-to-right and only the first occurrence of each alphabet letter is retained, preserving the order in which characters first appear. A sorted variant, SMGF, alphabetically sorts the resulting sequence before comparison, but the reported effect is strongly negative for precision, indicating that order is a crucial part of the signature. The paper also evaluates an Anchor Bigram Fingerprint (BGF), which uses a dictionary of the top frequent bigrams in the corpus, excluding highly ambiguous ones such as “at”, “it”, and “is” (Turenne, 2015).

Fingerprint	Construction	Reported result
SSF	First bigram of the first 8 title words, plus first-author bigram	Precision $D$0, Recall $D$1
MGF	First-occurrence title characters, order preserved	Precision $D$2, Recall $D$3
BGF	Title bigrams selected from a corpus dictionary of frequent anchors	Precision $D$4, Recall $D$5
SMGF	MGF characters sorted alphabetically before comparison	Precision $D$6

The motivation for such fingerprints is heterogeneity rather than semantic ambiguity. The paper illustrates cross-database variation with examples such as CAI-1 versus CAl-1, mTORC1 versus mTORCl, and token-boundary disruption in “5'-Triphosphate.” The fingerprint therefore functions as a compact normalization device: it is short, ordered, and robust to many superficial variations while remaining discriminative enough for exact-key matching (Turenne, 2015).

3. Textual fingerprinting as bibliographic evidence

The stylometric line of work defines a textual fingerprint, or author’s fingerprint, as a vector of numerical linguistic attributes that are characteristically and repeatably distributed for a given writer. The working hypothesis is that some stylistic attributes remain comparatively stable even under genre constraints or conscious manipulation and can therefore be diagnostic of authorship. In bibliographic practice, such fingerprints complement metadata-driven fingerprints rather than replace them (0802.2234).

The case study is genre-controlled. It focuses on German-language fairy tales by Parkin, Bassewitz, and Leander: multiple texts by Parkin, four texts from Bassewitz’s Peterchen’s Mondfahrt, and six texts from Leander’s Träumereien an französischen Kaminen. In a broader preliminary comparison, roughly 10 authors per genre with 3–5 documents per author were considered across Fairy Tale, Common Speech, and Technical Language. Fairy tales were chosen because they are short, have straightforward plots, and use simple language. This genre control is methodologically central, because the authors explicitly note that genre can dominate stylistic signal; comparison is therefore conducted primarily within the same genre to avoid genre-induced similarity across different authors (0802.2234).

The representation is deliberately broad. It includes token-level counts such as number of words and number of distinct word types; function-word and pronoun frequencies such as ich and wir; POS-category counts such as adjectives and verbs; sentence and clause structure measures including parataxis versus hypotaxis; vocabulary-distribution measures such as Type-Token Ratio (TTR), Yule’s characteristic $D$7, and Hapax Legomena; entropy of the word distribution computed on a fixed-length sample; stopword statistics; and “frequency class” from the University of Leipzig Wortschatz resource. Several of these measures are formalized directly:

$D$8

$D$9

$F_k$0

$F_k$1

Here $F_k$2 is the number of distinct types, $F_k$3 the total tokens, $F_k$4 the hapax count, and $F_k$5 the number of types occurring exactly $F_k$6 times (0802.2234).

The study also emphasizes attribute independence. Pairwise scatter plots are used to identify dependencies, and a bivariate chi-square ordering ranks variables by discrepancy from the global distribution within each genre. Entropy, TTR, and average word frequency are reported as mutually dependent, whereas Number of Hapax Legomena and Yule’s $F_k$7 are comparatively independent. This filtering is intended to reduce redundant evidence when forming an authorial fingerprint (0802.2234).

4. Computational procedures and models

The duplicate-detection framework is algorithmically simple. For a test database such as PubMed and a target database such as Web of Science, each record is preprocessed, a fingerprint key is computed, and a hash table indexes target records by that key. A query record is treated as a duplicate when its fingerprint exactly matches a key already present in the index. Because the decision rule is exact equality rather than thresholded similarity, the method is fast and has a low memory footprint relative to similarity-based matching. The paper reports approximately 7 seconds for $F_k$8 and $F_k$9 on a 2.3 GHz laptop, with memory approximately five times the raw dataset size due to token lists and hash indices (Turenne, 2015).

For comparison, the same study evaluates heavier similarity-based baselines. Salton Vector Space (SVS) uses TF or TF-IDF weighting with cosine similarity,

$O(NM)$0

with a duplicate decision at approximately $O(NM)$1. Collocation Similarity-Based (CSB) compares 2- and 3-word collocations by Jaccard similarity,

$O(NM)$2

again with a threshold near $O(NM)$3. A Sorted Neighborhood strategy with simple keys is also described, but the main method remains hashed fingerprint equality (Turenne, 2015).

The stylometric study uses a different computational repertoire. Principal component analysis and bivariate statistics are used for exploration and attribute prioritization. Two “dynamic” methods are then described: demographic clustering with IBM Intelligent Miner V8, and a custom genetic algorithm, ClusGen, that adaptively searches for a “representative median vector” per author. Its conceptual decision rule is a nearest-median classifier:

$O(NM)$4

where $O(NM)$5 is a robust central tendency for author $O(NM)$6. The paper does not specify the distance $O(NM)$7 or the genetic hyperparameters, and it does not report wrapper-based feature selection or SVM-based weighting. Validation details such as cross-validation folds are likewise not given; the emphasis is on cluster separability and overall classification rate within and across genres (0802.2234).

Taken together, these procedures show two different operational meanings of “fingerprint.” In duplicate detection it is a deterministic blocking key designed for exact equality. In stylometric attribution it is a multi-attribute vector interpreted through clustering or adaptive classification. This suggests that bibliographic fingerprinting is better understood as a family of reduced evidentiary representations than as a single algorithmic template (Turenne, 2015, 0802.2234).

5. Empirical findings

The duplicate-detection study evaluates PubMed and Web of Science corpora focused on the TOR molecule. After filtering, the PubMed corpus contains 7,709 records, 3.87M words, years 1951–2010, 7,709 author fields, 7,689 title fields, and 7,709 source fields. The Web of Science subset contains 12,658 records, 2.78M words, years 1951–2010, and 12,658 author, title, and source fields. The hand-curated gold standard consists of 374 PubMed records, of which 202 are exact duplicates with Web of Science. On this set, SSF achieves precision $O(NM)$8 and recall $O(NM)$9 with $O(N \log M)$0, $O(N \log M)$1, and $O(N \log M)$2; MGF achieves precision $O(N \log M)$3 and recall $O(N \log M)$4 with $O(N \log M)$5, $O(N \log M)$6, and $O(N \log M)$7. Baselines include SVS with precision $O(N \log M)$8 and recall $O(N \log M)$9, CSB with precision $W=(W_1,\ldots,W_{|W|})$0 and recall $W=(W_1,\ldots,W_{|W|})$1, and BGF with precision $W=(W_1,\ldots,W_{|W|})$2 and recall $W=(W_1,\ldots,W_{|W|})$3 (Turenne, 2015).

The intrinsic evaluations reinforce the same pattern. A Monte Carlo random baseline yields approximately $W=(W_1,\ldots,W_{|W|})$4 precision and recall. On Half PubMed, where no duplicates are expected, the system produces one quasi-plagiarism false positive out of 3,855. On Full PubMed, where each record is a duplicate of itself, precision and recall are both $W=(W_1,\ldots,W_{|W|})$5. The principal trade-off is therefore explicit: exact key matching produces no false positives in the reported gold standard but can miss some near-duplicates with slightly altered structure (Turenne, 2015).

The stylometric study reports overall classification rates above $W=(W_1,\ldots,W_{|W|})$6 over the whole test corpus. Within genres, Fairy Tale is reported at around $W=(W_1,\ldots,W_{|W|})$7, while Common Speech and Technical Language are nearly $W=(W_1,\ldots,W_{|W|})$8. Clustering yields genre-pure clusters. For Fairy Tale, Number of Hapax Legomena, number of adjectives, and Yule’s $W=(W_1,\ldots,W_{|W|})$9 show marked distributional behavior, whereas verbs and frequency class are low-distribution attributes in that cluster (0802.2234).

At the author level, the paper identifies distinct profiles for Parkin and Bassewitz but not a single compact cluster for Leander. A Parkin-only cluster exhibits low adjective counts, high parataxis proportion, unusually frequent wir, high TTR combined with low numbers of distinct words and low hapax counts, shorter sentences, repetition tendencies, and a preference for nouns over adjectives. The paper itself cautions that high TTR with low $A_1$0 suggests short $A_1$1 and that this combination reflects the instability of TTR for short texts. Bassewitz’s texts are marked by higher frequency-class values, disproportionately high verb usage, coherent behavior of “longest word class,” “most frequent word class,” “most frequent stopword class,” and hapax-related measures within the same cluster, and an overall preference for less common vocabulary. Leander’s texts are heterogeneous and do not form a single compact stylistic cluster, indicating greater intra-author variation in the sampled material (0802.2234).

6. Limitations, controversies, and bibliographic integration

Both forms of fingerprinting have strong boundary conditions. For metadata-based duplicate detection, exact key equality means that near-duplicates with title changes beyond the first eight words may fail SSF matching. The method relies on title availability and, for SSF, on a usable first-author string. The paper also notes multilingual and transliteration issues, the absence of diacritic normalization, and tokenization edge cases involving hyphenated chemical names or gene identifiers. SMGF’s poor precision shows that aggressive compaction without order preservation can destroy discriminative power (Turenne, 2015).

For textual fingerprinting, the principal limitation is confounding by genre and length. The authors explicitly state that genre can overwhelm author differences, which is why robustness claims are restricted mainly to within-genre comparisons, especially Fairy Tale. Many attributes are length-sensitive: TTR and entropy correlate and are sensitive to text length, while clause-structure estimation can be noisy without robust parsing of German subordinate clauses. The corpus is small, German-only, and genre-narrow; per-author confusion matrices, significance tests, and quantified overfitting risk for the genetic algorithm are not provided. The paper therefore presents a practical blueprint rather than a fully standardized benchmark (0802.2234).

The bibliographic significance lies in complementarity. Metadata fingerprints are efficient for large-scale blocking, exact duplicate detection, and heterogeneous data integration. Textual fingerprints are useful when paratext is missing, unreliable, or disputed, as in author disambiguation, edition identification, provenance checking, plagiarism or ghostwriting detection, and catalog enrichment by stylistic clustering. Combined approaches are described as strongest: metadata narrows candidates, while stylometry provides evidentiary support (Turenne, 2015, 0802.2234).

Future extensions are stated conservatively. The duplicate-detection work suggests practical fallback strategies such as using MGF when author data are missing and using alternative keys such as ARDF for weak matching when the title is absent. The stylometric work proposes scaling to larger corpora and more authors, adding “subjective linguistic attributes” such as phraseologies like I believe that, testing generalization across genres, and evaluating scoring engines more rigorously. A plausible implication is that bibliographic fingerprinting becomes more reliable as a layered workflow: short exact keys for efficient candidate generation, followed by linguistically richer evidence when bibliographic identity or attribution remains uncertain (Turenne, 2015, 0802.2234).