Bridging the Language Gap in Scholarly Data I: Enhancing Author Disambiguation Algorithms for Chinese Names

1. Problem Statement

Author name disambiguation (AND) is critical for accurate scientometric analysis, yet it faces unique and severe challenges in the context of Chinese names, particularly when transitioning between domestic and international databases.

• High Ambiguity in Pinyin: In international records, Chinese names are transliterated into Pinyin (e.g., "Wang Wei"). This process is highly ambiguous because a single Pinyin string can map to multiple distinct Chinese characters (e.g., 王伟，王威，王维), and different characters can share the same pronunciation.
• Structural Limitations: Chinese naming conventions exacerbate this: family names are highly concentrated (top 100 surnames cover ~87% of the population), full names lack middle initials, and the family-name-first structure conflicts with Western parsing assumptions.
• Gap in Existing Methods: Current disambiguation algorithms are predominantly optimized for Latin-script metadata. They often rely on supervised learning (requiring labeled data) or simple heuristics (co-authorship, affiliation) that fail to capture semantic similarity. Crucially, few methods address the specific challenges of Pinyin transliteration or compare performance across both Chinese character and Pinyin representations of the same corpus.

2. Methodology

The authors propose a rule-based, script-agnostic disambiguation framework that integrates social, institutional, and semantic signals. The methodology consists of three main components:

A. Dataset Construction

• Source: 65,241 physics papers from the China National Knowledge Infrastructure (CNKI), spanning 1953–2024.
• Parallel Corpora: The authors created two versions of the dataset:
  1. Native Chinese: Original Chinese characters for names and content.
  2. Translated English: Names converted to Pinyin (using a specialized dictionary to handle polyphonic characters) and content (titles, abstracts, keywords) translated to English.
• Ground Truth: A manually annotated dataset of 80 name pairs (40 "same identity," 40 "different identity") verified by native Chinese speakers to serve as the evaluation benchmark.

B. The Disambiguation Algorithm

The algorithm operates on distinct names associated with at least two publications. It evaluates pairs of papers against four sequential criteria. If any criterion is met, the authors are merged into a single identity:

1. Co-authorship: Do the authors share at least one co-author? (Strongest signal).
2. Affiliation Similarity: Do they share an institutional affiliation? (Measured via normalized Levenshtein similarity > 0.6 or substring containment).
3. Content Similarity: Do the papers share semantic similarity? (Computed using Word2Vec document embeddings from titles, abstracts, and keywords).
   • Thresholding: Separate thresholds were optimized for each language to ensure cross-corpus consistency (Chinese: 0.9207; English: 0.9621).
4. Citation Links: Has one author cited the other's work? (Indicating self-citation).

This approach extends existing rule-based methods (Sinatra et al.) by adding the content similarity component (inspired by Waqas & Qadir) to capture thematic links where social or institutional ties are absent.

C. Baselines

The proposed method was benchmarked against two established approaches:

• Sinatra et al.: Uses co-authorship, affiliation, and citations only.
• Waqas & Qadir: Uses co-authorship, affiliation, and metadata similarity (titles, abstracts, keywords) but excludes email/venue data not present in CNKI.

3. Key Contributions

1. Hybrid Rule-Based Framework: The integration of content similarity into a rule-based system significantly improves recall without sacrificing precision, addressing the limitation of methods that rely solely on social/institutional ties.
2. Script-Agnostic Evaluation: The study provides the first systematic comparison of disambiguation accuracy across native Chinese characters and Pinyin transliterations using the same underlying corpus.
3. Open Science: The authors released the dataset (65k papers with dual-language metadata) and all code/analysis scripts to ensure reproducibility.
4. Addressing the "Pinyin Gap": The research demonstrates that high-quality disambiguation is possible even with the high ambiguity introduced by Romanization, provided the algorithm leverages multiple signal types.

4. Results

• Performance Metrics:
  • F1-Score: The proposed method achieved 0.89 for Chinese character names and 0.88 for Pinyin names.
  • Comparison: It outperformed both baselines (Sinatra et al. and Waqas & Qadir) in both languages.
  • Precision vs. Recall: All methods achieved near-perfect precision (1.0 for Chinese, ~0.95 for English), meaning they rarely merged distinct authors. The proposed method's superiority came from higher recall (0.80–0.82), successfully identifying same-identity pairs that baselines missed.
• Cross-Language Consistency:
  • Jaccard Similarity: The algorithm showed high stability across representations, with a mean Jaccard similarity of 0.922 between the Chinese and Pinyin identity groupings.
  • Agreement: 87.3% of identities were assigned identical paper sets in both language representations, indicating the framework is robust to transliteration.
• Data Flow: The algorithm resolved 106,411 identities in the Chinese dataset and 99,222 in the Pinyin dataset. Co-authorship was the primary resolver (15,583 matches in Chinese), followed by affiliation, content similarity, and citations.

5. Significance and Implications

• Scientometric Reliability: The study highlights that name ambiguity leads to two critical errors: conflation (merging distinct people, inflating citation counts) and fragmentation (splitting one person's work, obscuring their impact). The proposed method significantly reduces fragmentation, which is crucial for fair research evaluation.
• Global Research Equity: By improving disambiguation for Chinese names, the work helps correct biases in global bibliometric systems where Chinese researchers are often under-counted or misattributed due to transliteration issues.
• Scalability: The rule-based, unsupervised nature of the algorithm allows it to be applied to large-scale datasets without the need for expensive labeled training data, making it suitable for other non-Western naming systems (e.g., Korean, Japanese).
• Future Directions: The authors suggest that while algorithmic improvements are vital, long-term solutions should include better metadata standards (dual representation of names) and wider adoption of persistent identifiers (ORCID).

In conclusion, this paper presents a robust, reproducible solution to a critical bottleneck in scientometrics, demonstrating that combining social network analysis with semantic content similarity can effectively bridge the language gap in author disambiguation.