Abstract
Chinese name disambiguation is a critical task in natural language processing due to the prevalence of polysemous names. This paper introduces an adaptive clustering framework that utilizes 12-dimensional attribute features extracted via Conditional Random Fields (CRF) and weighted by Maximum Entropy (ME) models. The core of the approach is an adaptive fuzzy clustering algorithm that employs the $F$-statistic to dynamically determine optimal similarity thresholds. Evaluated on the CLP2012 dataset, the system demonstrates robust performance in entity linking and identifying new entities, while error analysis highlights challenges in initial named entity recognition.
The Challenge of Chinese Name Disambiguation
In Chinese natural language processing, name repetition presents a significant hurdle. A single name string, such as "Xue Bai," can refer to a singer, a soldier, or even a common phrase depending on the context. The task of Chinese Name Disambiguation involves accurately determining which specific individual a name refers to within a given text.
The authors target the CLP2012 (CIPS-SIGHAN 2012) evaluation task, which requires three distinct actions:
- Entity Linking: Mapping a name to an existing entry in a Knowledge Base (KB).
- Clustering "Out" Entities: Grouping names that refer to the same person but are not present in the KB.
- Filtering "Other": Identifying when a name string is actually a common word rather than a person.
Methodology: Attribute Extraction and Adaptive Clustering
The researchers developed a multi-stage pipeline to address these challenges:
1. Attribute Recognition and Weighting
The system defines 12 key attributes (e.g., organization, career, position, education). Conditional Random Fields (CRFs) are used to extract these attributes from both raw text and KB files. Recognizing that not all attributes are equally discriminative, a Maximum Entropy (ME) model learns optimal weights for these 12 dimensions based on training data.
2. Semantic Similarity
To move beyond simple keyword matching, the system utilizes the Tongyici Cilin (a Chinese synonym dictionary). This allows the model to calculate semantic similarity between attributes, understanding, for instance, that "vocalist" and "singer" are semantically related.
3. Adaptive Fuzzy Clustering
The core innovation is the adaptive clustering method. Traditional fixed-threshold clustering often fails because different entities have different distributions in the data. The authors implement an iterative process using an $F$-statistic—a measure of the ratio between inter-cluster and intra-cluster distances—to find the mathematically optimal threshold ($\lambda$) for each specific name set.
Experimental Results
The system was evaluated using the CLP2012 dataset, which includes 16 training name sets and 32 test name sets. Performance was measured using Precision (P), Recall (R), and F-measure (F).
| Dataset | Precision | Recall | F-value |
|---|---|---|---|
| Training Data | 0.9256 | 0.9032 | 0.9143 |
| Test Data | 0.7424 | 0.7428 | 0.7426 |
While performance naturally decreased on the test data, the F-value of 0.7426 represents a competitive result for this complex task. The system's ability to "self-tune" to specific data distributions through the $F$-statistic is a key driver of its robustness.
Error Analysis and Future Directions
A detailed error analysis reveals that the primary bottleneck lies in the initial Named Entity Recognition (NER) phase. Over half of the errors occurred because names were incorrectly identified as common words.
Future work will focus on:
- Feature Selection: Incorporating non-attribute contextual features beyond the 12 defined dimensions.
- Clustering Optimization: Refining methods to better distinguish between "Out" (new entities) and "Other" (common words) categories.
- NER Improvement: Enhancing the initial recognition stage to reduce the misidentification of names as common vocabulary.
Summary updated