Abstract
The research question treated in this paper is centered on the idea of exploiting rich resources of one language to enhance the performance of a mention detection system of another one. We successfully achieve this goal by projecting information from one language to another via a parallel corpus. We examine the potential improvement using various degrees of linguistic information in a statistical framework and we show that the proposed technique is effective even when the target language model has access to a significantly rich feature set. Experimental results show up to 2.4F improvement in performance when the system has access to information obtained by projecting mentions from a resource-rich-language mention detection system via a parallel corpus.
Introduction
Mention Detection (MD) is a fundamental task in natural language processing that involves identifying and classifying references to entities—including proper names, nominal phrases, and pronouns. While English benefits from extensive annotated resources, languages like Arabic present significant challenges due to their complex morphology and relative scarcity of labeled data. This paper, originally presented at EMNLP 2010, proposes a method to "bootstrap" linguistic intelligence by projecting entity tags from English to Arabic across a parallel corpus.
Methodology
The researchers developed a three-step pipeline to transfer knowledge from a high-performing English MD system (82.7 F-measure) to an Arabic target:
- Annotation: Running the English MD system on the English side of a parallel corpus.
- Projection: Transferring tags to Arabic words using word-alignment data, with logic to handle complex mappings and filter noise.
- Feature Extraction: Integrating projected information as features into a Maximum Entropy Markov Model (MEMM).
Key features included gazetteers, model-based outputs, and syntactic features. The "Parser-based" features were particularly noteworthy, as they leveraged the labels of parent non-terminals in a parse tree to provide a strong cross-lingual signal.
Experimental Setup
The system was evaluated on the Arabic ACE 2007 dataset. The researchers utilized two types of parallel corpora: a 1-million-word hand-aligned corpus and a 22-million-word automatically aligned corpus. By treating projected tags as features rather than ground-truth labels, the model learned to weigh the reliability of the "borrowed" information against local linguistic cues.
Results and Discussion
The results demonstrated that cross-lingual projection provides significant gains, even when the Arabic baseline is already "rich" with morphological and syntactic information.
As shown in the results, the combined model ("Comb.") consistently outperformed the baseline across all feature sets (Lexical, Stem, and Syntactic). The maximum gain observed was a 2.4 F-measure point improvement for resource-poor baselines. Even in a "high-resource" scenario where the Arabic system already utilized stems and POS tags, the projection method provided a 1.7-point boost. This suggests that the information captured through cross-lingual alignment is complementary to local morphological analysis.
Conclusion
This research proves that "borrowing" knowledge from resource-rich languages is an effective strategy for enhancing performance in resource-poor or morphologically complex languages. The success of using projected data as features rather than direct labels highlights a robust path for cross-lingual transfer learning. Future work may involve scaling the parallel data further and refining alignment techniques to reduce noise, particularly for challenging nominal mentions.
Summary updated