Abstract
Dongxu Zhang, Boliang Zhang, Xiaoman Pan, Xiaocheng Feng, Heng Ji, Weiran Xu. Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers.
2016.
Introduction
Building robust Named Entity Recognition (NER) systems typically requires extensive human-annotated datasets. However, for many "incident languages" (IL)—those spoken in crisis-affected regions—such data is often unavailable. A common solution is annotation projection, which involves transferring entity labels from a high-resource language like English to a target IL using parallel corpora (bitexts).
Traditional projection methods rely heavily on word-alignment tools such as GIZA++. While useful, these tools are often noisy; a single alignment error can lead to lost or misplaced labels, which then propagate through the training pipeline. This paper introduces a weakly-supervised framework that moves beyond simple mapping, treating projection as a learning task through a neural "bitext name tagger" (BNT).
Methodology: The Bitext Name Tagger
The authors propose a two-step framework designed to improve the quality of automatically generated training data:
1. High-Confidence Seed Generation
Rather than projecting all labels immediately, the system first identifies "sure bets." Using string similarity (Levenshtein distance), phonetic similarity (Soundex), and translation tables, the framework identifies IL entities that clearly correspond to English entities. These high-confidence pairs serve as the initial training signal.
2. Neural Architecture
The core of the framework is the Bitext Name Tagger, a deep learning model based on a Bi-directional Long Short-Term Memory (Bi-LSTM) network coupled with a Conditional Random Field (CRF) layer.
As shown in the architecture diagram, the model processes English and IL sentences simultaneously. A critical innovation is the "bridge" between the two: the hidden states of the English LSTM are used to initialize the IL LSTM. This allows the model to transfer the semantic context of an entity from the English side to the IL side, enabling it to "find back" labels that traditional word aligners might miss due to structural differences or alignment errors.
Experiments and Results
The framework was evaluated on Turkish and Uzbek using data from the DARPA LORELEI program. The researchers compared the BNT against traditional word-alignment (WA) baselines and other state-of-the-art methods.
Key Performance Gains
The BNT demonstrated superior performance in both projection quality and downstream NER accuracy. In Turkish, the BNT achieved a projection F1-score of 57.2, a massive improvement over the 33.9 achieved by pure word alignment.
As illustrated in the results for Turkish, the BNT (Annotation with bitext name tagger) consistently outperforms both GIZA++ and high-confidence seed annotation alone across various training sizes. Notably, the framework is highly efficient, reaching a ~50% F-score with as few as 2,000 parallel sentences. This efficiency stems from the constrained nature of the NER task compared to full-text machine translation.
Discussion and Future Work
The core insight of this research is that projection should be a probabilistic learning task rather than a deterministic lookup. By using a neural network, the system learns the patterns of entity translation, making it robust to the failures of traditional mapping.
While the current system performs well for languages sharing similar scripts (like Turkish and English), the authors note that performance may vary for languages with different writing systems. Future work aims to integrate transliteration techniques and external gazetteers to broaden the framework's applicability to a wider range of global languages.
Summary updated