Domain Adaptation for Dependency Parsing via Self-Training

This paper presents a successful approach for domain adaptation of a dependency parser via self-training. We improve parsing accuracy for out-of-domain texts with a self-training approach that uses confidence-based methods to select additional training samples. We compare two confidence-based methods: The first method uses the parse score of the employed parser to measure the confidence into a parse tree. The second method calculates the score differences between the best tree and alternative trees. With these methods, we were able to improve the labeled accuracy score by 1.6 percentage points on texts from a chemical domain and by 0.6 on average on texts of three web domains. Our improvements on the chemical texts of 1.5% UAS is substantially higher than improvements reported in previous work of 0.5% UAS. For the three web domains, no positive results for self-training have been reported before.