Joint Modeling for Dependency Parsing

Опубликовано 7 ноября 2016, 20:57

Abstract: Parsing accuracy is greatly impacted by the quality of preprocessing steps such as part-of-speech (POS) tagging, word segmentation and morphological analysis. While prior researches have successfully demonstrated that joint modeling alleviates error propagation in pipeline architectures, their methods typically complicate the inference task and need to impose constraints on scoring functions to keep inference tractable. In this talk, I will present my work in joint modeling to address these challenges. In particular, I will first describe how to employ the randomized greedy algorithm to solve inference for joint segmentation, POS tagging and dependency parsing. The proposed method requires no constraints on the scoring function. Therefore it is able to handle arbitrary high-order and global features interleaving words, tags and parse trees. I will then introduce a neural network stacking method for joint dependency parsing and POS tagging which we call "stack-propagation". The underlying technique draws on recent advances in neural network parsing that allow us to address the error propagation issue while the inference complexity remaining linear. Experimental results show that both our randomized greedy and stack-propagation methods consistently outperform strong baselines across a wide range of datasets and languages.

See more on this video at microsoft.com/en-us/research/v...