Segmenting and Tagging Text with Neural Networks

  • Datum: 09 juni, kl. 10.15
  • Plats: Humanistiska teatern, Thunbergsvägen 3, Uppsala
  • Doktorand: Shao, Yan
  • Om avhandlingen
  • Arrangör: Institutionen för lingvistik och filologi
  • Kontaktperson: Shao, Yan
  • Disputation

Disputation

Segmentation and tagging of text are important preprocessing steps for higher-level natural language processing tasks. In this thesis, we apply a sequence labelling framework based on neural networks to various segmentation and tagging tasks, including sentence segmentation, word segmentation, morpheme segmentation, joint word segmentation and part-of-speech tagging, and named entity transliteration. We apply a general neural CRF model to different tasks by designing specific tag sets. In addition, we explore effective ways of representing input characters, such as utilising concatenated n-grams and sub-character features, and use ensemble decoding to mitigate the effects of random parameter initialisation.

The segmentation and tagging models are evaluated in a truly multilingual setup with more than 70 datasets. The experimental results indicate that the proposed neural CRF model is effective for segmentation and tagging in general as state-of-the-art accuracies are achieved on datasets in different languages, genres, and annotation schemes for various tasks. For word segmentation, we propose several typological factors to statistically characterise the difficulties posed by different languages and writing systems. Based on this analysis, we apply language-specific settings to the segmentation system for higher accuracy. Our system achieves substantially better results on languages that are more difficult to segment when compared to previous work. Moreover, we investigate conventionally adopted evaluation metrics for segmentation tasks. We propose that precision should be excluded and using recall alone is more adequate for sentence segmentation and word segmentation. The segmentation and tagging tools implemented along with this thesis are publicly available as experimental frameworks for future development as well as preprocessing tools for higher-level NLP tasks.