Creating POS Tagging and Dependency Parsing Experts via Topic Modeling

Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 347–355, Valencia, Spain, April 3-7, 2017. c⃝2017 Association for Computational Linguistics Creating POS Tagging and Dependency Parsing Experts via Topic Modeling Atreyee Mukherjee Indiana University atremukh@indiana.edu Sandra K¨ubler Indiana University skuebler@indiana.edu Matthias Scheutz Tufts University matthias.scheutz@tufts.edu Abstract Part of speech (POS) taggers and depen- dency parsers tend to work well on ho- mogeneous datasets but their performance suffers on datasets containing data from different genres. In our current work, we investigate how to create POS tagging and dependency parsing experts for het- erogeneous data by employing topic mod- eling. We create topic models (using Latent Dirichlet Allocation) to determine genres from a heterogeneous dataset and then train an expert for each of the gen- res. Our results show that the topic mod- eling experts reach substantial improve- ments when compared to the general ver- sions. For dependency parsing, the im- provement reaches 2 percent points over the full training baseline when we use two topics. 1 Introduction POS tagging and dependency parsing perform well when trained and tested on datasets that are predominantly in the same text domain. However, there is decrease in accuracy for heterogeneous datasets, i.e., for datasets that consist of a mix- ture of data from different domains. Our current work focuses on improving POS tagging and de- pendency parsing for such heterogeneous datasets from a variety of different genres by creating ex- perts for automatically detected topics. In our case, the datasets consist of newspaper reports on the one hand and biomedical extracts on the other. For determining the topic of a sentence, we use Latent Dirichlet Allocation (LDA), which finds the latent topic structure in a document. In our case, a document to be clustered consists of a sin- gle sentence. We then assign each sentence to the most likely topic, for both training and test sentences. We consequently train an expert for each topic and then use this expert to POS tag and parse the test sentences belonging to this topic. We assume that the topics detected by the topic modeler do not only pertain to lexical differences, which can be beneficial for the POS tagger and the parser, but also to syntactic phenomena. Thus, one topic may focus on “incomplete” sentences, such as headlines in a newspaper. Our work is related to domain adaptation since the aim is to improve (morpho-)syntactic analy- sis for different domains. However, our approach can be regarded as a more general approach to the problem of domains as it is based on automati- cally determining the genres present in the dataset. Thus, no manual work is involved. Our results show small to considerable im- provements over a competitive baseline of using the full training set. For POS tagging, there is an improvement of 0.3 percent points over the full training set. For dependency parsing, the gain is more pronounced: almost 2% over the full train- ing set. The …

Content truncated to 500 words