Towards Domain Adaptation for Parsing Web Data

Proceedings of Recent Advances in Natural Language Processing, pages 357–364, Hissar, Bulgaria, 7-13 September 2013. Towards Domain Adaptation for Parsing Web Data Mohammad Khan Indiana University Bloomington, IN USA khanms@indiana.edu Markus Dickinson Indiana University Bloomington, IN USA md7@indiana.edu Sandra K¨ubler Indiana University Bloomington, IN USA skuebler@indiana.edu Abstract We improve upon a previous line of work for parsing web data, by exploring the im- pact of different decisions regarding the training data. First, we compare train- ing on automatically POS-tagged data vs. gold POS data. Secondly, we compare the effect of training and testing within sub-genres, i.e., whether a close match of the genre is more important than training set size. Finally, we examine different ways to select out-of-domain parsed data to add to training, attempting to match the in-domain data in different shallow ways (sentence length, perplexity). In general, we find that approximating the in-domain data has a positive impact on parsing. 1 Introduction and Motivation Parsing data from the web is notoriously diffi- cult, as parsers are generally trained on news data (Petrov and McDonald, 2012). The problem, how- ever, varies greatly depending upon the particular piece of web data: what is often termed web data is generally a combination of different sub-genres, such as Facebook posts, Twitter feeds, YouTube comments, discussion forums, blogs, etc. The language used in such data does not follow stan- dard conventions in various respects (see Herring, 2011): 1) The data is edited to varying degrees, with Twitter on the lower end and professional emails and blog on the upper end of the scale. 2) The sub-genres often display characteristics of spoken language, including sentence fragments and colloquialisms. 3) Some web data, especially social media data, typically contains a high num- ber of emoticons and acronyms such as LOL. At the same time, there is a clear need to de- velop basic NLP technology for a variety of types of web data. To perform tasks such as sentiment analysis (Nakagawa et al., 2010) or information extraction (McClosky et al., 2011), it helps to part- of-speech (POS) tag and parse the data, as a step towards providing a shallow semantic analysis. We continue our work (Khan et al., 2013) on dependency parsing web data from the English Web Treebank (Bies et al., 2012). We previously showed that text normalization has a beneficial ef- fect on the quality of a parser on web data, that we can further improve the parser’s accuracy by a simple, n-gram-based parse revision method, and that having a balanced training set of out-of- domain and in-domain data provides the best re- sults when parsing web data. The current work extends this previous work by more closely exam- ining the data given as input for training the parser. Specifically, we take the following directions: 1. All previous experiments were carried out on gold part of speech (POS) tags. Here, we in- vestigate using a POS tagger trained on out- of-domain data, thus providing a more realis- tic setting for parsing web data. …

Content truncated to 500 words