Domain Adaptation for Parsing

Proceedings of Recent Advances in Natural Language Processing, pages 56–64, Hissar, Bulgaria, 7-13 September 2013. Domain Adaptation for Parsing Eric Baucom Indiana University Bloomington, IN, USA eabaucom@indiana.edu Levi King Indiana University Bloomington, IN, USA leviking@indiana.edu Sandra K¨ubler Indiana University Bloomington, IN, USA skuebler@indiana.edu Abstract We compare two different methods in domain adaptation applied to constituent parsing: parser combination and co- training, each used to transfer information from the source domain of news to the tar- get domain of natural dialogs, in a set- ting without annotated data. Both methods outperform the baselines and reach similar results. Parser combination profits most from the large amounts of training data combined with a robust probability model. Co-training, in contrast, relies on a small set of higher quality data. 1 Introduction Research on parsing has mostly concentrated on parsing the Penn Treebank (Marcus et al., 1993). As a consequence, most parsers have probability models that are optimized for the syntactic an- notations in this treebank and more generally for the language in the treebank. This means that a parser trained on the Penn Treebank will show a severe degradation in performance when used for parsing data from another domain (McClosky et al., 2010). More recently, research has started on adapting parsers to new domains so that the degra- dation in parsing is minimized. One of the first venues at which domain adaptation was targeted was the 2007 CoNLL shared task on dependency parsing (Nivre et al., 2007). One of the challenges in domain adaptation for parsing is the lack of annotated data in the target domain. Research has covered a range of differ- ent approaches, all geared towards providing auto- matically labeled data in the target domain to add as training data. Approaches include ensembles of parsers, self-training, and methods for selecting high quality sentences to reduce the noise (see sec- tion 2 for details). The most promising approach at present is an approach by McClosky et al. (2010), which automatically selects a domain that is the most similar to the target domain. In our current work, we investigate domain adaptation for constituent parsing, in a setting where no labeled data in the target domain is avail- able. More specifically, we compare two differ- ent approaches: One approach is based on an en- semble of parsers, the other one uses co-training with two different parsers. Both approaches reach moderate improvements over the baseline, and we are interested in seeing the advantages and disad- vantages of those two promising methods. The source domain for our experiments is the Penn Treebank; the target domain consists of sponta- neous dialogs based on cooperative tasks involv- ing navigation on a map or in a search envi- ronment. For the unlabeled target domain data, we use the Edinburgh Map Task (HCRC) corpus (Thompson et al., 1996), and the Indiana Cooper- ative Remote Search Task (CReST) corpus (Eber- hard et al., 2010) as test set. The remainder of the paper is structured as fol- lows: In section 2, we discuss …

Content truncated to 500 words