What’s in a Domain? Multi-Domain Learning for Multi-Attribute Data

Proceedings of NAACL-HLT 2013, pages 685–690, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics What’s in a Domain? Multi-Domain Learning for Multi-Attribute Data Mahesh Joshi⋆ Mark Dredze† William W. Cohen⋆ Carolyn P. Ros´e⋆ ⋆ School of Computer Science, Carnegie Mellon University Pittsburgh, PA, 15213, USA † Human Language Technology Center of Excellence, Johns Hopkins University Baltimore, MD, 21211, USA maheshj@cs.cmu.edu,mdredze@cs.jhu.edu wcohen@cs.cmu.edu,cprose@cs.cmu.edu Abstract Multi-Domain learning assumes that a sin- gle metadata attribute is used in order to di- vide the data into so-called domains. How- ever, real-world datasets often have multi- ple metadata attributes that can divide the data into domains. It is not always apparent which single attribute will lead to the best do- mains, and more than one attribute might im- pact classification. We propose extensions to two multi-domain learning techniques for our multi-attribute setting, enabling them to si- multaneously learn from several metadata at- tributes. Experimentally, they outperform the multi-domain learning baseline, even when it selects the single “best” attribute. 1 Introduction Multi-Domain Learning (Evgeniou and Pontil, 2004; Daum´e III, 2007; Dredze and Crammer, 2008; Finkel and Manning, 2009; Zhang and Yeung, 2010; Saha et al., 2011) algorithms learn when training in- stances are spread across many domains, which im- pact model parameters. These algorithms use exam- ples from each domain to learn a general model that is also sensitive to individual domain differences. However, many data sets include a host of meta- data attributes, many of which can potentially define the domains to use. Consider the case of restaurant reviews, which can be categorized into domains cor- responding to the cuisine, location, price range, or several other factors. For multi-domain learning, we should use the metadata attribute most likely to char- acterize a domain: a change in vocabulary (i.e. fea- tures) that most impacts the classification decision (Ben-David et al., 2009). This choice is not easy. First, we may not know which metadata attribute is most likely to fit this role. Perhaps the location most impacts the review language, but it could easily be the price of the meal. Second, multiple metadata attributes could impact the classification decision, and picking a single one might reduce classification accuracy. Therefore, we seek multi-domain learn- ing algorithms which can simultaneously learn from many types of domains (metadata attributes). We introduce the multi-attribute multi-domain (MAMD) learning problem, in which each learning instance is associated with multiple metadata at- tributes, each of which may impact feature behavior. We present extensions to two popular multi-domain learning algorithms, FEDA (Daum´e III, 2007) and MDR (Dredze et al., 2009). Rather than selecting a single domain division, our algorithms consider all attributes as possible distinctions and discover changes in features across attributes. We evaluate our algorithms using two different data sets – a data set of restaurant reviews (Chahuneau et al., 2012), and a dataset of transcribed speech segments from floor debates in the United States Congress (Thomas et al., 2006). We demonstrate that multi-attribute al- gorithms improve over their multi-domain counter- parts, which …

Content truncated to 500 words