Automatic Domain Partitioning for Multi-Domain Learning

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 869–873, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Automatic Domain Partitioning for Multi-Domain Learning Di Wang School of Computer Science Carnegie Mellon University Pittsburgh, PA 15213, USA diwang@cs.cmu.edu Chenyan Xiong School of Computer Science Carnegie Mellon University Pittsburgh, PA 15213, USA cx@cs.cmu.edu William Yang Wang School of Computer Science Carnegie Mellon University Pittsburgh, PA 15213, USA ww@cmu.edu Abstract Multi-Domain learning (MDL) assumes that the domain labels in the dataset are known. However, when there are multiple metadata at- tributes available, it is not always straightfor- ward to select a single best attribute for do- main partition, and it is possible that combin- ing more than one metadata attributes (includ- ing continuous attributes) can lead to better MDL performance. In this work, we propose an automatic domain partitioning approach that aims at providing better domain identi- ties for MDL. We use a supervised clustering approach that learns the domain distance be- tween data instances , and then cluster the data into better domains for MDL. Our experiment on real multi-domain datasets shows that us- ing our automatically generated domain parti- tion improves over popular MDL methods. 1 Introduction Instead of assuming data are i.i.d, Multi-domain learning (MDL) methods assumes that data come from several domains and make use of domain la- bels to improve modeling performance (Daum´e III, 2007). The motivation of using MDL is that datasets from different domains could be different, in two ways. First, the feature distribution p(x) could be domain specific, meaning that the importance of each feature is different across domains. Second, the distribution of label Y given X, p(y|x), of dif- ferent domains could be different. These differ- ences could create problems for traditional machine learning methods: models learned from one domain might not be generalizable to other domains (Ben- David et al., 2006; Ben-David et al., 2010). One common assumption of MDL methods is that the domain identities are pre-defined. For example, in the multi-domain Amazon product review dataset (Finkel and Manning, 2009), the product categories are typically used as the domain identities. How- ever, a question raised by Joshi et al. (2012) is that, in real-world data sets, there could be many ways to split data into domains, and it is hard to decide which one to use. Consider the Amazon product reviews, where we have multiple attributes attached to each review: for example, product category, reviewer lo- cation, price, and number of feedback. Which at- tribute is the most informative domain label? Or we should use all of these meta-data and partition the data into many small domains? In this paper, we investigate the problem of au- tomatic domain partitioning. We propose an em- pirical domain difference testing method to exam- ine whether two groups of data are i.i.d, or gener- ated from different distributions, and how different they are. Using this approach, we generate data pairs that belong to the same distribution, and data …

Content truncated to 500 words