Can Domain Adaptation be Handled as Analogies?

Can Domain Adaptation be Handled as Analogies? Núria Bel, Joel Pocostales Universitat Pompeu Fabra Roc Boronat, 138, 08018 Barcelona, Spain nuria.bel@upf.edu; joel.pocostales@upf.edu Abstract Aspect identification in user generated texts by supervised text classification might suffer degradation in performance when changing to other domains than the one used for training. For referring to aspects such as quality, price or customer services the vocabulary might differ and affect performance. In this paper, we present an experiment to validate a method to handle domain shifts when there is no available labeled data to retrain. The system is based on the offset method as used for solving word analogy problems in vector semantic models such as word embedding. Despite of the fact that the offset method indeed found relevant analogues in the new domain for the classifier initial selected features, the classifiers did not deliver the expected results. The analysis showed that a number of words were found as analogues for many different initial features. This phenomenon was already described in the literature as 'default words' or 'hubs'. However, our data showed that it cannot be explained in terms of word frequency or distance to the question word, as suggested. Keywords: Aspect identification, domain adaptation, document classification 1. Introduction Machine Learning in general, and classifiers in particular, might suffer degradation in performance when the data to handle belongs to a different domain than the data used for training. Domain adaptation addresses the problem of moving from a source distribution for which we have labelled training data to a target distribution for which we have no labels. Domain adaptation might be crucial for identifying aspects in Aspect Based Sentiment Analysis (ABSA). Note that while for polarity identification, it is likely that a common vocabulary is shared among different domains (e.g. good, bad), for identifying aspects such as quality, design, or support, different domains might exhibit different vocabulary. For instance, quality for a laptop would be described in terms of ‘performance’: fast, powerful, etc., while for a restaurant would be described in terms of ‘taste’: delicious, tasty, etc. In this paper, we present an experiment designed to validate a method to handle domain shifts when there is no available labelled data in a new target domain for an aspect classifier to be retrained. Our approach was to leverage the use of vector space models for semantics such as the one provided by word embeddings (WE, Mikolov et al. 2013). We experimented with using the offset method, as used for solving word analogy problems, to tackle domain adaptation. In a WE model, we found the examples like in (1), where responses provided by the offset method are in bold. (1) laptop : shop :: bread : bakery laptop : shop :: beer : brewery laptop : shop :: medicine: pharmacy Thus, we produced lists of analogue words to support domain adaptation of a system for classifying sentences. The task was classifying user-generated texts as according to the aspect (or attribute) of the product (or entity) the user …

Content truncated to 500 words