Domain Adaptation for Sentiment Analysis using Keywords in the Target Domain as the Learning Weight

Domain Adaptation for Sentiment Analysis using Keywords in the Target Domain as the Learning Weight Jing Bai Hiroyuki Shinnou Kanako Komiya Ibaraki University, Department of Computer and Information Sciences 4-12-1 Nakanarusawa, Hitachi, Ibaraki JAPAN 316-8511 {17nm719x, hiroyuki.shinnou.0828, kanako.komiya.nlp} @vc.ibaraki.ac.jp Abstract This paper proposes a new method of instance-based domain adaptation for senti- ment analysis. First, our method defines the likelihood of keywords, through the value of inverse document frequency (IDF), for each word in documents in the target domain. Next, the keyword content rate of a document is calculated using the likelihood of keywords and the domain adaptation is performed by giving the keyword content rate to each doc- ument in the source domain as the weight. The experiment used an Amazon dataset to demonstrate the effectiveness of our proposed method. Although the instance-based method has not shown great efficiency, the advan- tages combining instance-based method and feature-based method are shown in this paper. 1 Introduction This paper proposes a new method of instance-based domain adaptation for sentiment analysis. Sen- timent analysis involves judging a polarity, posi- tive or negative, of a review such as a movie re- view. This is one of the document classification tasks and supervised learning can be used to solve it. However, if the domain of the test data is dif- ferent from the domain for the learning data (for example, book reviews), the accuracy of the clas- sifier obtained through standard supervised learning reduced. This is the problem with domain shift. The solution to this problem is domain adaptation. Domain adaptation can be roughly divided into two categories: feature-based and instance-based (Pan and Yang, 2010). In summary, both are weighted- learning methods, but feature-based gives weights to features and instance-based gives weight to instance. Here, we present a new instance-based method. Generally, the instance-based method assumes a co- variate shift, and gives the weight based on the probability density ratio between target domain and source domain. However, the computational cost for the instance-based method is too high. The method presented here is simple and its effect is better than methods using a typical probability density ratio. Our method first defines lw, the likelihood of the keyword of the word w using the IDF in the target domain. Using lw, the weight of a review x in the source domain is set as the keyword content rate wx. After that, weighted-learning is performed by giv- ing wx to each document in the source domain x to overcome the domain shift. In the experiment, we used Amazon dataset (Blitzer et al., 2007), and compared our proposed method with two typical instance-based methods: unconstrained least squares importance fitting (uL- SIF) (Yamada et al., 2011) using the probability den- sity ratio and the method defining weight through Naive Bayes model (Shinnou and Sasaki, 2014), to demonstrate the effectiveness of the proposed method. 2 Related Work Domain adaptation is roughly divided into two types: the supervised approach using labeled data in the target domain and the unsupervised approach that …

Content truncated to 500 words