Robust Cross-Domain Sentiment Analysis for Low-Resource Languages

Abstract

While various approaches to domain adaptation exist, the majority of them requires knowledge of the target domain, and additional data, preferably labeled. For a language like English, it is often feasible to match most of those conditions, but in low-resource languages, it presents a problem. We explore the situation when neither data nor other information about the target domain is available. We use two samples of Danish, a low-resource language, from the consumer review domain (film vs. company reviews) in a sentiment analysis task. We observe dramatic performance drops when moving from one domain to the other. We then introduce a simple offline method that makes models more robust towards unseen domains, and observe relative improvements of more than 50%.

Demographic Factors Improve Classification Performance

“…To collect data for the sentiment analysis task, we select all reviews that contain the target variable (gender or age), and a star-rating. Following previous work on similar data (Blitzer et al, 2007;Hardt and Wulff, 2012;Elming et al, 2014), we use one, three, or five star ratings, corresponding to negative, neutral, and positive sentiment, respectively.…”

Opinion Analysis Corpora Across Languages

“…Many researchers have focused on a resourceless approach to sentiment analysis (Elming et al 2014;Le et al 2016). Blitzer et al (2007) proposed a domain adaptation approach for sentiment classification.…”

The Rating Game: Sentiment Rating Reproducibility from Text

“…Since Pang et al (2002) used author-labeled IMDb user reviews in their seminal study, author-labeled data has been used for a wide range of domains, like user-generated product reviews (Dave et al, 2003), restaurant reviews with several aspect ratings (Snyder and Barzilay, 2007), movie reviews from experienced film critics (Pang and Lee, 2005), business reviews (Hardt and Wulff, 2012;Elming et al, 2014;Hovy, 2015), and many more. Pang and Lee (2005) also argue that it is unreasonable to expect a learning algorithm to predict ratings on a fine-grained scale if humans are not able to do so.…”