Extracting word lists for domain-specific implicit opinions from corpora

Sentiment analysis relies to a large extent on lexical resources. While lists of words bearing a contextindependent evaluative polarity (‘great’, ‘bad’) are available for many languages now, the automatic extraction of domain-specific evaluative vocabulary still needs attention. This holds especially for implicit opinions or so-called polar facts. In our work, we focus on German and on a genre that has not received much attention yet: customer emails. As the prime downstream application is identifying customers’ complaints, we concentrate here on finding negative words, but our method applies to positive ones as well. Using a seed list approach, we provide a comparative analysis along three dimensions: effect of different seed lists, different linguistic analysis units, and different statistical correlation tests.