Rakesh Gupta, Lev Ratinov
Multi-category classification of short dialogues is a common task performed by humans. When assigning a question to an expert, a customer service operator tries to classify the customer query into one of N different classes for which experts are available. Similarly, questions on the web (for example questions at Yahoo Answers) can be automatically forwarded to a restricted group of people with a specific expertise. Typical questions are short and assume background world knowledge for correct classification. With exponentially increasing amount of knowledge available, with distinct properties (labeled vs unlabeled, structured vs unstructured), no single knowledge-transfer algorithm such as transfer learning, multi-task learning or selftaught learning can be applied universally. In this work we show that bag-of-words classifiers performs poorly on noisy short conversational text snippets. We present an algorithm for leveraging heterogeneous data sources and algorithms with significant improvements over any single algorithm, rivaling human performance. Using different algorithms for each knowledge source we use mutual information to aggressively prune features. With heterogeneous data sources including Wikipedia, Open Directory Project (ODP), and Yahoo Answers, we show 89.4% and 96.8% correct classification on Google Answers corpus and Switchboard corpus using only 200 features/class. This reflects a huge improvement over bag of words approaches and 48-65% error reduction over previously published state of art (Gabrilovich et. al. 2006).
Subjects: 13. Natural Language Processing
Submitted: Apr 15, 2008