Domain adaptation in Natural Language Processing

Domain adaptation has received much attention in the past decade. It has been shown that domain knowledge is paramount for building successful Natural Language Processing (NLP) applications. To investigate the domain adaptation problem, we conduct several experiments from different perspectives. First, we automatically adapt sentiment dictionaries for predicting the financial outcomes “excess return” and “volatility”. In these experiments, we compare manual adaptation of the domain-general dictionary with automatic adaptation, and manual adaptation with a combination consisting of first manual, then automatic adaptation. We demonstrate that automatic adaptation performs better than manual adaptation, namely the automatically adapted sentiment dictionary outperforms the previous state of the art in predicting excess return and volatility. Furthermore, we perform qualitative and quantitative analyses finding that annotation based on an expert’s a priori belief about a word’s meaning is error-prone – the meaning of a word can only be recognized in the context that it appears in. Second, we develop the temporal transfer learning approach to account for the language change in social media. The language of social media is changing rapidly – new words appear in the vocabulary, and new trends are constantly emerging. Temporal transfer-learning allows us to model these temporal dynamics in the document collection. We show that this method significantly improves the prediction of movie sales from discussions on social media forums. In particular, we illustrate the success of parameter transfer, the importance of textual information for financial prediction, and show that temporal transfer learning can capture temporal trends in the data by focusing on those features that are relevant in a particular time step, i.e., we obtain more robust models preventing overfitting. Third, we compare the performance of various domain adaptation models in low-resource settings, i.e., when there is a lack of large amounts of high-quality training data. This is an important issue in computational linguistics since the success of NLP applications primarily depends on the availability of training data. In real-world scenarios, the data is often too restricted and specialized. In our experiments, we evaluate different domain adaptation methods under these assumptions and find the most appropriate techniques for such a low-data problem. Furthermore, we discuss the conditions under which one approach substantially outperforms the other. Finally, we summarize our work on domain adaptation in NLP and discuss possible future work topics.