| Summary: | Fine grained extraction is an important subtask in natural language processing and sentiment analysis which aims to extract ‘aspect’ terms that describe properties of entities and ‘opinion’ terms that convey user emotion and sentiment from natural language text. While multiple models and research techniques have been proposed recently to solve this task, these techniques are confined to the ‘source domain’ or the inherent structure of training data.
Application of these models in new or ‘target domains’ constitutes the significant overheads of human effort in labelling of target domain data and computational time for retraining of model. This limits the potential of such models in the industry and is also unlike the human mind, which is adept at identifying commonalities between different domains. Thus, developing techniques for domain adaptation for fine grained extraction models is an extremely relevant sub-problem for industrial applications as well as development of general intelligent machines.
In this final year project, we first give an overview of the problems of fine grained natural language extraction and domain adaptation, and review corresponding literature and related research fields. We then decompose the primary problem of domain adaptation of fine grained extraction models into relevant subtasks, and review and document performance of existing research methods for domain adaptation on the “Laptop” and “Restaurant” domains of the Semeval Challenge 2014 Task 4 dataset. Finally, we experiment with the usage of unsupervised techniques to measure the syntactic, semantic (statistical) and conceptual impact of removing a word on its sentence, and document the resulting performance of using the mentioned word removal measure as an additional word feature.
|
|---|