This thesis proposal outlines the use of unsu-pervised data-driven methods for paraphrasing tasks. We motivate the development of knowledge-free methods at the guiding use case of multi-document summarization, which requires a domain-adaptable system for both the detection and generation of sentential paraphrases. First, we define a number of guiding research questions that will be addressed in the scope of this thesis. We continue to present ongoing work in unsupervised lexical substitution. An existing supervised approach is first adapted to a new language and dataset. We observe that supervised lexical substitution relies heavily on lexical semantic resources, and present an approach to overcome this dependency. We describe a method for unsupervised relation extraction, which we aim to leverage in lexical substitution as a replacement for knowledge-based resources.
展开▼
