I am building a question answering system that given a legal document attempts to answer questions related to the document. For example a tenancy agreement is given to the system and the user asks questions about the agreement.
I am splitting the legal document into small segments and try to find the most similar segment to each question. I am using a combination of
tf-idf and word-embeddings based similarity measures.
The problem I have at the moment is the questions are asked by non-legal experts and almost entirely are in everyday language, therefore there is a huge mismatch between terms used in the questions and the answers (segments). Because I do not have access to a huge collection of legal documents, I am using the prebuilt GloVe vectors. Even if I manage to estimate the word embeddings based on relevant legal documents the mismatch problem would still remain.
How can I deal with this problem? Is there a method to augment the word embeddings to fit a certain domain, like this?