CS224N Final Project: QA

We extend existing question answering (QA) system to deal with words that are unseen or do not have enough examples during training. Instead of learning word embedding from scratch for a specific QA dataset, we decompose the embedding into two components: The first captures the general semantic meaning of a word and can be trained using readily available large corpuses. The second component reweights the general embedding so that it is optimized for our QA tasks. The advantage is twofold. First, the number of parameters to learn using the QA dataset does not grow linearly with the size of the vocabulary. Second, the system can be easily generalized to unseen words by borrowing the knowledge from outside corpuses.