For the first week of our research project I have been learning as much as I can about BiMPM. I spent most of the week studying a paper by Zhiguo Wang, Wael Hamza, and Radu Florian titled “Bilateral Multi-Perspective Matching for Natural Language Sentences”. The paper explained how BiMPM can be used for NLSM. This can be used to check whether or not two sentences mean the same thing. We will be using BiMPM throughout our project to see if news stories are saying the same thing. I, along with other members of this team, are planning on training BiMPM and learning how to run it on the server we are using for our project.
It works by first encoding two sentences, say P and Q, with a bidirectional Long ShortTerm Memory Network (BiLSTM). It then matches both of the encoded sentences in two directions, P against Q and Q against P. Suppose we have P against Q, every time step of Q will be matched against all the time-steps of P from multiple perspectives. Another layer of BiLSTM is then used to aggregate the matching results into a fixed-length matching vector. Lastly, depending on the matching vector, a conclusion is made through a fully connected layer.