Modeling the meaning of words and sentences is core to NLP. In this talk, I discuss several neural-net models we developed recently, including models that consider both compositional and non-compositional semantics to estimate the meaning of a sentence. I will also discuss several specific architectures, i.e., LSTMs with tree and DAG (directed acyclic graph) structures, which achieved state-of-the-art performance on two semantic composition tasks.
The content is based on an ICML, a *SEM paper, and our on-going work: http://www.xiaodanzhu.com/publications.html
This is joint work with Parinaz Sobhani and Harry (Hongyu) Guo.