Temporal Framework for Natural Language Processing with Convolutional Networks

In this research, we propose a model learning documents to fixed-length embedding vector space. This is meaningful, because in vector space, we can find similar documents or measure the relations between documents by simple linear algebra calculation. One of state-of-the-art methods is to apply deep Convolutional Neural Network on language sequences and thus learn different levels of features. In these methods, aggregating functions (e.g. max) are used to address the variable-length problem of documents. However, the aggregating functions also result disadvantages. They destroy the temporal structure of the documents, which is not ideal for a language model. In this work we propose an approach that leverages the advantages of the CNN architecture but avoids most of the disadvantages. TO BE CONT’D

Faculty Supervisor:

Sanja Fidler

Student:

Partner:

Layer 6 AI

Discipline:

Computer science

Sector:

Information and Communications Technology; Technology; Other

University:

University of Toronto

Program: