Natural Language Processing Lab
From left to right: Vikash Balasubramanian, Lili Mou, Amirpasha Ghabussi, Raphael Schumann, Yao Lu, Kashif Khan, Gaurav Sahu. Front: Olga Vechtomova
The NLP Lab at the University of Waterloo conducts research to solve problems involving natural language text. The ultimate goal of NLP research is to achieve a human-like machine understanding of natural language. While this goal is yet to be reached, big strides have been made in the NLP field to automate many important tasks involving language, such as question answering and opinion mining. Below is a selection of our current and recent research projects:
Task based information retrieval
Supported by an Amazon Research Award, this project aims to design neural IR models for recommending products to a user based on the descriptions of activities and tasks that he/she wants to perform.
Neural generative models, such as variational autoencoders (VAEs) and variational encoder-decoders (VEDs) are powerful models that are trained in an unsupervised manner, and can be used to generate text. One of the research problems we are interested in is controlled text generation, whereby we can influence specific characteristics of the generated text, such as style and emotion. Other examples of current research problems are variational attention and latent feature disentanglement.
Text style transfer
The goal of style transfer models is to take text written in one style, and re-write it in a different style, while maintaining the meaning of the original text. An example of a practical application of text style transfer is automatic editing of a manuscript to conform to the writing style of a journal.
The research on sentiment analysis aims to develop methods for identifying sentiment (positive or negative) and emotions expressed in text. Our research projects in this field include: detecting contextual sentiment polarity of ambiguous words (e.g. "cold", "warm"); identifying polarity of opinions expressed about a specific aspect of an entity; determining sentiment of financial news headlines; identifying the strength of emotions expressed in tweets; detecting stance towards rumours expressed in tweets.
The field of entity retrieval is concerned with extracting and retrieving information about entities, such as products, people and organizations. We did a number of projects in this field, such as retrieval of entities that have a specific relation to an entity in the user query (e.g., "Find artists who won the Turner prize") or unsupervised extraction and clustering of product aspects from user reviews.
Current and past research support from:
Amazon Research Award
Natural Sciences and Engineering Research Council (NSERC)
Federal Economic Development Agency for the Southern Ontario
InsightNG Solutions Ltd.
Canada Foundation for Innovation