There is a large amount of information that can be extracted automatically from social media messages. Of particular interest are the topics discussed by the users, the opinions and emotions expressed, and the events and the locations mentioned. This work focuses on machine learning methods for detecting locations from Twitter messages, because the extracted locations can be useful in business, marketing and defence applications (Farzindar and Inkpen, 2015).
There are two types of locations that we are
interested in: location entities mentioned in the text
of each message and the physical locations of the
users. For the first type of locations (task 1), we
detected expressions that denote locations and
we classified them into names of cities,
provinces/states, and countries. We approached the task
in a novel way, consisting in two stages. In the
first stage, we trained Conditional Random Field
models with various sets of features. We
collected and annotated our own dataset for training and
testing. In the second stage, we resolved cases
when more than one place with the same name
exists, by applying a set of heuristics
For the second type of locations (task 2), we put
together all the tweets written by a user, in order
to predict his/her physical location. Only a few
users declare their locations in their Twitter
profiles, but this is sufficient to automatically
produce training and test data for our classifiers. We
experimented with two existing datasets
collected from users located in the U.S. We propose a
deep learning architecture for the solving the
task, because deep learning was shown to work
well for other natural language processing tasks,
and because standard classifiers were already
tested for the user location task. We designed a
model that predicts the U.S. region of the user
and his/her U.S. state, and another model that
predicts the longitude and latitude of the user's
location. We found that stacked denoising
autoencoders are well suited for this task, with results
comparable to the state-of-the-art
Diana Inkpen is a Professor at the University of Ottawa, in the School of Electrical Engineering and Computer Science. Her research is in applications of Computational Linguistics and Text Mining. She organized seven international workshops and she was a program co-chair for the AI 2012 conference. She is in the program committees of many conferences and an associate editor of the Computational Intelligence and the Natural Language Engineering journals. She published a book on Natural Language Processing for Social Media (Morgan and Claypool Publishers, Synthesis Lectures on Human Language Technologies), 8 book chapters, more than 25 journal articles and more than 90 conference papers. She received many research grants, including intensive industrial collaborations.
I want to thank my collaborators on the two tasks addressed in this work: Ji Rex Liu for his implementation of the proposed methods for both tasks and Atefeh Farzindar for her insight on the first task. I also thank the two annotators of the corpus for task 1, Farzaneh Kazemi and Ji Rex Liu. This work is funded by the Natural Sciences and Engineering Research Council of Canada.