10 July 2018- Hady Lauw: Seminar
Representation Learning for Subjective Expressions of Preferences
Users' preferences increasingly manifest themselves through multiple forms of preference signals, including consumption, ratings, text reviews, images, etc. There are now large corpora of such preference expressions that inevitably embody some level of subjectivity. Representation learning from such corpora could be affected by the inherent subjectivity, which may impact follow-on tasks that rely on the learnt representations.
In the first part of the talk, we explore the notion of subjectivity for word representations. Distributional word embedding methods, such as Word2Vec or Glove, have been critical for the success of many large scale natural language processing applications. We hypothesize that the varying levels of subjectivity in input corpora may affect word embeddings for text classification (e.g., sentiment, subjectivity, topic). Through systematic comparative analysis, we discover the outsize role that sentiment words play on subjectivity-sensitive tasks, and develop a novel word embedding method, SentiVec, which is infused with sentiment information from a lexical resource and is shown to outperform baselines on such tasks.
In the second part of the talk, we explore the notion of subjectivity for image representations. Deep learning methods, such as Convolutional Neural Networks or CNN, are effective in deriving representations for data with spatial characteristics, such as images. Visual sentiment analysis seeks to infer whether a given image included as part of an online review expresses an overall positive or negative sentiment. Observing that the sentiment captured within an image may be affected by the user or item associated with the review, we develop CNN-based models for visual sentiment analysis that incorporate the potential subjectivity due to user or item, which are shown to be more effective on images from restaurant reviews.
Hady W. Lauw is a faculty member at the School of Information Systems, Singapore Management University. Currently, he is also an NRF Fellow of the Singapore National Research Foundation. Formerly, he served as postdoctoral researcher at Microsoft Research in Silicon Valley, as well as scientist at A*STAR's Institute for Infocomm Research. He received his PhD from Nanyang Technological University on A*STAR Graduate Scholarship. At SMU, he leads the Preferred.AI research project, whose research activities span data mining and machine learning, focusing on preference analytics and recommender systems. More information may be found at http://www.hadylauw.com.