Location: 3105, Lambton Tower 

Data-driven learning of document vectors that capture linkage between them is of immense importance in natural language processing (NLP) for downstream tasks like information retrieval, classification, and clustering. Inherently, documents are linked together in the form of links or citations in case of web pages or academic papers respectively. Methods like PV-DBOW try to capture the semantic representation of the document using only the textual information and ignore the linkage information altogether. Similarly, methods developed for network representation learning like Node2Vec capture the linkage information between the documents but they ignore the textual information. In this thesis, we proposed a method based on Retrofit for learning word embeddings using a semantic lexicon. Our approach tries to incorporate both the textual and network information together while learning the document representation. Our experimentation result shows that our method improves the classification score by 4% and we analyze the best weight for adding network information to the embeddings. Furthermore, we also introduce a new dataset containing both network and content information. 

 

 

Thesis Committee:

Thesis Defense Announcement