Contextual Embedding for Distributed Representations of Entities in a Text Corpus
; Proceedings of the 5th International Workshop on Big Data, Streams and Heterogeneous Source Mining: Algorithms, Systems, Programming Models and Applications at KDD 2016, PMLR 53:35-50, 2016.
Distributed representations of textual elements in low dimensional vector space to capture context has gained great attention recently. Current state-of-the-art word embedding techniques compute distributed representations using co-occurrences of words within a contextual window discounting the flexibility to incorporate other contextual phenomena like temporal, geographical, and topical contexts. In this paper, we present a flexible framework that has the ability to leverage temporal, geographical, and topical information of documents along with the textual content to produce more effective vector representations of entities or words within a document collection. The framework first captures contextual relationships between entities collected from different relevant documents and then leverages these relationships to produce inputs of a graph, or to train a neural network to produce vectors for the entities. Through a set of rigorous experiments we test the performance of our approach and results show that our proposed solution can produce more meaningful vectors than the state-of-the-art methods.