Distant-supervised algorithms with applications to text mining, product search, and scholarly networks [thesis]
Author
Saurav Manchanda (Ph.D. 2020)
Abstract
In recent times, data has become the lifeblood of pretty much all businesses. As such, the real-world impact of data-driven machine learning has grown in leaps and bounds. It has set up itself as a standard tool for organizations to draw insights from the data at scale, and hence, to enhance their profits. However, one of the key-bottlenecks in deploying machine learning models in practice is the unavailability of labeled training data. The manually-labeled training sets are expensive and it can be a tedious exercise to create them. Besides, they cannot be practically reused for new objectives, if the underlying distribution of data changes with time. As such, distant-supervision provides a solution to using expensive hand-labeled datasets, which means leveraging alternative sources of weak-supervision. In this thesis, we identify and provide solutions to some of the challenges that can benefit from distant-supervised approaches. First, we present a distant-supervised approach to accurately and efficiently estimate a vector representation for each sense of the multi-sense words. Second, we present approaches for distant-supervised text-segmentation and annotation, which is the task of associating individual parts in a multilabel document with their most appropriate class labels. Third, we present approaches for query understanding in product search. Specifically, we developed distant-supervised solutions to three challenges in query understanding: (i) when multiple terms are present in a query, determining the relevant terms that are representative of the query’s product intent, (ii) vocabulary gap between the terms in the query and the product’s description, and (iii) annotating individual terms in a query with the corresponding intended product characteristics (product type, brand, gender, size, color, etc.). Fourth, we present approaches to estimate content-aware bibliometrics to accurately quantitatively measure the scholarly impact of a publication. Our proposed metric assigns content-aware weights to the edges of a citation network, that quantify the extent to which the cited-node informs the citing-node. Consequently, this weighted network can be used to derive impact metrics for the various involved entities, like the publications, authors, etc.
Link to full paper
Keywords
machine learning, text mining