Date of Award

Spring 2022

Project Type

Dissertation

Program or Major

Computer Science

Degree Name

Doctor of Philosophy

First Advisor

Laura Dietz

Second Advisor

Dongpeng Xu

Third Advisor

Elizabeth Varki

Abstract

Information Retrieval (IR) refers to obtaining valuable and relevant information from various sources in response to a specific information need. For the textual domain, the most common form of information sources is a collection of textual documents or text corpus. Depending on the scope of the information need, also referred to as the query, the relevant information can span a wide range of topical themes. Hence, the relevant information may often be scattered through multiple documents in the corpus, and each satisfies the information need to varying degrees. Traditional IR systems present the relevant set of documents in the form of a ranking where the rank of a particular document corresponds to its degree of relevance to the query.

If the query is sufficiently specific, the set of relevant documents will be more or less about similar topics. However, they will be much more topically diverse when the query is vague or about a generalized topic, e.g., ``Computer science." In such cases, multiple documents may be of equal importance as each represents a specific facade of the broad topic of the query. Consider, for example, documents related to information retrieval and machine learning for the query ``Computer Science." In this case, the decision to rank documents from these two subtopics would be ambiguous. Instead, presenting the retrieved results as a cluster of documents where each cluster represents one subtopic would be more appropriate. Subtopic clustering of search results has been explored in the domain of Web-search, where users receive relevant clusters of search results in response to their query.

This thesis explores query-specific subtopic clustering that incorporates queries into the clustering framework. We develop a query-specific similarity metric that governs a hierarchical clustering algorithm. The similarity metric is trained to predict whether a pair of relevant documents should also share the same subtopic cluster in the context of the query. Our empirical study shows that direct involvement of the query in the clustering model significantly improves the clustering performance over a state-of-the-art neural approach on two publicly available datasets. Further qualitative studies provide insights into the strengths and limitations of our proposed approach.

In addition to query-specific similarity metrics, this thesis also explores a new supervised clustering paradigm that directly optimizes for a clustering metric. Being discrete functions, existing approaches for supervised clustering find it difficult to use a clustering metric for optimization. We propose a scalable training strategy for document embedding models that directly optimizes for the RAND index, a clustering quality metric. Our method outperforms a strong neural approach and other unsupervised baselines on two publicly available datasets. This suggests that optimizing directly for the clustering outcome indeed yields better document representations suitable for clustering.

This thesis also studies the generalizability of our findings by incorporating the query-specific clustering approach and our clustering metric-based optimization technique into a single end-to-end supervised clustering model. Also, we extend our methods to different clustering algorithms to show that our approaches are not dependent on any specific clustering algorithm. Having such a generalized query-specific clustering model will help to revolutionize the way digital information is organized, archived, and presented to the user in a context-aware manner.

Share

COinS