PhD Thesis Defense: Prasanna Subramanyam, A Computational Approach To Understanding Online Community Dynamics and Opinion
Content
Speaker:
Abstract:
This research advances methodologies for analyzing textual content from large-scale online datasets, interpreting message intent, retrieving knowledge, and leveraging these insights to address practical challenges across diverse domains.
Firstly, I use two complementary approaches to improve Temporal Question Answering. Temporal Question Reasoning (TempoQR) uses a Retrieval Augmented Generation (RAG) inspired approach to enhance the question representation using relevant facts from a large knowledge graph. Language Aware SPARQL Retrieval (LASR), on the other hand, focuses on understanding the intent of the question to pick the best SPARQL query to answer temporal questions. Through extensive experiments on the CronQuestions benchmark, I demonstrate that TempoQR and LASR improve Hits@1 by 27 and 33 percentage points over state-of-the-art approaches, respectively. Building on this approach, I use intent classification to classify messages with labels of interest to investigators to help them triage massive corpora of conversations. The classification model enables a novel conversation clustering technique, revealing distinct conversational patterns and providing insights into different survivor experiences.
Secondly, I advance a method for understanding the opinions expressed within large-scale online communities. I propose a self-supervised approach to model collective community opinion by leveraging readily available data. This overall solution is based on two novel techniques for community comparison: Binary Opinion Transfer Similarity (BOTS), which calculates similarity based on expressed opinions; and Embedding-based Post Similarity Ratio (Emb-PSR), which calculates similarity based on content. My quantitative results demonstrate that both BOTS and Emb-PSR outperform existing methods by 13 and 36 percentage points, respectively, and facilitate cross-platform comparisons between communities.
Finally, to move beyond the black box nature of global similarity, I introduce methods that capture content and opinion similarities between communities at the level of topics and subtopics. Hierarchical Topic Similarity (HiT-Sim), a fine-grained content similarity method, not only improves the subreddit retrieval performance by 12 percentage points over existing methods, but is also interpretable, decomposable, and allows users to condition similarity on specific topics. The representation used in HiT-Sim enables exploratory analyses to identify topics that define specific categories of communities and those that connect distinct groups. Similarly, Topic-BOTS extends BOTS by providing opinion similarity at the topic level while offering greater visibility into the direction of agreement. Together, these methods allow for the interpretation and explanation of community similarity at a granular level.
Advisor:
Brian Levine