Export
Research Overview
Anupam Guha focuses on natural language processing, coreference resolution, and multimodal narrative understanding. Early work addressed the creation of challenging linguistic datasets and the application of active learning to complex coreference tasks. More recently, this focus has expanded to include unsupervised models for literary analysis and visual reasoning across stylized media formats.
Research Trajectory
Guha’s research trajectory began with a focus on foundational challenges in natural language processing, particularly in building robust systems for coreference resolution. In 2015, Guha led the development of a novel dataset designed to push beyond the limitations of standard newswire text. In the first-author paper Removing the Training Wheels: A Coreference Dataset that Entertains Humans and Challenges Computers, Guha and collaborators utilized quiz bowl questions to create a dense, highly referential corpus. This work demonstrated that state-of-the-art coreference systems of the time underperformed simple classifiers on this new domain. To optimize the data collection process, the project integrated techniques from the research direction of Active Learning via Model Disagreement, ensuring that human annotation efforts were directed toward the most informative examples. This foundational work also intersected with broader challenges in the field, such as the area of Resolving Split Antecedents in Plural Anaphora and themes related to the Expert Annotation for Specialized Machine Learning.
In 2016, Guha’s focus expanded into the digital humanities, applying unsupervised machine learning techniques to literary analysis. As the second author of Feuding Families and Former Friends: Unsupervised Learning for Dynamic Fictional Relationships, Guha contributed to a novel neural network architecture designed to track how relationships between characters evolve over the course of a novel. Rather than relying on pre-annotated plot summaries, the model utilized dictionary learning to jointly generate interpretable relationship descriptors and trajectories directly from raw text. This approach to disentangling complex, overlapping textual features aligns with the broader methodological area of Orthogonality Constraints for Diverse Representation Learning.
By 2017, Guha’s collaborative work shifted toward multimodal systems and visual narrative understanding. As a middle author, Guha participated in the development of The Amazing Mysteries of the Gutter: Drawing Inferences Between Panels in Comic Book Narratives. This research investigated whether computational models could replicate the human cognitive process of "closure"—inferring unseen actions and temporal movements in the gutters between comic book panels. The project introduced a massive dataset of 1.2 million stylized comic panels paired with dialogue transcriptions. This work required models to navigate the research direction of Generalizing Vision Models to Stylized and Abstract Domains, as the models had to interpret stylized artwork rather than natural photographs, while also engaging with the theme of Relational Reasoning for Spatial Disambiguation to logically connect sequential scenes.
Current Directions
The most recent paper in the evidence is from 2017. Based on this work, Guha’s latest research directions investigate multimodal narrative understanding, specifically focusing on how computational models can infer implicit actions and transitions in stylized visual media. By examining sequential art and comic book panels, this direction explores the intersection of computer vision and natural language processing. The primary open problems in this area involve teaching models to perform logical closure between disjointed scenes and adapting visual reasoning architectures to succeed on non-photographic, highly stylized domains where traditional object recognition models frequently fail.
Profile updated
Loading new papers…
Loading research neighbors…
See something wrong? Report a mistake