Skip to main content

‘Bold,’ ‘positive’ and ‘unparalleled’: Allen School Ph.D. graduates Ashish Sharma and Sewon Min recognized with ACM Doctoral Dissertation Awards

Each year, the Association for Computing Machinery (ACM) recognizes the best Ph.D. dissertations in computer science with its ACM Doctoral Dissertation Award. In the 2024 competition, two of the recognized dissertations were the work of Allen School students: award winner Ashish Sharma (Ph.D., ‘24), now a senior applied scientist at Microsoft, and honorable mention recipient Sewon Min (Ph.D., ‘24), a research scientist at the Allen Institute for AI (Ai2) and incoming faculty member at the University of California, Berkeley. 

Both Sharma and Min contributed to advances in artificial intelligence, albeit in different domains — highlighting both the variety and quality of AI research in the Allen School.

For his dissertation titled “Human-AI Collaboration to Support Mental Health and Well-Being,” Sharma devised ways to address a fundamental challenge in health care by leveraging AI to make high-quality mental health support available to more people. Meanwhile, in her dissertation titled “Rethinking Data Use in Large Language Models,” Min addressed fundamental challenges in natural language processing (NLP) by developing a new class of language models (LMs) and alternative approaches for how such models are trained.

Ashish Sharma: “Human-AI Collaboration to Support Mental Health and Well-Being”

Headshot of Ashish Sharma
Ashish Sharma

Therapy can be an effective tool for supporting those with mental health challenges, but barriers such as an ongoing shortage of clinicians, high costs as well as stigma with seeking care can limit access. Instead of fully replacing therapists and clinicians with AI, given the significant risks inherent in this domain, Sharma proposes an alternative: he developed two novel human-AI collaboration systems designed to augment, rather than replace, human providers.

“Augmenting mental health interventions with AI and NLP-based methods has the potential to provide scaffolding that could make quality mental health care accessible to all,” said Sharma,  who completed his dissertation as part of the Allen School’s Behavioral Data Science Group. “By carefully designing human-AI collaboration that is grounded in psychology expertise to truly understand the complexities of mental health, human behavior and user needs, and is rigorously tested for safety and effectiveness, we can empower both those seeking help and those providing it.”

Sharma introduced reinforcement learning-based methods that can understand, measure and give feedback on how empathy is expressed in online peer-to-peer mental health support platforms. While many peer supporters are well-intentioned in helping those who reach out, they may be untrained and unaware of key psychotherapy skills such as empathy that can foster more effective conversations. He then leveraged and evaluated these methods in a randomized trial of 300 real-world peer supporters from TalkLife, one of the largest, global peer support platforms, and found that the AI-based feedback helped peer supporters express empathy more effectively in their conversations. The research received the Best Paper Award at The Web Conference 2021.

Human-AI collaboration can also enhance the accessibility and engagement of self-guided mental health interventions. These “do-it-yourself” methods to learn and practice coping skills are often cognitively demanding and emotionally triggering, making it difficult to implement them on a wider scale, Sharma explained. Building on psychological and cognitive science theories, he developed human-centered NLP methods to help “debug” human thought and support people through the process of cognitive reframing – that is, identifying and overcoming negative thoughts. In a randomized study of more than 15,000 participants, Sharma showed that the system helped participants reframe negative thoughts and informed psychology theory about the processes that lead to positive outcomes. He deployed this system at Mental Health America, which provides mental health tools and resources, and it has been used by over 160,000 users.

“Ashish’s dissertation is highly interdisciplinary and unparalleled in combining fundamental advances in natural language processing with large-scale, positive, immediate impacts on the mental health of large populations,” said Allen School professor Tim Althoff, who advised Sharma. “To date, his research has directly improved mental health services that support more than 10 million people yearly — an exceptional feat for any researcher.”

Prior to the ACM’s recognition of his work, Sharma received one of two William Chan Memorial Dissertation Awards, which are named for the late Allen School graduate student William Chan and recognize dissertations of exceptional merit, as well as a JP Morgan AI Ph.D. Fellowship.

Sewon Min: “Rethinking Data Use in Large Language Models”

Headshot of Sewon Min
Sewon Min

Although current LMs including ChatGPT have transformed NLP progress, they still have fundamental issues, such as factuality and privacy, that arise from how they learn to perform new tasks after training. The widespread belief was that LMs obtain new skills on the fly without additional training through in-context learning; however, Min showed that LMs’ in-context learning capabilities are actually based on patterns they learn in their training data, which can be activated in certain ways. Based on this understanding, she introduced a new class of models called nonparametric LMs. 

“This new class of LMs includes learned parameters and a datastore, from which they retrieve information for improved accuracy and updatability,” Min explained. 

During inference, a nonparametric model can identify and reason with relevant text from its datastore, unlike a conventional model that must remember every relevant detail from its training set. Having a datastore present at inference time can help lead to more efficient and flexible LMs. 

“My Ph.D. thesis is about understanding and advancing large language models centered around how they use the very large text corpora they are trained on,” said Min, who was part of the UW NLP group. “My research established the foundations of nonparametric models, and also opened up new avenues for responsible data use, such as enabling data opt-out and credit assignment to data creators.”

These nonparametric LMs include retrieval-augmented generation (RAG), and her research has helped establish the technique. However, “the recent use of RAG has been mainly using an off-the-shelf retrieval model and an off-the-shelf LM and plugging them together without training the model, whereas my research advocates for developing new architectures and training methods that allow for more effective and efficient use of the datastore,” Min explained.

Nonparametric LMs can lead to new approaches to avoid the legal constraints that traditional LMs often run into. It is common practice to train LMs using all available online data, but this approach can lead to concerns with copyrights and crediting data creators. Instead, Min developed a new method based on nonparametric LMs — training LMs using public domain data, while keeping copyrighted or other high-risk data in a datastore that is only accessed during inference and can be modified at any time. 

“Sewon’s thesis identifies bold, impactful and challenging problems that many researchers shy away from, and then designs creative technical solutions to address these problems,” said Allen School professor Hannaneh Hajishirzi, who is also a senior director for NLP research at Ai2 and co-advised Min alongside faculty colleague Luke Zettlemoyer. “Her ambitious, creative and forward-thinking vision is complemented by a foundation of technical, structured, mathematical and analytical strengths, leading to groundbreaking and pioneering research.”

The ACM honor is not the only accolade Min has earned for her dissertation work; she, too, earned a William Chan Memorial Dissertation Award from the Allen School, as well as the 2024 Western Association of Graduate Schools (WAGS) ProQuest Innovation in Technology Award, which recognizes research that introduces innovative technology as a creative solution to a major problem. During her time at the Allen School, Min received a JP Morgan Ph.D. Fellowship in AI and was also named a 2022 EECS Rising Star.

Read more about the ACM Doctoral Dissertation Awards as well as a related GeekWire story.