Odds are, at some point, you've communicated with a chatbot or other digital customer service rep. These neural models for question answering (QA) have improved significantly in the past few years, but there's still a ways to go.


Enter Futureforce intern Sewon Min, a student at Seoul National University in South Korea who is planning to work on her Ph.D. at the University of Washington. Sewon worked with our Salesforce Research team this past year studying the minimal conversational context required for QA and found that most questions can be answered within a small set of sentences.



We sat down with her to learn more about this:


Can you describe the work you did during your three-month Salesforce internship?

I focused on Natural Language Processing (NLP) and worked with Salesforce Research Scientists to build an efficient system that could read a document and answer a question. We published our work as a conference paper at the 56th Annual Meeting of the Association for Computational Linguistics (ACL).

How exactly does this QA system work, and what sort of impact might it have on Salesforce's business?

Our system selects sentences from a document based on a related question, then predicts the answer based on a complex interaction between the selected sentences and the question. We were motivated by the observation that even when a document is large, most questions can be answered by examining 1-2 sentences. This is closely related to customer services in that it can help the businesses build automatic services to answer the questions that customers ask. Since our system is fast and scalable, it can allow a system to answer a wide range of questions in real-time.

Your research paper “Efficient and Robust Question Answering from Minimal Context over Documents” got selected for the ACL conference, how did that feel?

I was so happy that people found our paper interesting and helpful, and thought it worthy to appear at the conference.

How would you describe the team collaboration on this research and your internship as a whole?

Salesforce Research is a leading research team in AI, and being a part of it was an invaluable experience for me. The teams' projects varied on topics from NLP, computer vision, and machine learning — thus, I had the chance to learn about other areas and topics. The team also gave a lot of feedback on others' projects, often leading to collaborations. From working on the paper, to learning how people work efficiently, and even how to balance life and research, the internship with the Salesforce Research team was a wonderful opportunity for me.


Learn more about opportunities to join our Salesforce Research Team as a Ph.D. intern here.