UMass NLP Seminar: Spring 2025

This webpage has information about two NLP seminar activities for Spring 2025: (1) a list of NLP research talks by visitors, which are open to anyone at UMass or the Five Colleges; and (2) an accompanying graduate seminar, COMPSCI 692L, for reading and discussing related research papers.

NLP Research Talks

Talks are typically in either CS 151 or LGRC (likely A112). Please see the CICS Events webpage (linked from talk titles), and/or emails to seminars@cs.umass.edu, for the room number. Talks are open to anyone at UMass or the Five Colleges.

• Wed Feb 5, 12pm-1pm: Alexander Spangher: Planning in Creative Contexts
• Wed Feb 12, 12pm-1pm: Sarah Wiegreffe: Demystifying the Inner Workings of Language Models
• Wed Feb 19, 12pm-1pm: Weixin Liang, Societal Impact of Large Language Models
• Thu Feb 20, 4pm-5pm: Yizhong Wang, Building a Sustainable Data Foundation for AI
• Tue Feb 25, 12pm-1pm: Akari Asai, Beyond Scaling: Frontiers of Retrieval-Augmented Language Models
• Thu Feb 27, 12pm-1pm: Wenting Zhao
• Tue Mar 4, 12pm-1pm: TBA
• Tue Mar 11, 12pm-1pm: Nicholas Tomlin: Reasoning with Language Models. (Originally scheduled for Feb 10)

COMPSCI 692L: Discussion course for the NLP Seminar

Schedule

The course meets Wednesdays, 11:30am-12:45pm, room CS 142. On days with a talk, we end early at 11:50am to head over to the talk. See full course description further below.

• Class session, Feb 5 (short session): Brief introduction and small amount of paper discussion.
  • Nothing must be submitted before class, but there is an assignment. Read the following research paper before class, and bring a printout or electronic copy. Formulate at least two questions and be prepared to discuss in class: (1) a clarification question, about something you’d like to understand better about the paper; and (2) a question or comment about something you found interesting, would like to know more about, or an idea for future work.
    • Paper: Do LLMs Plan Like Human Writers? Comparing Journalist Coverage of Press Releases with LLMs. Spangher et al., EMNLP 2024.
    • No paper review is due for this paper.
    • But, a talk review is due tomorrow, 2/6 (see below)
• Class session, Feb 12: Discussion on Wiegreffe and Liang papers. Note talk on Feb 12.
  • Thus two paper reviews are due Feb 11. For each author, choose one of their papers as your target paper for reading and comments.

• Class session, Feb 19: Discussion on Wang and Akari papers. Note talks on Feb 19, Feb 20, and Feb 25.

• Class session, Feb 26: Discussion on Zhao papers. Note talks on Feb 27 and Mar 4.
  • See slack message on in-class presentations.
• Class session, Mar 5: Discussions on TBA and Tomlin papers.
  • In-class within-group presentations again. See slack message for details.

• Class session, Mar 12: TBD

• No class Mar 19. Enjoy spring break!

• One or two meetings after spring break, to be scheduled.

• Tentative additional talk: Fri Apr 11, 12pm-1:30: Laura Biester, in LGRC A104 (in the CSSI seminar series)

Course description

This 1-credit discussion course meets Wednesdays, 11:30am-12:45pm, room CS 142 during the semester. Instructor: Brendan O’Connor (email)

For spring 2025, the NLP seminar will focus on reading and discussing the research papers of the NLP research visitors. Visitor talks will often not be at the class discussion time, but enrolled students will be required to attend or do make-up activities. When the schedule allows, our goal is to read papers before the researcher’s visit, and be prepared to engage with questions and discussion.

Each week’s session will focus on one or a few papers, such as the work authored by an upcoming visitor. All participants will choose and read one or more papers for each week’s session, and write reactions, comments, and bring questions to class. Activities will primarily include participating in discussions.

Note the course schedule is front loaded for early in the semester. Due to the required attendance (or equivlant makeup) at research talks on other days, we will end earlier in the semester. The number of class sessions is chosen so that the total number of (class sessions + non-class talks) = the normal number of class sessions in a standard weekly 1-credit course.

Depending on the schedule, the later part of the semester may feature talks and discussions of other NLP research topics.

The course is recommended for graduate students with experience in NLP, such as one from the UMass NLP courses page. COMPSCI 485 and 685 may be the most relevant.

Course requirements

This course is structured about NLP research seminar visitors. For each visitor, students are required to read a paper, and go to the talk.

For each researcher, you’ll submit two reviews:

1. Paper review: Read in detail a research paper, and submit a summary and comments on this Google Form. This is due either the night before the class session where the author’s work is being discussed (as per the schedule). (No paper review is due for Spangher 2/5 since it’s the first class session.)
2. Talk review: Take some notes during the talk. Afterward, write and submit comments on this Google Form. This is due the day after the talk (but it might be easiest to submit it the same day, while it’s still fresh).

Comments should be written in your own words.

Attendance is required at class sessions and talks, except for conflicting classes (for the talks), or for extenuating circumstances (in general). Allowable extenuating circumstances are those according to official University policy, including personal, health, religious observances, official University travel, etc. To make up a missed class session or talk, read an additional paper and submit a review through the form, and clearly mark what it’s a makeup for.

Paper reading

How long does it take to read a paper? Anywhere from 10 minutes to 4 hours or more, depending on how much familiarity you have with a topic, background information, the paper’s length, nature of contribution, writing quality, and many other factors.

As stated in the introductory session, for this course please aim for at least an hour of reading per paper.

I (Brendan) usually advise reading via multiple passes through the paper. For example: (1) Read the abstract and quickly skim through the rest of the paper, to get a feel for what they did overall, and what the major resesarch questions are that they are intending to address. For typical experimentally-driven or empirical NLP papers, this includes seeing what all the tables and figures are. (2) Do a moderately deep read, taking notes on terminology, concepts, or various details you want to come back to. (3) Go deeper on parts that you’d like to understand better, possibly including looking up previous work and such that is necessary to understand the main things you want to learn.

The important thing is to NOT start with a deep read, which can bog you down and cause you to spend too much time on less important details.

Other course details

If you’d like to participate but cannot sign up in SPIRE, please submit your information on this form.

We’ll do communication through a Slack group.

Speakers from some previous offerings of the NLP Seminar: Fall 2024, Spring 2024, Fall 2023, Spring 2023.