NLP MS Capstone Workshop Summer 2021

Welcome to the 2021 NLP Capstone Workshop!

About the Workshop

Natural Language Processing (NLP) is a rapidly growing field with applications in many of the technologies we use every day, from virtual assistants and smart speakers to autocorrect. In the NLP MS program at UCSC, students engage in a one-year intensive program that includes a 3-quarter capstone project. As part of a capstone team, students collaborate with experts from industry giants like IBM, Microsoft, Amazon, and Bloomberg (and more!) to address industry-relevant and research-focused topics in the NLP field.


On Friday, August 27th, 2021, NLP MS students will showcase their capstone projects at the inaugural NLP Capstone Workshop. Each team will be given time to present their work and answer questions from workshop attendees. The workshop will start at 9:30am PDT and will conclude at 2:30pm PDT.



To attend the workshop, please complete the following registration form by 5pm PDT on Thursday, August 26th:


After registering, you will receive a confirmation email containing information about joining the workshop through Zoom.


Workshop Schedule

9:30am - 9:45am

Welcome Remarks

9:45am - 10:15am

Capstone Team 1 Presentation

10:15am - 10:45am

Capstone Team 2 Presentation

10:45am - 11:00am


11:00am - 11:30am

Capstone Team 3 Presentation

11:30am - 12:00pm

Capstone Team 4 Presentation

12:00pm - 12:30pm

Lunch Break

12:30pm - 1:00pm

Capstone Team 5 Presentation

1:00pm - 1:30pm

Capstone Team 6 Presentation

1:30pm - 1:40pm


1:40pm - 2:25pm

Panel Discussion

2:25pm - 2:30pm

Closing Remarks

NOTE: All times are listed in Pacific Daylight Time


Contact Us

If you have any questions about the NLP Capstone Workshop, please email the NLP Program Team at

If you have any questions about the program, please visit the UCSC NLP MS website.


Preview the 2021 NLP Capstone Projects

Team 1: Amazon Alexa Prize Discourse Model & Coreference Resolution

Presenters: Jeshwanth Bheemanpally, Phillip Lee, Cecilia Li, Angela Ramirez, Eduardo Zamora

Project Mentor: Dr. Marilyn Walker & Dr. Adwait Ratnaparkhi, UCSC

Abstract: Our goal is to create a coreference resolution model that can identify entities being referred to in a social bot's user utterance using information found in our discourse model. We implement both a rule-based, neural, and ensemble model, and compare how they perform on a set of pronominal types: it, they, that, she, he, her, him, his, and hers. To enhance our models, we harness the information found in our discourse model that tracks and holds information of which entities are currently in focus in the conversation, independently of whether they are realized by a pronoun or a definite referring expression or a proper noun, and independently of whether they are introduced into the discourse by the user or the system. By doing so, we hope to improve the system's ability to acknowledge the entity being referred to by the user and respond with more salience.

Team 2: Identifying SRL Errors Using Weak Supervision

Presenters: Kit Lao, Alex Lue, Sam Shamsan

Project Mentor: Ishan Jindal & Frederick Reiss, IBM

Abstract: In datasets collected from real word data, noise and mislabelings in the corpora are almost always inevitable, and are especially prominent in large datasets. Performance of learned models from these datasets relies heavily on correctly labeled data to produce significant results. Our research project aims to investigate the noise rate in large semantic role labeling datasets, EWT and OntoNotes. Using a novel weak supervision mechanism for noise detection called Confident Learning, we propose an end-to-end system to characterize, identify, and correct the noisy labels in these datasets. Our goal is to generate corrected versions of SRL datasets that show performance improvement in state of the art SRL models trained on these datasets.

Team 3: Domain Adaptation for Covid-19 Question Answering

Presenters: Morgan Eidam, Adam Fidler, John Lara

Project Mentor: Arafat Sultan & Radu Florian, IBM; Vittorio Castelli, Amazon

Abstract: Covid-19 has affected the lives of billions globally. Experts were required to make significant decisions affecting hundreds of thousands at a time with limited data. It’s crucial for researchers and the general public to have a way to easily navigate existing research to identify effective treatments and management policies. QA systems can be an effective tool to leverage existing research to answer important Covid-19 questions. In this project, we utilize various domain adaptation tools such as several novel approaches to Gradual fine tuning as well as Adversarial training in order to improve the performance of our Covid QA model with very limited data.

Team 4: MKD-Ultra: Compressing Causal Language Models in Multiple Steps

Presenters: Mamon Alsalihy, Austin King, Nilay Patel

Project Mentor: Sarangarajan “Partha” Parthasarathy, Microsoft

Abstract: Modern deep neural networks have immensely powerful predictive power at the cost of equally great size and compute requirements. A recent surge of work has erupted in compressing these large models into smaller versions with similar predictive capabilities. Particularly, transformer language models such as BERT and XLNet are targeted due to their unparalleled performances across the board. In this paper, we present two compression techniques to train a small, autoregressive language model. First, we use a bidirectional teacher model to increase the information available. Second, we use a multi-step distillation approach to lessen the gap between large and small models.

Team 5: Evaluating Explanations

Presenters: Raghav Chaudhary, Christopher Garcia-Cordova, Kaleen Shrestha, Zachary Sweet

Project Mentor: Marina Danilevsky, IBM

Abstract: ​​We have reached a point in technological advances where outcomes of important decisions can be determined by the output of a machine​ ​learning model. This has motivated the development of methods that can​ ​generate explanations for these models. However, when it comes to​ ​black-box models, post-hoc explanations for such models need to be​ ​viewed critically. In this project, we developed a library to evaluate​ ​post-hoc explanations for machine learning models. Our library has​ ​multiple criteria for evaluating explanation quality, such as​ ​consistency, faithfulness, and model confidence indication. Our​ ​library also contains tools to generate perturbations of text based on​ ​adjuncts and synonyms in order to automatically create synthetic data​ ​for further analysis. We also show how effective and applicable our​ ​metrics are via novel user studies to understand the usefulness of​ ​metrics to an end-user. With this, we hope to help both researchers​ ​and end-users understand their models and explanations, as well as​ ​propose some guidance for evaluating explanation techniques.

Team 6: Information Extraction of Corporate Events from the Web

Presenters: Tianxiao Jiang, David Li, Liren Wu

Project Mentor: Yuval Marton & Swapnil Khedekar, Bloomberg

Abstract: Publicly traded companies are required to report earnings and hold certain types of events. During these events stocks are most volatile and draw huge interests from analysts, investors, shareholders and journalists. Therefore collecting this information is valuable, but the problem in extracting this information is that companies announce their events on the web in various ways and formats. We are building a pipeline that extracts key information, such as the event type, date, and time, from corporate event announcements on websites. This extracted data is then normalized so it can be easily used later. We are using a variety of natural language processing techniques ranging from rule-based or regular expression based approaches to Transformer-based techniques such as BERT and GPT to extract events from websites, normalize the data, and classify the type of these events.