My research lies at the intersections of machine learning, natural language processing, and social media. I focus on designing algorithms for learning semantics from large data for natural language understanding, and natural language generation in particular with stylistic variations. I recently received the NSF CRII Award, NSF CAREER Award, Criteo Faculty Research Award, CrowdFlower AI for Everyone Award, Best Paper Award at COLING'18, as well as research funds from DARPA and IARPA. I was a postdoctoral researcher at the University of Pennsylvania. I received my PhD in Computer Science from New York University, MS and BS from Tsinghua University.
I am a senior area chair for EMNLP 2022 (generation), NAACL 2022 (machine learning for NLP), 2021 (generation), and ACL 2020 (generation), and an area chair for EMNLP 2021 (computational social science), EMNLP 2020 (generation), AAAI 2020 (NLP), ACL 2019 (semantics), NAACL 2019 (generation), EMNLP 2018 (social media), COLING 2018 (semantics), EMNLP 2016 (generation), a workshop chair for ACL 2017, and the publicity chair for EMNLP 2019, NAACL 2018 and 2016. I also created a new undergraduate course on Social Media and Text Analytics.
I'm recruiting 1-2 new PhD students who will start in Fall 2023 (apply to PhD program and list me as a potential advisor). I also advise undergraduate and MS students with research thesis option (apply to MSCS; email me).
Spring 2021, work with my phd student Mounica Maddela, Sebastian Gehrmann (Google Research), Dhruv Kumar (Borealis AI), and others on the 💎GEM Benchmark for natural language generation, evaluation, and metrics.
Apr 29, invited talk at NYU NLP/Text-as-Data Speaker Series
April 2020, three long papers accepted to ACL 2020! We are releasing (1) new high-quality dataset and Transformer-based model for text simplification; (2) fine-grained named entity and code recognition for StackOverflow; (3) a unified span-based neural network framework and benchmark leaderboard for 10+ NLP tasks.
Many text-to-text generation problems can be thought of as sentential paraphrasing or monolingual machine translation. It faces an exponential search space larger than bilingual translation, but a much smaller optimal solution space due to specific task requirements. I am interested in a variety of generation problems, including style transfer [COLING'12] and stylistics in general (e.g., historic ↔ modern, non-standard ↔ standard [BUCC'13], feminine ↔ masculine [AAAI'16]). Our latest work focuses on controllable text generation [NAACL'21]. My work uncovered multiple serious problems in previous research (from 2010 to 2014) on text simplification [TACL'15] , designed a new tunable metric SARI [TACL'16] which is effective for evaluation and as a learning objective for training (now added to TensorFlow by the Google AI group), optimized syntax-based machine translation models [TACL'16], created pairwise neural ranking models to for lexical simplification [EMNLP'18], and studied document-level simplification [AAAI'20]. Our newest Transformer-based model initialized with BERT is the current state-of-the-art for automatic text simplification [ACL'20a].
Natural Language Understanding / Semantics
My approach to natural language understanding is learning and modeling paraphrases on a much larger scale and with a much broader range than previous work, essentially by developing more robust machine learning models and leveraging social media data. These paraphrase can enable natural language systems to handle errors (e.g., “everytime” ↔ “every time”), lexical variations (e.g., “oscar nom’d doc” ↔ “Oscar-nominated documentary”), rare words (e.g “NetsBulls series” ↔ “Nets and Bulls games”), and language shifts (e.g. “is bananas” ↔ “is great”). We designed a series of unsupervised and supervised learning approaches for paraphrase identification in social media data (also applicable to question/answer pairs for QA systems), ranging from neural network models [COLING'18][NAACL'18a] to multi-instance learning [TACL'14][EMNLP'16], and crowdsourcing large-scale datasets [SemEval'15][EMNLP'17].
Noisy User-generated Data / Social Media
For AI to truly understand human language and help people (e.g., instructing a robot), we ought to study the language people actually use in their daily life (e.g., posting on social media), besides the formally written texts that are well supported by existing NLP software. I thus focus on specially designed learning algorithms and the data for training these algorithms to develop tools to process and analyze noisy user-generated data. I have worked a lot with Twitter data [EMNLP'19][EMNLP'17][EMNLP'16][TACL'14], given its importance and large scale coverage. Social media also contains very diverse languages for studying stylistics and semantics, carrying information that is important for both people’s everyday lives and national security. In the past three years, with my students, I have expanded my scope to cover a wider range of user-generated data, including biology lab protocols [NAACL'18b], GitHub, and StackOverflow [ACL'20b].
PhD Thesis Committee: Sarah Wiegreffe (PhD @GaTech, 2022; interpretability/explainable AI - advisor: Mark Riedl) Yuval Pinter (PhD @GaTech, 2021; interpretability/semantics/morphology - advisor: Jacob Eisenstein) Sanqiang Zhao (PhD @UPitt, 2021; text simplification - advisor: Daqing He) Kai Cao (PhD @NYU, 2017; information extraction - advisor: Ralph Grishman) Maria Pershina (PhD @NYU, 2014; information extraction ACL'14 - advisor: Ralph Grishman)