My research focuses on making machine learning more useful, more
interpretable, and able to learn and interact from humans. This helps
users sift through decades of documents; discover when individuals lie,
reframe, or change the topic in a conversation; or to compete against
humans in games that are based in natural language.
@article{Guo:Zhang:Zhang:He:Zhang:Xie:Boyd-Graber-2022,
Author = {Fenfei Guo and Chen Zhang and Zhirui Zhang and Qixin He and Kejun Zhang and Jun Xie and Jordan Boyd-Graber},
Title = {Automatic Song Translation for Tonal Languages},
Journal = {Findings of the Association for Computational Linguistics},
Year = {2022},
Location = {Dublin},
Url = {http://umiacs.umd.edu/~jbg//docs/2022_acl_ast.pdf},
}
@article{Si:Zhao:Min:Boyd-Graber-2022,
Title = {Re-Examining Calibration: The Case of Question Answering},
Author = {Chenglei Si and Chen Zhao and Sewon Min and Jordan Boyd-Graber},
Journal = {Findings of Empirical Methods in Natural Language Processing},
Year = {2022},
Location = {Abu Dhabi},
Url = {http://umiacs.umd.edu/~jbg//docs/2022_emnlp_calibration.pdf},
}
Accessible Abstract: Calibration is an important problem in question answering: if a search engine or virtual assistant doesn't know the answer to a question, you should probably abstain from showing an answer (to save embarassment, as when Google said a horse had six legs). This EMNLP Findings paper shows that existing metrics to test how good a QA calibration push calibrated confidence toward the average confidence. We proposed an alternate method both for evaluation and to generate better calibration by looking how models change as they learn.
@article{He:Mao:Boyd-Graber-2022,
Title = {Cheater's Bowl: Human vs. Computer Search Strategies for Open-Domain QA},
Author = {Wanrong He and Andrew Mao and Jordan Boyd-Graber},
Journal = {Findings of Empirical Methods in Natural Language Processing},
Year = {2022},
Location = {Abu Dhabi},
Url = {http://umiacs.umd.edu/~jbg//docs/2022_emnlp_cheaters.pdf},
}
Accessible Abstract: When the Covid pandemic it, trivia games moved online. With it came cheating: people tried to quickly Google answers. This is bad for sportsmanship, but a good source of training data for helping teach computers how to find answers. We built an interface to harvest this training data from trivia players, fed these into retrieval-based QA systems, showing that these queries were better than the automatically generated queries used by the current state of the art.
@article{Jansen:Boyd-Graber-2022,
Title = {Picard understanding Darmok: A Dataset and Model for Metaphor-Rich Translation in a Constructed Language},
Author = {Peter Jansen and Jordan Boyd-Graber},
Booktitle = {Figurative Language Workshop 2022 @EMNLP},
Year = {2022},
Url = {https://arxiv.org/abs/2107.08146},
}
@inproceedings{Han:Carpuat:Boyd-Graber-2022,
Title = {SimQA: Detecting Simultaneous MT Errors through Word-by-Word Question Answering},
Author = {HyoJung Han and Marine Carpuat and Jordan Boyd-Graber},
Booktitle = {Empirical Methods in Natural Language Processing},
Year = {2022},
Location = {Abu Dhabi},
Url = {http://umiacs.umd.edu/~jbg//docs/2022_emnlp_simqa.pdf},
}
Accessible Abstract: Simultaneous interpretation (where a translation happens word by word before the source sentence is finished) is difficult to evaluate. We created a new evaluation framework based on the following scenario: imagine that you're thrown into a trivia gameshow where you don't know the language. Specifically, it's a game format where you interrupt the question word by word as soon as possible. Our hypothesis is that a monolingual player (who doesn't speak the source language) will be able to do better in the game with a better simultaneous translation system. In this 2022 EMNLP publication, we show that this evaluation is not only cheaper (you just need to translate the answer) but can also detect hallucinations and undertranslations better than existing evaluation methods.
@inproceedings{Feng:Boyd-Graber-2022,
Title = {Learning to Explain Selectively: A Case Study on Question Answering},
Author = {Shi Feng and Jordan Boyd-Graber},
Booktitle = {Empirical Methods in Natural Language Processing},
Year = {2022},
Location = {Abu Dhabi},
Url = {http://umiacs.umd.edu/~jbg//docs/2022_emnlp_augment.pdf},
}
Accessible Abstract: Many AI methods are a black box: input goes in, predictions come out. While there are many AI explanation tools that you can add to these predictions, how do you know if they are any good. In this work presented at EMNLP, if you put a human in front of a AI that's trying to answer questions, our hypothesis is that you can measure how good the underlying explanations are by how much the human's score goes up. This 2022 EMNLP publication not just measures which combinations of explanations are most effective for an individual. We use bandit exploration to quickly figure out what set of explanations best help a specific user.
@inproceedings{Yuan:Xia:May:Durme:Boyd-Graber-2022,
Author = {Michelle Yuan and Patrick Xia and Chandler May and Benjamin Van Durme and Jordan Boyd-Graber},
Title = {Adapting Coreference Resolution Models through Active Learning},
Booktitle = {Association for Computational Linguistics},
Year = {2022},
Location = {Dublin},
Url = {http://umiacs.umd.edu/~jbg//docs/2022_acl_alcoref.pdf},
}
Benjamin Börschinger, Jordan Boyd-Graber, Christian Buck, Jannis Bulian, Massimiliano Ciaramita, Michelle Chen Huebscher, Wojciech Gajewski, Yannic Kilcher, Rodrigo Nogueira, and Lierni Sestorain Saralegu. Meta Answering for Machine Reading. ArXiv, Preprint. [Preprint] [Bibtex]
@article{B\"orschinger:Boyd-Graber:Buck:Bulian:Ciaramita:Huebscher:Gajewski:Kilcher:Nogueira:Saralegu-Preprint,
Title = {Meta Answering for Machine Reading},
Author = {Benjamin B\"orschinger and Jordan Boyd-Graber and Christian Buck and Jannis Bulian and Massimiliano Ciaramita and Michelle Chen Huebscher and Wojciech Gajewski and Yannic Kilcher and Rodrigo Nogueira and Lierni Sestorain Saralegu},
Journal = {ArXiv},
Year = {Preprint},
Url = {https://arxiv.org/abs/1911.04156},
}
@article{Rodriguez:Feng:Iyyer:He:Boyd-Graber-Preprint,
Title = {Quizbowl: The Case for Incremental Question Answering},
Author = {Pedro Rodriguez and Shi Feng and Mohit Iyyer and He He and Jordan Boyd-Graber},
Journal = {ArXiv},
Year = {Preprint},
Url = {https://arxiv.org/abs/1904.04792},
}