A5078390032- Natural Language Processing Techniques
- Topic Modeling
- Text Readability and Simplification
- Multimodal Machine Learning Applications
- Semantic Web and Ontologies
- Text and Document Classification Technologies
- Speech and dialogue systems
- Software Engineering Research
- Authorship Attribution and Profiling
- Biomedical Text Mining and Ontologies
- Explainable Artificial Intelligence (XAI)
- Speech Recognition and Synthesis
- Adversarial Robustness in Machine Learning
- Hate Speech and Cyberbullying Detection
- Software Testing and Debugging Techniques
- Advanced Text Analysis Techniques
- Software Reliability and Analysis Research
- Machine Learning and Algorithms
- Logic, Reasoning, and Knowledge
- Music and Audio Processing
- Team Dynamics and Performance
- Speech and Audio Processing
- Translation Studies and Practices
- Educational Systems and Policies
- Bayesian Modeling and Causal Inference
University of Maryland, College Park
2016-2025
University of Maryland, Baltimore
2024
University of Baltimore
2024
Microsoft (United States)
2021
University of Southern California
2020
Carnegie Mellon University
2017
The University of Tokyo
2017
Karlsruhe Institute of Technology
2017
Laboratoire d'Informatique de Paris-Nord
2017
Johns Hopkins University
2017
Milind Agarwal, Sweta Agrawal, Antonios Anastasopoulos, Luisa Bentivogli, Ondřej Bojar, Claudia Borg, Marine Carpuat, Roldano Cattoni, Mauro Cettolo, Mingda Chen, William Khalid Choukri, Alexandra Chronopoulou, Anna Currey, Thierry Declerck, Qianqian Dong, Kevin Duh, Yannick Estève, Marcello Federico, Souhir Gahbiche, Barry Haddow, Benjamin Hsu, Phu Mon Htut, Hirofumi Inaguma, Dávid Javorský, John Judge, Yasumasa Kano, Tom Ko, Rishu Kumar, Pengwei Li, Xutai Ma, Prashant Mathur, Evgeny...
Xuan Zhang, Pamela Shapiro, Gaurav Kumar, Paul McNamee, Marine Carpuat, Kevin Duh. Proceedings of the 2019 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). 2019.
Machine translation systems based on deep neural networks are expensive to train. Curriculum learning aims address this issue by choosing the order in which samples presented during training help train better models faster. We adopt a probabilistic view of curriculum learning, lets us flexibly evaluate impact curricula design, and perform an extensive exploration German-English task. Results show that it is possible improve convergence time at no loss quality. However, results highly...
Generative Artificial Intelligence (GenAI) systems are being increasingly deployed across all parts of industry and research settings. Developers end users interact with these through the use prompting or prompt engineering. While is a widespread highly researched concept, there exists conflicting terminology poor ontological understanding what constitutes due to area's nascency. This paper establishes structured prompts, by assembling taxonomy techniques analyzing their use. We present...
We directly investigate a subject of much recent debate: do word sense disambiguation models help statistical machine translation quality? present empirical results casting doubt on this common, but unproved, assumption. Using state-of-the-art Chinese model to choose candidates for typical IBM MT system, we find that does not yield significantly better quality than the system alone. Error analysis suggests several key factors behind surprising finding, including inherent limitations current...
Despite impressive progress in high-resource settings, Neural Machine Translation (NMT) still struggles low-resource and out-of-domain scenarios, often failing to match the quality of phrase-based translation. We propose a novel technique that combines back-translation multilingual NMT improve performance these difficult cases. Our trains single model for both directions language pair, allowing us back-translate source or target monolingual data without requiring an auxiliary model. then...
We develop two techniques for analyzing the effect of porting a machine translation system to new domain. One is macro-level analysis that measures how domain shift affects corpus-level evaluation; second micro-level word-level errors. apply these methods understand what happens when Parliament-trained phrase-based applied in four very different domains: news, medical texts, scientific articles and movie subtitles. present quantitative qualitative experiments highlight opportunities future...
Parallel corpora are often not as parallel one might assume: non-literal translations and noisy abound, even in curated routinely used for training evaluation. We use a cross-lingual textual entailment system to distinguish sentence pairs that meaning from those not, show filtering out divergent examples improves translation quality.
Stylistic variations of language, such as formality, carry speakers’ intention beyond literal meaning and should be conveyed adequately in translation. We propose to use lexical formality models control the level machine translation output. demonstrate effectiveness our approach empirical evaluations, measured by automatic metrics human assessments.
Abstract Neural sequence generation models are known to “hallucinate”, by producing outputs that unrelated the source text. These hallucinations potentially harmful, yet it remains unclear in what conditions they arise and how mitigate their impact. In this work, we first identify internal model symptoms of analyzing relative token contributions contrastive hallucinated vs. non-hallucinated generated via perturbations. We then show these reliable indicators natural hallucinations, using them...
This paper reports on the shared tasks organized by 21st IWSLT Conference. The address 7 scientific challenges in spoken language translation: simultaneous and offline translation, automatic subtitling dubbing, speech-to-speech dialect low-resource speech Indic languages. attracted 18 teams whose submissions are documented 26 system papers. growing interest towards translation is also witnessed constantly increasing number of task organizers contributors to overview paper, almost evenly...
We describe the system built by National Research Council Canada for "Discriminating between similar languages" (DSL) shared task.Our uses various statistical classifiers and makes predictions based on a two-stage process: we first predict language group, then discriminate languages or variants within group.Language groups are predicted using generative classifier with 99.99% accuracy five target groups.Within each group (except English), use voting combination of discriminative trained...
This task combines the labeling of multiword expressions and supersenses (coarse-grained classes) in an explicit, yet broad-coverage paradigm for lexical semantics.Nine systems participated; best scored 57.7% F 1 a multi-domain evaluation setting, indicating that remains largely unresolved.An error analysis reveals large number instances data set are either hard cases, which no get right, or easy all correctly solve.
Generating natural language requires conveying content in an appropriate style. We explore two related tasks on generating text of varying formality: monolingual formality transfer and formality-sensitive machine translation. propose to solve these jointly using multi-task learning, show that our models achieve state-of-the-art performance for are able perform translation without being explicitly trained style-annotated examples.
Sweta Agrawal, Marine Carpuat. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and 9th International Joint (EMNLP-IJCNLP). 2019.
Abstract We introduce an Edit-Based TransfOrmer with Repositioning (EDITOR), which makes sequence generation flexible by seamlessly allowing users to specify preferences in output lexical choice. Building on recent models for non-autoregressive (Gu et al., 2019), EDITOR generates new sequences iteratively editing hypotheses. It relies a novel reposition operation designed disentangle choice from word positioning decisions, while enabling efficient oracles imitation learning and parallel...
We revisit the one sense per discourse hypothesis of Gale et al. in context machine translation. Since a given can be lexicalized differently translation, do we observe translation discourse? Analysis manual translations reveals that still holds when using parallel text as annotation, thus confirming translational differences represent useful distinctions. Statistical Machine Translation (SMT) output showed despite ignoring document structure, is strongly supported part because low...
Yogarshi Vyas, Xing Niu, Marine Carpuat. Proceedings of the 2018 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers). 2018.