A5021000040- Topic Modeling
- Natural Language Processing Techniques
- Multimodal Machine Learning Applications
- Personal Information Management and User Behavior
- Speech and dialogue systems
- Speech Recognition and Synthesis
- Information Retrieval and Search Behavior
- Advanced Text Analysis Techniques
- Software Engineering Research
- Mobile Crowdsensing and Crowdsourcing
- Web Data Mining and Analysis
- Recommender Systems and Techniques
- Spam and Phishing Detection
- Expert finding and Q&A systems
- Mental Health via Writing
- Domain Adaptation and Few-Shot Learning
- Semantic Web and Ontologies
- Misinformation and Its Impacts
- Machine Learning and Data Classification
- Text Readability and Simplification
- Software System Performance and Reliability
- Advanced Image and Video Retrieval Techniques
- Data Stream Mining Techniques
- AI in Service Interactions
- Adversarial Robustness in Machine Learning
Microsoft Research (United Kingdom)
2016-2023
Microsoft (United States)
2014-2023
Microsoft (Finland)
2023
Mohamed bin Zayed University of Artificial Intelligence
2023
Pennsylvania State University
2021-2022
Purdue University West Lafayette
2022
University of Maryland, College Park
2020-2021
Microsoft Research (India)
2021
Rice University
2020-2021
Yale University
2021
Leveraging weak or noisy supervision for building effective machine learning models has long been an important research problem. Its importance further increased recently due to the growing need large-scale datasets train deep models. Weak could originate from multiple sources including non-expert annotators automatic labeling based on heuristics user interaction signals. There is extensive amount of previous work focusing leveraging labels. Most notably, recent shown impressive gains by...
Ming Zhong, Da Yin, Tao Yu, Ahmad Zaidi, Mutethia Mutuma, Rahul Jha, Ahmed Hassan Awadallah, Asli Celikyilmaz, Yang Liu, Xipeng Qiu, Dragomir Radev. Proceedings of the 2021 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies. 2021.
Recent research has focused on enhancing the capability of smaller models through imitation learning, drawing outputs generated by large foundation (LFMs). A number issues impact quality these models, ranging from limited signals shallow LFM outputs; small scale homogeneous training data; and most notably a lack rigorous evaluation resulting in overestimating model's as they tend to learn imitate style, but not reasoning process LFMs. To address challenges, we develop Orca (We are working...
We introduce phi-3-mini, a 3.8 billion parameter language model trained on 3.3 trillion tokens, whose overall performance, as measured by both academic benchmarks and internal testing, rivals that of models such Mixtral 8x7B GPT-3.5 (e.g., phi-3-mini achieves 69% MMLU 8.38 MT-bench), despite being small enough to be deployed phone. The innovation lies entirely in our dataset for training, scaled-up version the one used phi-2, composed heavily filtered web data synthetic data. is also further...
Voice-activated intelligent assistants, such as Siri, Google Now, and Cortana, are prevalent on mobile devices. However, it is challenging to evaluate them due the varied evolving number of tasks supported, e.g., voice command, web search, chat. Since each task may have its own procedure a unique form correct answers, expensive individually. This paper first attempt solve this challenge. We develop consistent automatic approaches that can different in voice-activated assistants. use implicit...
Voice-controlled intelligent personal assistants, such as Cortana, Google Now, Siri and Alexa, are increasingly becoming a part of users' daily lives, especially on mobile devices. They introduce significant change in information access, not only by introducing voice control touch gestures but also enabling dialogues where the context is preserved. This raises need for evaluation their effectiveness assisting users with tasks. However, order to understand which type user interactions reflect...
Modern NLP applications have enjoyed a great boost utilizing neural networks models. Such deep models, however, are not applicable to most human languages due the lack of annotated training data for various tasks. Cross-lingual transfer learning (CLTL) is viable method building models low-resource target language by leveraging labeled from other (source) languages. In this work, we focus on multilingual setting where in multiple source leveraged further performance. Unlike existing methods...
There is a rapid growth in the use of voice-controlled intelligent personal assistants on mobile devices, such as Microsoft's Cortana, Google Now, and Apple's Siri. They significantly change way users interact with search systems, not only because voice control touch gestures, but also due to dialogue-style nature interactions their ability preserve context across different queries. Predicting success failure dialogues new problem, an important one for evaluating further improving...
Learning to capture text-table alignment is essential for tasks like text-to-SQL. A model needs correctly recognize natural language references columns and values ground them in the given database schema. In this paper, we present a novel weakly supervised Structure-Grounded pretraining framework (StruG) text-to-SQL that can effectively learn based on parallel corpus. We identify set of prediction tasks: column grounding, value grounding column-value mapping, leverage pretrain encoder....
Giannis Karamanolakis, Subhabrata Mukherjee, Guoqing Zheng, Ahmed Hassan Awadallah. Proceedings of the 2021 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies. 2021.
Orca 1 learns from rich signals, such as explanation traces, allowing it to outperform conventional instruction-tuned models on benchmarks like BigBench Hard and AGIEval. In 2, we continue exploring how improved training signals can enhance smaller LMs' reasoning abilities. Research small LMs has often relied imitation learning replicate the output of more capable models. We contend that excessive emphasis may restrict potential seek teach employ different solution strategies for tasks,...
We present methods to automatically identify and recommend sub-tasks help people explore accomplish complex search tasks. Although Web searchers often exhibit directed behaviors such as navigating a particular Website or locating item of information, many scenarios involve more tasks learning about new topic planning vacation. These multiple queries can span sessions. Current systems do not provide adequate support for tackling these Instead, they place most the burden on searcher...
Understanding and estimating satisfaction with search engines is an important aspect of evaluating retrieval performance. Research to date has modeled predicted on a binary scale, i.e., the searchers are either satisfied or dissatisfied their outcome. However, users' experience complex construct there different degrees satisfaction. As such, classification may be limiting. To best our knowledge, we first study problem understanding predicting graded (multi-level) We ex-amine sessions mined...
Web searchers sometimes struggle to find relevant information. Struggling leads frustrating and dissatisfying search experiences, even if ultimately meet their objectives. Better understanding of tasks where people is important in improving systems. We address this issue using a mixed methods study large-scale logs, crowd-sourced labeling, predictive modeling. analyze anonymized logs from the Microsoft Bing engine characterize aspects struggling searches better explain relationship between...
As the Web evolves towards a service-oriented architecture, application program interfaces (APIs) are becoming an increasingly important way to provide access data, services, and devices. We study problem of natural language interface APIs (NL2APIs), with focus on web for services. Such NL2APIs have many potential benefits, example, facilitating integration services into virtual assistants.
Multilingual representations embed words from many languages into a single semantic space such that with similar meanings are close to each other regardless of the language. These embeddings have been widely used in various settings, as cross-lingual transfer, where natural language processing (NLP) model trained on one is deployed another While transfer techniques powerful, they carry gender bias source target languages. In this paper, we study multilingual and how it affects learning for...
Neural sequence labeling is widely adopted for many Natural Language Processing (NLP) tasks, such as Named Entity Recognition (NER) and slot tagging dialog systems semantic parsing. Recent advances with large-scale pre-trained language models have shown remarkable success in these tasks when fine-tuned on large amounts of task-specific labeled data. However, obtaining training data not only costly, but also may be feasible sensitive user applications due to access privacy constraints. This...
Yusen Zhang, Ansong Ni, Ziming Mao, Chen Henry Wu, Chenguang Zhu, Budhaditya Deb, Ahmed Awadallah, Dragomir Radev, Rui Zhang. Proceedings of the 60th Annual Meeting Association for Computational Linguistics (Volume 1: Long Papers). 2022.
Ziming Mao, Chen Henry Wu, Ansong Ni, Yusen Zhang, Rui Tao Yu, Budhaditya Deb, Chenguang Zhu, Ahmed Awadallah, Dragomir Radev. Proceedings of the 60th Annual Meeting Association for Computational Linguistics (Volume 1: Long Papers). 2022.
Email is still among the most popular online activities. People spend a significant amount of time sending, reading and responding to email in order communicate with others, manage tasks archive personal information. Most previous research on based either relatively small data samples from user surveys interviews, or consumer accounts such as those Yahoo! Mail Gmail. Much less has been published how people interact enterprise even though it contains automatically generated commercial...
Deep and large pre-trained language models are the state-of-the-art for various natural processing tasks. However, huge size of these could be a deterrent to using them in practice. Some recent works use knowledge distillation compress into shallow ones. In this work we study with focus on multilingual Named Entity Recognition (NER). particular, several strategies propose stage-wise optimization scheme leveraging teacher internal representations, that is agnostic architecture, show it...
Social media has greatly enabled people to participate in online activities at an unprecedented rate. However, this unrestricted access also exacerbates the spread of misinformation and fake news which might cause confusion chaos unless being detected early for its mitigation. Given rapidly evolving nature events limited amount annotated data, state-of-the-art systems on detection face challenges due lack large numbers training instances that are hard come by detection. In work, we exploit...