A5056169983- Topic Modeling
- Natural Language Processing Techniques
- Speech and dialogue systems
- Multimodal Machine Learning Applications
- Speech Recognition and Synthesis
- Text Readability and Simplification
- Domain Adaptation and Few-Shot Learning
- Text and Document Classification Technologies
- Advanced Text Analysis Techniques
- Video Analysis and Summarization
- Generative Adversarial Networks and Image Synthesis
- Handwritten Text Recognition Techniques
Nanyang Technological University
2018-2021
Han Cheol Moon, Tasnim Mohiuddin, Shafiq Joty, Chi Xu. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and 9th International Joint (EMNLP-IJCNLP). 2019.
Tasnim Mohiuddin, Shafiq Joty. Proceedings of the 2019 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). 2019.
Most of the successful and predominant methods for Bilingual Lexicon Induction (BLI) are mapping-based, where a linear mapping function is learned with assumption that word embedding spaces different languages exhibit similar geometric structures (i.e. approximately isomorphic). However, several recent studies have criticized this simplified showing it does not hold in general even closely related languages. In work, we propose novel semi-supervised method to learn cross-lingual embeddings...
We present Fanar, a platform for Arabic-centric multimodal generative AI systems, that supports language, speech and image generation tasks. At the heart of Fanar are Star Prime, two highly capable Arabic Large Language Models (LLMs) best in class on well established benchmarks similar sized models. is 7B (billion) parameter model was trained from scratch nearly 1 trillion clean deduplicated Arabic, English Code tokens. Prime 9B continually Gemma-2 base same token set. Both models...
Crosslingual word embeddings learned from monolingual have a crucial role in many downstream tasks, ranging machine translation to transfer learning. Adversarial training has shown impressive success learning crosslingual and the associated task without any parallel data by mapping shared space. However, recent work superior performance for non-adversarial methods more challenging language pairs. In this article, we investigate adversarial autoencoder unsupervised propose two novel...
M Saiful Bari, Tasnim Mohiuddin, Shafiq Joty. Proceedings of the 59th Annual Meeting Association for Computational Linguistics and 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.
Neural Machine Translation (NMT) models are typically trained on heterogeneous data that concatenated and randomly shuffled. However, not all of the training equally useful to model. Curriculum aims present NMT in a meaningful order. In this work, we introduce two-stage framework for where fine-tune base model subsets data, selected by both deterministic scoring using pre-trained methods online considers prediction scores emerging Through comprehensive experiments six language pairs...
Participants in an asynchronous conversation (e.g., forum, e-mail) interact with each other at different times, performing certain communicative acts, called speech acts question, request). In this article, we propose a hybrid approach to act recognition conversations. Our works two main steps: long short-term memory recurrent neural network (LSTM-RNN) first encodes sentence separately into task-specific distributed representation, and is then used conditional random field (CRF) model...
The success of Neural Machine Translation (NMT) largely depends on the availability large bitext training corpora.Due to lack such corpora in low-resource language pairs, NMT systems often exhibit poor performance.Extra relevant monolingual data helps, but acquiring it could be quite expensive, especially for languages.Moreover, domain mismatch between (train/test) and might degrade performance.To alleviate issues, we propose AUGVIC, a novel augmentation framework which exploits vicinal...
Although coherence modeling has come a long way in developing novel models, their evaluation on downstream applications for which they are purportedly developed largely been neglected. With the advancements made by neural approaches such as machine translation (MT), summarization and dialog systems, need of these tasks is now more crucial than ever. However, models typically evaluated only synthetic tasks, may not be representative performance applications. To investigate how use cases, we...
Transfer learning has yielded state-of-the-art (SoTA) results in many supervised NLP tasks. However, annotated data for every target task language is rare, especially low-resource languages. We propose UXLA, a novel unsupervised augmentation framework zero-resource transfer scenarios. In particular, UXLA aims to solve cross-lingual adaptation problems from source distribution an unknown distribution, assuming no training label the language. At its core, performs simultaneous self-training...
Tasnim Mohiuddin, Thanh-Tung Nguyen, Shafiq Joty. Proceedings of the 2019 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). 2019.
Most of the successful and predominant methods for bilingual lexicon induction (BLI) are mapping-based, where a linear mapping function is learned with assumption that word embedding spaces different languages exhibit similar geometric structures (i.e., approximately isomorphic). However, several recent studies have criticized this simplified showing it does not hold in general even closely related languages. In work, we propose novel semi-supervised method to learn cross-lingual embeddings...
Recently, neural approaches to coherence modeling have achieved state-of-the-art results in several evaluation tasks. However, we show that most of these models often fail on harder tasks with more realistic application scenarios. In particular, the existing underperform require model be sensitive local contexts such as candidate ranking conversational dialogue and machine translation. this paper, propose a unified incorporates sentence grammar, inter-sentence relations, global patterns into...
Recent advancements in speech-language models have yielded significant improvements speech tokenization and synthesis. However, effectively mapping the complex, multidimensional attributes of into discrete tokens remains challenging. This process demands acoustic, semantic, contextual information for precise representations. Existing representations generally fall two categories: acoustic from audio codecs semantic self-supervised learning models. Although recent efforts unified improved...
This paper presents a comprehensive overview of the first edition Academic Essay Authenticity Challenge, organized as part GenAI Content Detection shared tasks collocated with COLING 2025. challenge focuses on detecting machine-generated vs. human-authored essays for academic purposes. The task is defined follows: "Given an essay, identify whether it generated by machine or authored human.'' involves two languages: English and Arabic. During evaluation phase, 25 teams submitted systems 21...
The ability to edit images in a realistic and visually appealing manner is fundamental requirement various computer vision applications. In this paper, we present ImEW, unified framework designed for solving image editing tasks. ImEW utilizes off-the-shelf foundation models address four essential tasks: object removal, translation, replacement, generative fill beyond the frame. These tasks are accomplished by leveraging capabilities of state-of-the-art models, namely Segment Anything Model,...
We propose a novel coherence model for written asynchronous conversations (e.g., forums, emails), and show its applications in assessment thread reconstruction tasks. conduct our research two steps. First, we improvements to the recently proposed neural entity grid by lexicalizing transitions. Then, extend incorporating underlying conversational structure representation feature computation. Our achieves state of art results on standard tasks monologue outperforming existing models. also...