A5078530959- Topic Modeling
- Natural Language Processing Techniques
- Multimodal Machine Learning Applications
- Speech Recognition and Synthesis
- Advanced Text Analysis Techniques
- Text Readability and Simplification
- Speech and dialogue systems
- Adversarial Robustness in Machine Learning
- Biomedical Text Mining and Ontologies
- Algorithms and Data Compression
- Hate Speech and Cyberbullying Detection
- Semantic Web and Ontologies
- Authorship Attribution and Profiling
- Sentiment Analysis and Opinion Mining
- Computational and Text Analysis Methods
- Domain Adaptation and Few-Shot Learning
- Ethics and Social Impacts of AI
- Gaussian Processes and Bayesian Inference
- Misinformation and Its Impacts
- Complex Network Analysis Techniques
- Human Mobility and Location-Based Analysis
- Machine Learning and Data Classification
- Geographic Information Systems Studies
- Text and Document Classification Technologies
- Software Engineering Research
The University of Melbourne
2015-2024
IT University of Copenhagen
2023
Tokyo Institute of Technology
2023
Administration for Community Living
2023
American Jewish Committee
2023
Australian National University
2023
Iran University of Science and Technology
2022
Information Technology University
2021
Amazon (Germany)
2021
Twitter (United States)
2021
We describe DyNet, a toolkit for implementing neural network models based on dynamic declaration of structure. In the static strategy that is used in toolkits like Theano, CNTK, and TensorFlow, user first defines computation graph (a symbolic representation computation), then examples are fed into an engine executes this computes its derivatives. DyNet's strategy, construction mostly transparent, being implicitly constructed by executing procedural code outputs, free to use different...
Many NLP applications can be framed as a graph-to-sequence learning problem. Previous work proposing neural architectures on obtained promising results compared to grammar-based approaches but still rely linearisation heuristics and/or standard recurrent networks achieve the best performance. In this propose new model that encodes full structural information contained in graph. Our architecture couples recently proposed Gated Graph Neural Networks with an input transformation allows nodes...
Long Duong, Trevor Cohn, Steven Bird, Paul Cook. Proceedings of the 53rd Annual Meeting Association for Computational Linguistics and 7th International Joint Conference on Natural Language Processing (Volume 2: Short Papers). 2015.
Active learning aims to select a small subset of data for annotation such that classifier learned on the is highly accurate. This usually done using heuristic selection methods, however effectiveness methods limited and moreover, performance heuristics varies between datasets. To address these shortcomings, we introduce novel formulation by reframing active as reinforcement problem explicitly policy, where policy takes role heuristic. Importantly, our method allows simulation one language be...
We present iterative back-translation, a method for generating increasingly better synthetic parallel data from monolingual to train neural machine translation systems. Our proposed is very simple yet effective and highly applicable in practice. demonstrate improvements quality both high low resourced scenarios, including the best reported BLEU scores WMT 2017 German↔English tasks.
In cross-lingual transfer, NLP models over one or more source languages are applied to a low-resource target language. While most prior work has used single model few carefully selected models, here we consider "massive" setting with many such models. This raises the problem of poor particularly from distant languages. We propose two techniques for modulating suitable zero-shot few-shot learning, respectively. Evaluating on named entity recognition, show that our much effective than strong...
Neural encoder-decoder models of machine translation have achieved impressive results, rivalling traditional models. However their modelling formulation is overly simplistic, and omits several key inductive biases built into In this paper we extend the attentional neural model to include structural from word based alignment models, including positional bias, Markov conditioning, fertility agreement over directions. We show improvements a baseline standard phrase-based language pairs,...
Automatic metrics are fundamental for the development and evaluation of machine translation systems. Judging whether, to what extent, automatic concur with gold standard human is not a straightforward problem. We show that current methods judging highly sensitive translations used assessment, particularly presence outliers, which often leads falsely confident conclusions about metric's efficacy. Finally, we turn pairwise system ranking, developing method thresholding performance improvement...
Social media user geolocation is vital to many applications such as event detection. In this paper, we propose GCN, a multiview model based on Graph Convolutional Networks, that uses both text and network context. We compare GCN the state-of-the-art, two baselines propose, show our achieves or competitive with state-of-the-art over three benchmark datasets when sufficient supervision available. also evaluate under minimal scenario, it outperforms baselines. find highway gates are essential...
In this paper we generalise the sentence compression task.Rather than simply shorten a by deleting words or constituents, as in previous work, rewrite it using additional operations such substitution, reordering, and insertion.We present new corpus that is suited to our task discriminative tree-totree transduction model can naturally account for structural lexical mismatches.The incorporates novel grammar extraction method, uses language coherent output, be easily tuned wide range of...
Long Duong, Antonios Anastasopoulos, David Chiang, Steven Bird, Trevor Cohn. Proceedings of the 2016 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies. 2016.
Written text often provides sufficient clues to identify the author, their gender, age, and other important attributes. Consequently, authorship of training evaluation corpora can have unforeseen impacts, including differing model performance for different user groups, as well privacy implications. In this paper, we propose an approach explicitly obscure author characteristics at time, such that representations learned are invariant these Evaluating on two tasks, show leads increased in...
Oliver Adams, Adam Makarucha, Graham Neubig, Steven Bird, Trevor Cohn. Proceedings of the 15th Conference European Chapter Association for Computational Linguistics: Volume 1, Long Papers. 2017.
Michal Lukasik, P. K. Srijith, Duy Vu, Kalina Bontcheva, Arkaitz Zubiaga, Trevor Cohn. Proceedings of the 54th Annual Meeting Association for Computational Linguistics (Volume 2: Short Papers). 2016.
Recent work has shown that simple vector subtraction over word embeddings is surprisingly effective at capturing different lexical relations, despite lacking explicit supervision.Prior evaluated this intriguing result using a analogy prediction formulation and hand-selected but the generality of finding broader range relation types learning settings not been evaluated.In paper, we carry out such an evaluation in two settings:(1) spectral clustering to induce (2) supervised classify...
This paper presents an approach to geolocating users of online social networks, based solely on their 'friendship' connections. We observe that interact more regularly with those closer themselves and hypothesise that, in many cases, a person's network is sufficient reveal location.
Crosslingual word embeddings represent lexical items from different languages in the same vector space, enabling transfer of NLP tools. However, previous attempts had expensive resource requirements, difficulty incorporating monolingual data or were unable to handle polysemy. We address these drawbacks our method which takes advantage a high coverage dictionary an EM style training algorithm over corpora two languages. Our model achieves state-of-the-art performance on bilingual lexicon...
Cross-lingual model transfer is a compelling and popular method for predicting annotations in low-resource language, whereby parallel corpora provide bridge to high-resource its associated annotated corpora. However, data not readily available many languages, limiting the applicability of these approaches. We address drawbacks our framework which takes advantage cross-lingual word embeddings trained solely on high coverage dictionary. propose novel neural network joint training from both...
The emergence of online social networks (OSNs) and the accompanying availability large amounts data, pose a number new natural language processing (NLP) computational challenges. Data from OSNs is different to data traditional sources (e.g. newswire). texts are short, noisy conversational. Another important issue that occurs in real-time streams, needing immediate analysis grounded time context. In this paper we describe open-source framework for efficient text streaming OSN (available at...
In this paper we present a novel approach for inducing word alignments from sentence aligned data. We use Conditional Random Field (CRF), discriminative model, which is estimated on small supervised training set. The CRF conditioned both the source and target texts, thus allows of arbitrary overlapping features over these Moreover, has efficient decoding processes find globally optimal solutions.We apply alignment model to French-English Romanian-English language pairs. show how large number...