Dat Tien Nguyen

Alex Tamkin, Dat Nguyen, Salil Deshpande et al. · stanford

Models can fail in unpredictable ways during deployment due to task ambiguity, when multiple behaviors are consistent with the provided training data. An example is an object classifier trained on red squares and blue circles: when encountering blue squares, the intended behavior is undefined. We investigate whether pretrained models are better active learners, capable of disambiguating between the possible tasks a user may be trying to specify. Intriguingly, we find that better active learning is an emergent property of the pretraining process: pretrained models require up to 5 times fewer labels when using uncertainty-based active learning, while non-pretrained models see no or even negative benefit. We find these gains come from an ability to select examples with attributes that disambiguate the intended behavior, such as rare product categories or atypical backgrounds. These attributes are far more linearly separable in pretrained model's representation spaces vs non-pretrained models, suggesting a possible mechanism for this behavior.

Archana Warrier, Dat Nguyen, Michelangelo Naim et al. · cambridge

Model-learning agents should gather information to learn world models that support many downstream tasks and inferences, such as predicting unobserved states, estimating near- and far-term consequences of actions, planning action sequences, and detecting changes in dynamics. Current methods for learning and evaluating world models diverge from this goal: training and evaluation are anchored to next-frame prediction, and success is scored by reward maximization in the same environment. We propose WorldTest, a protocol to evaluate model-learning agents that separates reward-free interaction from a scored test phase in a different but related environment. WorldTest is open-ended$\unicode{x2014}$models should support many different tasks unknown ahead of time$\unicode{x2014}$and agnostic to model representation, allowing comparison across approaches. We instantiated WorldTest with AutumnBench, a suite of 43 interactive grid-world environments and 129 tasks across three families: masked-frame prediction, planning, and predicting changes to the causal dynamics. We compared 517 human participants and three frontier models on AutumnBench. We found that humans outperform the models, and scaling compute improves performance only in some environments but not others. WorldTest provides a novel template$\unicode{x2014}$reward-free exploration, derived tests, and behavior-based scoring$\unicode{x2014}$to evaluate what agents learn about environment dynamics, and AutumnBench exposes significant headroom in world-model learning.

Hieu Nguyen Van, Dat Nguyen, Phuong Minh Nguyen et al.

We introduce efficient deep learning-based methods for legal document processing including Legal Document Retrieval and Legal Question Answering tasks in the Automated Legal Question Answering Competition (ALQAC 2022). In this competition, we achieve 1\textsuperscript{st} place in the first task and 3\textsuperscript{rd} place in the second task. Our method is based on the XLM-RoBERTa model that is pre-trained from a large amount of unlabeled corpus before fine-tuning to the specific tasks. The experimental results showed that our method works well in legal retrieval information tasks with limited labeled data. Besides, this method can be applied to other information retrieval tasks in low-resource languages.

Masaharu Mizumoto, Dat Nguyen, Zhiheng Han et al.

Benchmark saturation and contamination undermine confidence in LLM evaluation. We present Nazonazo, a cost-effective and extensible benchmark built from Japanese children's riddles to test insight-based reasoning. Items are short (mostly one sentence), require no specialized domain knowledge, and can be generated at scale, enabling rapid refresh of blind sets when leakage is suspected. We evaluate 38 frontier models and 126 adults on 120 riddles. No model except for GPT-5 is comparable to human performance, which achieves a 52.9% mean accuracy. Model comparison on extended 201 items shows that reasoning models significantly outperform non-reasoning peers, while model size shows no reliable association with accuracy. Beyond aggregate accuracy, an informal candidate-tracking analysis of thought logs reveals many cases of verification failure: models often produce the correct solution among intermediate candidates yet fail to select it as the final answer, which we illustrate with representative examples observed in multiple models. Nazonazo thus offers a cost-effective, scalable, and easily renewable benchmark format that addresses the current evaluation crisis while also suggesting a recurrent meta-cognitive weakness, providing clear targets for future control and calibration methods.

Ines Chami, Albert Gu, Dat Nguyen et al.

This paper studies Principal Component Analysis (PCA) for data lying in hyperbolic spaces. Given directions, PCA relies on: (1) a parameterization of subspaces spanned by these directions, (2) a method of projection onto subspaces that preserves information in these directions, and (3) an objective to optimize, namely the variance explained by projections. We generalize each of these concepts to the hyperbolic space and propose HoroPCA, a method for hyperbolic dimensionality reduction. By focusing on the core problem of extracting principal directions, HoroPCA theoretically better preserves information in the original data such as distances, compared to previous generalizations of PCA. Empirically, we validate that HoroPCA outperforms existing dimensionality reduction methods, significantly reducing error in distance preservation. As a data whitening method, it improves downstream classification by up to 3.9% compared to methods that don't use whitening. Finally, we show that HoroPCA can be used to visualize hyperbolic data in two dimensions.

Dat Tien Nguyen, Shikhar Sharma, Hannes Schulz et al.

Understanding audio-visual content and the ability to have an informative conversation about it have both been challenging areas for intelligent systems. The Audio Visual Scene-aware Dialog (AVSD) challenge, organized as a track of the Dialog System Technology Challenge 7 (DSTC7), proposes a combined task, where a system has to answer questions pertaining to a video given a dialogue with previous question-answer pairs and the video itself. We propose for this task a hierarchical encoder-decoder model which computes a multi-modal embedding of the dialogue context. It first embeds the dialogue history using two LSTMs. We extract video and audio frames at regular intervals and compute semantic features using pre-trained I3D and VGGish models, respectively. Before summarizing both modalities into fixed-length vectors using LSTMs, we use FiLM blocks to condition them on the embeddings of the current question, which allows us to reduce the dimensionality considerably. Finally, we use an LSTM decoder that we train with scheduled sampling and evaluate using beam search. Compared to the modality-fusing baseline model released by the AVSD challenge organizers, our model achieves a relative improvements of more than 16%, scoring 0.36 BLEU-4 and more than 33%, scoring 0.997 CIDEr.

Tasnim Mohiuddin, Shafiq Joty, Dat Tien Nguyen

We propose a novel coherence model for written asynchronous conversations (e.g., forums, emails), and show its applications in coherence assessment and thread reconstruction tasks. We conduct our research in two steps. First, we propose improvements to the recently proposed neural entity grid model by lexicalizing its entity transitions. Then, we extend the model to asynchronous conversations by incorporating the underlying conversational structure in the entity grid representation and feature computation. Our model achieves state of the art results on standard coherence assessment tasks in monologue and conversations outperforming existing models. We also demonstrate its effectiveness in reconstructing thread structures.

Dat Tien Nguyen, Shafiq Joty, Basma El Amel Boussaha et al.

Discussion forums are an important source of information. They are often used to answer specific questions a user might have and to discover more about a topic of interest. Discussions in these forums may evolve in intricate ways, making it difficult for users to follow the flow of ideas. We propose a novel approach for automatically identifying the underlying thread structure of a forum discussion. Our approach is based on a neural model that computes coherence scores of possible reconstructions and then selects the highest scoring, i.e., the most coherent one. Preliminary experiments demonstrate promising results outperforming a number of strong baseline methods.

Dat Tien Nguyen, Firoj Alam, Ferda Ofli et al.

The extensive use of social media platforms, especially during disasters, creates unique opportunities for humanitarian organizations to gain situational awareness and launch relief operations accordingly. In addition to the textual content, people post overwhelming amounts of imagery data on social networks within minutes of a disaster hit. Studies point to the importance of this online imagery content for emergency response. Despite recent advances in the computer vision field, automatic processing of the crisis-related social media imagery data remains a challenging task. It is because a majority of which consists of redundant and irrelevant content. In this paper, we present an image processing pipeline that comprises de-duplication and relevancy filtering mechanisms to collect and filter social media image content in real-time during a crisis event. Results obtained from extensive experiments on real-world crisis datasets demonstrate the significance of the proposed pipeline for optimal utilization of both human and machine computing resources.

Dat Tien Nguyen, Shafiq Joty, Muhammad Imran et al.

During natural or man-made disasters, humanitarian response organizations look for useful information to support their decision-making processes. Social media platforms such as Twitter have been considered as a vital source of useful information for disaster response and management. Despite advances in natural language processing techniques, processing short and informal Twitter messages is a challenging task. In this paper, we propose to use Deep Neural Network (DNN) to address two types of information needs of response organizations: 1) identifying informative tweets and 2) classifying them into topical classes. DNNs use distributed representation of words and learn the representation as well as higher level features automatically for the classification task. We propose a new online algorithm based on stochastic gradient descent to train DNNs in an online fashion during disaster situations. We test our models using a crisis-related real-world Twitter dataset.

Dat Tien Nguyen, Kamela Ali Al Mannai, Shafiq Joty et al.

The role of social media, in particular microblogging platforms such as Twitter, as a conduit for actionable and tactical information during disasters is increasingly acknowledged. However, time-critical analysis of big crisis data on social media streams brings challenges to machine learning techniques, especially the ones that use supervised learning. The Scarcity of labeled data, particularly in the early hours of a crisis, delays the machine learning process. The current state-of-the-art classification methods require a significant amount of labeled data specific to a particular event for training plus a lot of feature engineering to achieve best results. In this work, we introduce neural network based classification methods for binary and multi-class tweet classification task. We show that neural network based models do not require any feature engineering and perform better than state-of-the-art methods. In the early hours of a disaster when no labeled data is available, our proposed method makes the best use of the out-of-event data and achieves good results.

Angeliki Lazaridou, Dat Tien Nguyen, Raffaella Bernardi et al.

We introduce language-driven image generation, the task of generating an image visualizing the semantic contents of a word embedding, e.g., given the word embedding of grasshopper, we generate a natural image of a grasshopper. We implement a simple method based on two mapping functions. The first takes as input a word embedding (as produced, e.g., by the word2vec toolkit) and maps it onto a high-level visual space (e.g., the space defined by one of the top layers of a Convolutional Neural Network). The second function maps this abstract visual representation to pixel space, in order to generate the target image. Several user studies suggest that the current system produces images that capture general visual properties of the concepts encoded in the word embedding, such as color or typical environment, and are sufficient to discriminate between general categories of objects.