Yoonseok Heo

Video captioning is a task of describing the visual scene of a given video in natural language. There have been several lines of research focused on developing large-scale models in a transfer learning paradigm, with major challenge being the tradeoff between scalability and performance in limited environments. To address this problem, we propose a simple yet effective encoder–decoder-based video captioning model integrating transformers and CLIP, both of which are widely adopted in the vision and language domains, together with appropriate temporal feature embedding modules. Taking this proposal a step further, we also address the challenge of human-interactive video captioning, where the captions are tailored to specific information desired by humans. To design a human-interactive environment, we assume that a human offers an object or action in the video as a short prompt; in turn, the system then provides a detailed explanation regarding the prompt. We embed human prompts within an LSTM-based prompt encoder and leverage soft prompting to tune the model effectively. We extensively evaluated our model on benchmark datasets, demonstrating comparable results, particularly on the MSR-VTT dataset, where we achieve state-of-the-art performance with 4% improvement. In addition, we also show potential for human-interactive video captioning through quantitative and qualitative analysis.

With the development of multimedia systems in wireless environments, the rising need for artificial intelligence is to design a system that can properly communicate with humans with a comprehensive understanding of various types of information in a human-like manner. Therefore, this paper addresses an audio-visual scene-aware dialog system that can communicate with users about audio-visual scenes. It is essential to understand not only visual and textual information but also audio information in a comprehensive way. Despite the substantial progress in multimodal representation learning with language and visual modalities, there are still two caveats: ineffective use of auditory information and the lack of interpretability of the deep learning systems’ reasoning. To address these issues, we propose a novel audio-visual scene-aware dialog system that utilizes a set of explicit information from each modality as a form of natural language, which can be fused into a language model in a natural way. It leverages a transformer-based decoder to generate a coherent and correct response based on multimodal knowledge in a multitask learning setting. In addition, we also address the way of interpreting the model with a response-driven temporal moment localization method to verify how the system generates the response. The system itself provides the user with the evidence referred to in the system response process as a form of the timestamp of the scene. We show the superiority of the proposed model in all quantitative and qualitative measurements compared to the baseline. In particular, the proposed model achieved robust performance even in environments using all three modalities, including audio. We also conducted extensive experiments to investigate the proposed model. In addition, we obtained state-of-the-art performance in the system response reasoning task.

A rapidly expanding multimedia environment in recent years has led to an explosive increase in demand for multimodality that can communicate with humans in various ways. Even though the convergence of vision and language intelligence has shed light on the remarkable success over the last few years, there is still a caveat: it is unknown whether they truly understand the semantics of the image. More specifically, how they correctly capture relationships between objects represented within the image is still regarded as a black box. In order to testify whether such relationships are well understood, this work mainly focuses on the Graph-structured visual Question Answering (GQA) task which evaluates the understanding of an image by reasoning a scene graph describing the structural characteristics of an image in the form of natural language together with the image. Unlike the existing approaches that have been accompanied by an additional encoder for scene graphs, we propose a simple yet effective framework using pre-trained multimodal transformers for scene graph reasoning. Inspired by the fact that a scene graph can be regarded as a set of sentences describing two related objects with a relationship, we fuse them into the framework separately from the question. In addition, we propose a multi-task learning method that utilizes evaluating the grammatical validity of questions as an auxiliary task to better understand a question with complex structures. This utilizes the semantic role labels of the question to randomly shuffle the sentence structure of the question. We have conducted extensive experiments to evaluate the effectiveness in terms of task capabilities, ablation studies, and generalization.

An automatic story generation system continuously generates stories with a natural plot. The major challenge of automatic story generation is to maintain coherence between consecutive generated stories without the need for human intervention. To address this, we propose a novel multi-modal story generation framework that includes automated storyline decision-making capabilities. Our framework consists of three independent models: a transformer encoder-based storyline guidance model, which predicts a storyline using a multiple-choice question-answering problem; a transformer decoder-based story generation model that creates a story that describes the storyline determined by the guidance model; and a diffusion-based story visualization model that generates a representative image visually describing a scene to help readers better understand the story flow. Our proposed framework was extensively evaluated through both automatic and human evaluations, which demonstrate that our model outperforms the previous approach, suggesting the effectiveness of our storyline guidance model in making proper plans.

Word sense disambiguation (WSD) is a task of determining a reasonable sense of a word in a particular context. Although recent studies have demonstrated some progress in the advancement of neural language models, the scope of research is still such that the senses of several words can only be determined in a few domains. Therefore, it is necessary to move toward developing a highly scalable process that can address a lot of senses occurring in various domains. This paper introduces a new large WSD dataset that is automatically constructed from the Oxford Dictionary, which is widely used as a standard source for the meaning of words. We propose a new WSD model that individually determines the sense of the word in accordance with its part of speech in the context. In addition, we introduce a hybrid sense prediction method that separately classifies the less frequently used senses for achieving a reasonable performance. We have conducted comparative experiments to demonstrate that the proposed method is more reliable compared with the baseline approaches. Also, we investigated the adaptation of the method to a realistic environment with the use of news articles.

Machine translation refers to a fully automated process that translates a user's input text into a target language. To improve the accuracy of machine translation, studies usually exploit not only the input text itself but also various background knowledge related to the text, such as visual information or prior knowledge. Herein, in this paper, we propose a multimodal neural machine translation system that uses both texts and their related images to translate Korean image captions into English. The data in the experiment is a set of unlabeled images only containing bilingual captions. To train the system with a supervised learning approach, we propose a weak-labeling method that selects a keyword from an image caption using feature selection methods. The keywords are used to roughly determine an image label. We also introduce an improved feature selection method using sentence clustering to select keywords that reflect the characteristics of the image captions more accurately. We found that our multimodal system achieves an improved performance compared to a text-only neural machine translation system (baseline). Furthermore, the additional images have positive impacts on addressing the issue of under-translation, where some words in a source sentence are falsely translated or not translated at all.