신유현 교수

Data scarcity is one of the main obstacles of domain adaptation in spoken language understanding (SLU) due to the high cost of creating manually tagged SLU datasets. Recent works in neural text generative models, particularly latent variable models such as variational autoencoder (VAE), have shown promising results in regards to generating plausible and natural sentences. In this paper, we propose a novel generative architecture which leverages the generative power of latent variable models to j

Many approaches have been proposed to teach the basic concepts of big data and artificial intelligence to K-12 students based on block-based programming languages, such as Scratch. Using these approaches, young students can easily experience big data and artificial intelligence through a drag-and-drop approach. However, it remains difficult for them to perform more complex tasks, such as directly collecting data from the web or exploiting custom-made machine learning algorithms. In this paper, w

Scratch and App Inventor are two of the most widely used block-based programming languages for young students. These are educational languages which allow students to program easily by dragging and dropping their code blocks. One question that arises in relation to these educational languages is which of them would be more helpful in fostering computational thinking. It is difficult to answer this question because each language has its own advantages. In this paper, we propose a novel rubric bas

In this paper, we propose a Korean abstractive text summarization approach that uses a multi -encoder transformer. Recently, in many natural language processing (NLP) tasks, the use of the pre-trained language models (PLMs) for transfer learning has achieved remarkable performance. In particular, transformer-based models such as Bidirectional Encoder Representations from Transformers (BERT) are used for pre-training and applied to downstream tasks, showing state-of-the-art performance including

Slot filling must be trained using human-labeled data that are expensive and only a limited amount of labeled utterances are readily available for learning. Data generation methods can help increase the size of the dataset and make variations to the training dataset by means of emerging new instances. We propose a novel labeled utterance generation algorithm to augment training data. Our hypothesis is that words in an utterance can be separated into the two parts, namely, slot values that are in

We propose a novel unsupervised keyphrase extraction approach, called SAMRank, which uses only a self-attention map in a pre-trained language model (PLM) to determine the importance of phrases. Most recent approaches for unsupervised keyphrase extraction mainly utilize contextualized embeddings to capture semantic relevance between words, sentences, and documents. However, due to the anisotropic nature of contextual embeddings, these approaches may not be optimal for semantic similarity measurem

The existing block-based machine learning educational environments have a drawback in that they do not support model training based on large-scale data. This makes it difficult for young students to learn the importance of large amounts of data when creating machine learning models. In this paper, we present a novel programming environment in which students can easily train machine learning models based on large-scale data using a block-based programming language. We redefine the interfaces of e

This article introduces an approach that learns segment-level context for sequence labeling in natural language processing (NLP). Previous approaches limit their basic unit to a word for feature extraction because sequence labeling is a tokenlevel task in which labels are annotated word-by-word. However, the text segment is an ultimate unit for labeling, and we are easily able to obtain segment information from annotated labels in a IOB/IOBES format. Most neural sequence labeling models expand t

Blended learning has steadily gained in popularity at the higher levels of education. This marks a change in pedagogical approaches from one-directional instruction to an interactive and technology-aided class. However, to manage fluent in-class activities and proper data analysis, real-time and fine-grained data collection activities are still needed. We propose an approach which provides rich information about student activities and automates processes which are time-consuming and which otherw

Although Scratch is the most widely used block-based educational programming language, it is not easy for students to create various types of Scratch programs based on real-life data because it does not provide web scraping capabilities. In this paper, we present novel Scratch blocks for web scraping. Using these blocks, students can not only scrape the contents of HTML elements in a web page by using CSS selectors but also automate their keyboard and mouse in a number of ways, such as by using

In this paper, we present a novel approach to teach text processing for primary and secondary school students using a block-based programming language such as Scratch. Our main idea is to have students (1) build “basic building blocks” for text processing, and then (2) use them to create our example text processing applications. Here, we slightly modified Scratch to make it easy for students to create these basic building blocks. Also, because our example applications are built on the Data & Ana

Sentence representation models trained only on language could potentially suffer from the grounding problem. Recent work has shown promising results in improving the qualities of sentence representations by jointly training them with associated image features. However, the grounding capability is limited due to distant connection between input sentences and image features by the design of the architecture. In order to further close the gap, we propose applying self-attention mechanism to the sen

The importance of semantic similarity measures between sentences is increasingly growing in text mining, text clustering, and question answering. Many studies have focused on finding exact term matching to predict sentence similarity. In this paper, we present a method for measuring sematic similarity of sentences based on constructed synonymy graph to avoid considering just exactly matching terms. When we construct graph which has terms as nodes and synonymy relation as edges, we use WordNet an

In this paper, we present a novel approach to optical character recognition that incorporates various supplementary techniques, including the gradual detection of texts and gradual filtering of inaccurately recognized texts. To minimize false negatives, we attempt to detect all text by incrementally lowering the relevant thresholds. To mitigate false positives, we implement a novel filtering method that dynamically adjusts based on the confidence levels of recognized texts and their correspondin