| 氏家 翔吾 | M, 2回目発表 | ソーシャル・コンピューティング | 荒牧 英治, | 渡辺 太郎, | 田中 宏季, | 若宮 翔子 |
|
title: Biomedical Entity Linking via Contrastive Context Matching
abstract: We introduce BioCoM, a contrastive learning framework for biomedical entity linking that uses only two resources: a small-sized dictionary and a large number of raw biomedical articles. Specifically, we build the training instances from raw PubMed articles by dictionary matching and use them to train a context-aware entity linking model with contrastive learning. We predict the normalized biomedical entity at inference time through a nearest-neighbor search. Results found that BioCoM substantially outperforms state-of-the-art models, especially in low-resource settings, by effectively using the context of the entities. language of the presentation: Japanese 発表題目: 文脈化埋め込み表現を用いた対照学習による病名正規化 発表概要: 病名の表記揺れ解消（本稿では病名正規化と呼ぶ）は，医学論文や診療録などのテキスト（医療文書）の解析における重要な要素技術の一つである。従来の正規化手法は、大量の学習データや大規模な辞書などの言語資源を利用して高い精度を達成してきたが、そのような言語資源は日本語をはじめとする非英語言語では整備されていない．本研究では，大規模な医学文書に対して自動でコーパスを作成し，それを用いて文脈化埋め込み表現を学習することで，低リソース時においても精度よく正規化可能なモデルを提案する．実験の結果，辞書サイズが小さい場合に精度が向上していることがわかった． | ||||||
| AZUAJE SUAREZ GAMAR IVAN | M, 2回目発表 | ソーシャル・コンピューティング | 荒牧 英治, | 渡辺 太郎, | 須藤 克仁, | 若宮 翔子 |
|
title: Multimodal visualization of Lyrics
abstract: With the advent of social networks, there has been an increase in the number of aspiring artists sharing their music online. As such, there is an interest in creating suitable album arts to attract new listeners. However, not all artists have the skills or funds to create album art for their tracks. We present a preliminary application that allows artists to create album arts in a multimodal approach: using textual information from lyrics to generate images and aural information from song files to find a suitable style. language of the presentation: English | ||||||
| GAO ZHIWEI | M, 2回目発表 | ソーシャル・コンピューティング | 荒牧 英治, | 渡辺 太郎, | 田中 宏季, | 若宮 翔子 |
|
title: Offensive Language Detection on Live Streaming Chats Using Active Learning
abstract: Given the growing popularity of online games and eSports, the young generation is increasingly enjoying its video live streaming service. Streaming channels are usually combined with chat rooms, where offensive conversations often appear against the streamer or audience. Such offensive language may cause many severe impacts on a victim’s life and even lead to teen suicide. This research aims to detect offensive language appearing in live streaming chats. Focusing on Twitch, the most popular live streaming platform, we created a dataset containing ten games with a total of 10,000 data for the task of detecting offensive language. In particular, we used active learning to address the scarcity of well-labeled data on emerging platforms. In addition, we proposed to use dataset merging to solve the problem of limited model performance when selecting data from only one domain. Our results showed that active learning is able to detecting offensive language on live streaming platforms when the data is limited. language of the presentation: English | ||||||
| TAZEBE DANAY TASEW | M, 1回目発表 | ソーシャル・コンピューティング | 荒牧 英治, | 渡辺 太郎, | 若宮 翔子, | Md. Altaf-ul-Amin |
|
title: Diagnosis Prediction Using Electronic Health Records
abstract: Electronic Health Records (EHRs) are a longitudinal record of a patient's medical history. The data in EHR are represented as a sequence of temporally ordered visits, where each visit contains a set of medical codes within it. Diagnosis prediction using EHR data is a core research area in medical informatics and the development of personalized healthcare systems. Our goal is to predict a patient's diagnosis of next visit (i.e., future diagnosis) by using the previous sequential diagnosis data assigned through the past sequence of visits. We plan to approach diagnosis prediction by incorporating the overall health characterization of a patient with the local sequence diagnosis information to capture contextual dependency and temporal relationships within the patient’s timeline. language of the presentation: English | ||||||
| TAN RENZO ROEL PEREZ | D, 中間発表 | 数理情報学 | 池田 和司, | 笠原 正治, | 吉本 潤一郎, | 福嶋 誠, | 日永田智絵 |
|
Title: On the Utility of the Zero-Suppressed Binary Decision Diagram
Abstract: Decision-diagram-based solutions for discrete optimization have been persistently studied. Among these is the use of the zero-suppressed binary decision diagram, a compact graph-based representation for a specified family of sets. Such a diagram may work out problems in combinatorics by efficient enumeration. A wide range of combinatorial problems in operations research falls under arc routing problems, a domain which focuses on arc or edge features rather than node or vertex attributes. The generalized directed rural postman problem is a generic type of problem with the goal of finding the shortest path utilizing at least an edge from each category in a graph with labeled edges. Another is the undirected rural postman problem, a well-known problem in arc routing that seeks to determine a minimum cost walk that traverses a certain set of required edges on a given graph. The problems, arising in numerous real-world applications, are nondeterministic polynomial-time hard. In brief, an extension to the frontier-based search approach for zero-suppressed binary decision diagram construction is proposed. The modification allows for the inclusion of a class-determined constraint in formulation. Variations of the generalized directed rural postman problem, proven to be nondeterministic polynomial-time hard, are solved on some rapid transit systems as illustration. Results are juxtaposed against standard integer programming in establishing the relative superiority of the new technique. A solution to the undirected rural postman problem based on the zero-suppressed binary decision diagram is also presented. Through an extension to the frontier-based search method of diagram construction, the approach solves the problem by efficient enumeration, producing all feasible routes in addition to the optimal route. Instances of the problem put forward in literature are then solved as benchmark for the decision-diagram-based solution. As reasonable time is consumed, the method also proves to be a practicable candidate in solving the problem. Language of the presentation: English | |||||||
| DODU ALBRECHT YORDANUS ERWIN | D, 中間発表 | 生体医用画像 | 佐藤 嘉伸, | 金谷 重彦, | 大竹 義人, | Soufi Mazen, | 上村 圭亮 |
|
Title: Comparing the Segmentation Accuracy of Diseased and Unaffected Leg Sides Using CT Images of the Hip and Thigh Bones
abstract: Bone segmentation in CT images is necessary for several applications, such as hip surgery. The accuracy automated bone segmentation in the hips and thighs can be affected by the variability in structure and density between diseased and unaffected leg sides. This occurs because the resulting CT image will be imperfect due to the bone damage, making it difficult for the network to segment automatically. The goal of this study is to compare the segmentation accuracy of the diseased and unaffected leg sides of CT images of the pelvis, sacrum, femur and patella bones and the sacral canal in the hips and thighs using a state-of-the-art automated segmentation tool (i.e. Bayesian U-Net). The dataset consisted of CT images of 20 hip-diseased females. The accuracy of segmentation is assessed using the Dice Coefficient (DC) and Average Surface Distance (ASD). language of the presentation: English | |||||||