コロキアムB発表

日時: 6月18日(金)3限(13:30~15:00)


会場: L1

司会: 花田 研太
NGUYEN VAN CAM D, 中間発表 コンピューティング・アーキテクチャ 中島 康彦, 林 優一, TRAN THI HONG, 張 任遠
Title: A Power-Efficient Domain Specific Architecture (DSA) for Skeleton-based Human Action Recognition

Abstract: In recent years, Skeleton-based Human Action Recognition (S-HAR) has achieved great interest. For real-time processing, S-HAR are usually accelerated by parallel processors such as graphic processing units (GPUs). However, due to substantial power consuming, they are applied limitedly for embedded systems. On this research, a power-efficient domain specific architecture for S-HAR is developed based on Field Programmable Gate Arrays (FPGA). In addition, the on-chip memory in current FPGAs are not sufficient to entirely store large-scale model. Therefore, a High Bandwidth Memory is used to expand bandwidth and speed up load/store manner from off-chip memory. After finishing development, we will evaluate about power consuming and performance by comparing with GPU and CPU.

Language of the presentation: English

 
WU MAN D, 中間発表 コンピューティング・アーキテクチャ 中島 康彦, 林 優一, TRAN THI HONG, 張 任遠
Title: Temporal-Spatial Elastic Neural Network Topology Towards Parallel Hardware Implementation
Abstract: Deep neural networks (DNNs) have demonstrated state-of-the-art performances in the broad field of artificial intelligence (AI) applications. Although the DNNs appear potentials comparable performance to human brains in some specific tasks, their hardware implementations still suffer from extremely high cost over processing time, power consumption, and resource explosion. A temporal-spatial elastic neural network (ENN) is developed and evolved that enables fully parallel and dynamic composition through multi-grained re-configurable (MGRA) to improve energy efficiency. Several evolutionary methods of ENN are developed, including I/O layer integration for deep reduction, skip connection for improving the performance, and accumulated spiking neural networks (SNNs) for further reducing power consumption. For a single task, compared with the LeNet5 baseline model (ANN and SNN), our work achieves the 90.86% parameter reduction on MNIST and NMIST dataset with a negligible loss of accuracy. Moreover, the preliminary experiment shows that the feasibility of processing multiple applications in fully parallel on FPGA.
Language of the presentation: English
 

会場: L2

司会: SOUFI Mazen
TRAN VAN HIEN D, 中間発表 自然言語処理学 渡辺 太郎, 中村 哲, 進藤 裕之
Title: Research on End-to-end Relation Extraction

Abstract: Extracting relation triplets is a critical and challenging task in Natural Language Processing (NLP). Given an unstructured text, it aims to extract pairs of entities with semantic relations, in the form of (head, relation, tail). It has attracted many NLP applications such as Information Extraction, Knowledge Base Population, and Question Answering.
In this research, we investigate end-to-end relation extraction systems which aim to extract such relation triplets in both supervised and unsupervised manners. In a supervised learning manner, we propose an improved decomposition strategy that helps to solve effectively the overlapping triplet problem, especially in the entity pair overlap (EPO) scenario, which is one of the biggest challenges of the joint entity and relation extraction. In an unsupervised learning manner, we build a CovRelex system, a scientific paper retrieval system targeting entities and relations via relation extraction on COVID-19 scientific papers. This system supports users to acquire knowledge across a huge number of COVID-19 scientific papers and can be accessed at this link.

Language of the presentation: English

 
新妻 巧朗 M, 2回目発表 自然言語処理学 渡辺 太郎, 中村 哲, 進藤 裕之
Title: Learning Method for Alleviating Social Bias

Abstract: The research area of natural language processing has made great progress by deep learning methods. However, these models are likely affected by datasets, so the models learning biased data can discriminate against some identities of social groups.
In this research, we tried to alleviate to let model learn social biases by devising learning technique. As a result, it was found that evaluation metrics of social bias were improved without degrading the classification performance. Furthermore, when we analyzed these results, we also confirmed that biases per token were decreasing.

Language of the presentation: Japanese
 
田口 智大 M, 1回目発表 自然言語処理学 渡辺 太郎, 中村 哲, 進藤 裕之
Title: Transliteration for Low-Resource Code-Switching Texts: Building an Automatic Cyrillic-to-Latin Converter for Tatar

Abstract: The Tatar language, mainly spoken in the Republic of Tatarstan, Russia, has two orthographies: Cyrillic and Latin. In my research, I aim to build a Cyrillic-to-Latin transliterator for Tatar based on subword-level language identification. The transliteration is a challenging task due to the following two reasons. First, because modern Tatar texts often contain Russian morphemes and words, a different set of transliteration rules needs to be applied to each morpheme depending on the language, which necessitates morpheme-level language identification. Second, the fact that Tatar is a low-resource language, with most of the texts in Cyrillic, makes it difficult to prepare a sufficient dataset. Given this situation, I propose a transliteration method based on subword-level language identification. A language classifier is trained with monolingual Tatar and Russian texts, and then different transliteration rules are applied in accord with the identified language. At this moment, the proposed method has already outscored other Tatar transliteration tools, and shows that some Russian loanwords were correctly predicted in the transliteration.

Language of the presentation: English
 
前田 拓哉 M, 1回目発表 ソーシャル・コンピューティング 荒牧 英治, 中村 哲, 若宮翔子, 矢田竣太郎
Title: Improving the Accountability and Responsibility of NLP Research on Social Issues

Abstract: Although NLP research on online abuse is often motivated by humanistic reasons, it is unclear to what extent these humanistic concerns manifest in research practices and orientations. Because researchers’ ideologies and values shape automated detection systems, leading to potentially adverse offline consequences, it is important to investigate not only the design of NLP research, but also the way in which this research is reported. This will help us to understand how researchers’ implicit theoretical approaches influence their engagement with this social problem. In this systematic review, we analyze 107 papers written about NLP-based approaches to online abuse published between January 2019 and January 2021. We focus particularly on how researchers conceptualize and contextualize the given problem, and to what extent they exercise transparency and accountability. Our findings reveal a tendency toward exclusion, reductivism, and dissociation in recent NLP research on online abuse, all of which present barriers to a methodologically humanistic approach. We then discuss the systemic bases for these obstacles and how they may extend beyond the context of NLP, and we suggest routes forward to ensure greater social responsibility in future research.

Language of the presentation: English