| Budi Irmawati | 1261204: D, 中間発表 | 松本 裕治, 松本 健一, 新保 仁, Kevin Duh |
|
title: *** Corpus based Error Identification System for Indonesian Second Language Learning ***
abstract: Nowadays, many research on grammatical error correction for second language learner emerge to use learner's corpus because mistakes created by learners and native speakers are unlikely. However, these corpora are limited for underdevelop language resources such as Indonesian. Considering the number of Indonesian learner's [http://www.ethnologue.com/language/ind], automatic error identification system for Indonesian second language learner is substantially necessary. Implies that the corpora are compulsory. Using lang-8 (Language Learning SNS) data crawled 2011, Indonesian learners' corpus is starting to be developed in one-word-to-one-word alignment. Several problems of native corrected sentence should be tackled such as a lot of spelling errors, ambiguity of corrections, long distance reordering, paraphrases corrections, as well as over corrected. We also computed the annotator agreement to show that native speakers have similar consideration whether a sentence has errors in such position while the error types and the corrections are subjective. Because Indonesian is freely word order language, error identification based on surface word, POS tag, and stem word by running a classifier showed low precision. It happens because the features share plenty of ambiguities. Therefore, syntactic information are tried to be involv to precisely identify learner's mistakes. As no data with syntactic information available publically, dependency annotation schema is tried to be proposed; resulted 81.18% precision with 39.39 % complete and 73% precision with 19.19% complete for UAS and LAS respectively. Moreover, some POS sequences are extracted based on syntactic information from large raw corpus as a gold grammar reference. Then, rule-based evaluation is proposed to identify word order and incorrect register. The next works are finding out ill-form by comparing nGram and dependency relation of learner and native speaker sentences. In general, I want to investigate if syntactic information helps the identification with some conditions such as with and without dependency relations as well as by manual and automatic dependency annotation. Furthermore, if it can be proved, it means that interactive system in which learners can write sentence and assign dependency relation will help system to identify error position and give feedback about the error type. language of the presentation: *** English *** | ||
| Sornchumni Nut | 1251210: M, 2回目発表 | 伊藤 実, 松本 健一, 楫 勇一, 関 浩之 |
|
title: Differentially Private Parameter Estimation for Hidden Markov Model
abstract: Privacy is one of the main concerns in database query processing and data mining. In this study, we investigate about the adaptation of differential privacy, which is newly proposed privacy definition and privacy preserving mechanism, to the vastly used stateful data mining method, hidden Markov model. We will show the differential privacy mechanism and then evaluate performance of differentiallu private version of hidden markov model. language of the presentation: English | ||
| Koganti Nishanth | 1251209: M, 2回目発表 | 池田 和司, 小笠原 司, 柴田 智広 |
|
Title: Real-time Estimation of Human-Cloth Topological Relationship using Depth Sensor for Robotic Clothing Assistance
Abstract: In the domain of Robotic Clothing Assistance, the human-cloth relationship needs to be estimated accurately to ensure efficient learning of motor skills. In this study, we propose a novel method for the real-time estimation of Human-Cloth relationship. This system relies on the use of a low cost depth sensor, which provides color and depth images without requiring an elaborate setup, making it suitable for real-world applications. We present an efficient algorithm to estimate the parameters that represent the topological relationship between human and the clothing article. We further evaluated the performance of our proposed method by applying it to real-time clothing assistance tasks and compared the estimates provided by our method with the ground truth. Language of the presentation: English | ||