| 佐藤 太清 | 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: Evaluating Japanese poem Tanka Automatically
abstract: Tanka is a Japanese poem expressed in form of 57577. Due to the lack of publicly available data on tanka, research hasn't been done on tanka. As for the study of poetry, only generation has been done, and no study has been done to evaluate poetry. Therefore, I collected tanka automatically from the site that created the ranking data for posting and evaluation of tanka, and I constructed a data set. Also, using the ranking, create an automatic evaluation system. language of the presentation: Japanese 発表題目: 短歌の自動評価 発表概要: 短歌は57577の定型で表現される日本の詩である。短歌の公開されているデータが存在しないため、これまで短歌に関する研究はほとんど行われてこなかった。詩の研究に関しても生成しか行われおらず、詩を評価する研究は行われていない。そこで短歌の投稿、評価を行ランキングデータ作成しているサイト上から自動的に短歌を収集し、データセットを構築した。またランキングであることを利用し、自動評価システムを作成する。 | |||||
| 佐藤 義貴 | 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: Grammatical error correction using pseudo data considering learner's first language
abstract: Grammatical error correction is a natural language processing task that takes a non-grammatical sentence as input and outputs the corrected sentence. Some of the errors of English learners are influenced by their native language. By training a correction model using the data of non-grammatical sentences written by a common native language speaker and the corrected sentences, the model becomes robust to the errors of the language speaker. However, at present, there is little data that can be used for model training, and if we focus only on English sentences written by learners with a specific language, it will be even less. The purpose of this study is to expand the training data and improve the performance of the correction model by generating pseudo data including errors that learners with a specific language make. language of the presentation: Japanese 発表題目: 英語学習者の母語を考慮した文法誤り訂正のための擬似データ生成 発表概要: 文法誤り訂正は入力が文法的に誤っている文，出力がその誤りを訂正した文であり，この入出力の系列変換を行う自然言語処理タスクである．英語学習者の誤りの中には母語の影響を受けているものがあり，共通の母語をもつ学習者によって書かれた誤り英文とその修正後の英文のデータを用いて訂正モデルを訓練することで，その母語をもつ学習者の誤りに対して頑健になることが知られている．しかし現状，モデルの訓練に使えるデータが少なく，特定の母語をもつ学習者が書いた英文のみに絞るとさらに少なくなる．本研究では，特定の母語をもつ学習者が起こすような誤りを含む擬似的なデータの生成を行うことで訓練データを拡張し，訂正モデルの性能を向上させることを目的とする． | |||||
| 澤田 悠冶 | 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: Coordination Identification without Labelled Data for Composite Named Entity Recognition
abstract: Named Entity Recognition is a fandamental task in natural language processing for detecting entity such as proper nouns and terms from a given text. Named entities in the biomedical domain appear frequently along with a coordinating conjunction and ellipses, e.g., ``alpha- and beta- globin''. This expression embeds two distinct entities while many named entity recognizer treat such an expression as a single large entity or as a fragment, thus there is a problem that recent recognizer is severe to correctly extract them. The goal of this study is to propose a simple method to retrieve named entities embedded in elliptical named entity involved in coordination. language of the presentation: Japanese 発表題目: 複合⽤語抽出のためのラベルなしデータを⽤いた並列名詞句の範囲同定 発表概要: 固有表現抽出は，入力された 文から固有名詞や専門用語といった固有表現を検出する自然言語処理タスクの一つで ある．バイオ分野における固有表現には，複数の固有表現が一部単語が省略されなが ら並列的に記述されるケースがあり，従来の抽出手法では対応が困難になる問題があ った．本研究の目的は，固有表現抽出器と組み合わせ可能な並列構造解析器を構築し，抽出器の性能を向上させることを目的とする． | |||||
| 山口 泰弘 | 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: Coreference-aware End-to-end Relation Extracion
abstract: Information extraction is one of the natural language processing task to extract structured information from documents. The task consists of some subtasks: named entity recognition, coreference resolution, and relation extraction, which often addressed independently as components of a pipeline. However, these subtasks are related each other, in particular, coreference is important for both named entity recognition and relation extraction. Additionally, pipeline systems are prone to error propagation. The goal of this study is to build coreference-aware end-to-end relation extraction system and improve the performance. language of the presentation: Japanese 発表題目: 共参照を考慮したEnd-to-end関係抽出 発表概要: 情報抽出とは文書中から構造化された情報を抽出する自然言語処理タスクである．情報抽出タスクは，固有表現認識，共参照解決，関係抽出など複数のサブタスクから構成される．多くの場合，これらのタスクは個別に処理され，パイプライン処理することで情報抽出システムを実現する．しかしながら，これらのタスクは互いに関係していて，特に共参照は固有表現認識，関係抽出の両方のタスクで重要な役割を果たす．加えて，パイプラインシステムは各タスクの間で誤りを伝搬することで性能の低下を引き起こす恐れがある．本研究では，共参照を考慮したEnd-to-endの関係抽出システムを構築し，関係抽出タスクにおける性能を向上させることを目的とする． | |||||
| 大木 麻里衣 | M, 2回目発表 | 数理情報学 | 池田 和司, | 松本 健一, | 吉本 潤一郎, | 久保 孝富（特任准教授）, | 福嶋 誠, | 日永田智絵 |
|
Title: Evaluation of Emotional Impact of Music Based on Heart Rate Variability Analysis
Abstract: Music affects our emotion and causes physiological changes. Previous studies have shown the effects of elements of music, e.g., music style and tempo, on emotion during listening to music. However, the contribution of other factors including lyrics, vocals, instruments are not yet clarified sufficiently. We are considering a hypothesis that lyrics induce the emotional change, and vocals and instruments modulate it. In this study, we aimed to verify this hypothesis by an experiment using questionnaires heart rate variability. We conducted the experiment to record the heart rate variability during listening music in different three presentation manner: lyrics only, vocal condition (lyrics and vocal), music (lyrics, vocal, and instruments). In this presentation, I will report the current results about emotional influences of the lyrics, vocals, and instruments. language of the presentation: Japanese | ||||||||
| 伊藤 健史 | D, 中間発表 | 数理情報学 | 池田 和司, | 松本 健一, | 吉本 潤一郎, | 久保 孝富, | 福嶋 誠, | 日永田智絵 |
|
title: Learning Semantics of Program with Function Call Contexts
abstract: Modern social infrastructure is built on diverse software. However, the complexity of software development processes costs programmers a heavy mental burden. Thus, there is a need for automated tools to help them comprehend existing programs and develop high-quality software. Recently, researchers have focused on applying machine learning techniques for program comprehension tasks, such as source code summarization or bug detection. Existing studies used single program functions independently as the unit of their model inputs, even though programmers implement nearly all practical software as large sets of functions calling one another. The problem here is that they failed to use the knowledge of the semantic contexts of a function defined by its callers and callees --we call them function call contexts-- to train their program comprehension models. The purpose of the present study is to show that incorporating function call contexts improves the performance of program comprehension models. To this end, we propose a program comprehension model using the "function call abstract syntax tree," or FCAST. FCAST is a graph data structure that can represent the semantics of program functions with their function call contexts. We present a graph neural network architecture for modeling program function semantics using FCAST, and compare its performance with existing models. language of the presentation: English | ||||||||
| 鈴木 文丈 | D, 中間発表 | 数理情報学 | 池田 和司☆, | 松本 健一, | 川鍋 一晃(客員）, | 森本 淳（客員）, | 福嶋 誠 | |
|
TItle:vmPFC and midbrain under stress regulate trait and state anxiety.
Abstract: In society, we are often confronted with stress-induced anxiety in exams and interviews, but humans exhibit the ability to suppress anxiety when under stress. Humans show the ability to down-regulate anxiety under stress, but anxiety can be either trait or state depending on the situation, and it is necessary to prove how the brain achieved the regulation of anxiety into traits and states under stress. Following acute social stress, the ventromedial prefrontal cortex (vmPFC) and midbrain were found to be activated for each emotional image or self-assessment of an image. The vmPFC also showed significant correlation with trait anxiety. Surprisingly, smaller changes in stress-induced anxiety states indicated increased larger functional connectivity between the vmPFC and the midbrain. It suggests that the vmPFC and midbrain cope with various types of anxiety under stress. Language of the presentation: Japanese | ||||||||