ゼミナール発表

日時: 10月05日(Wed)3限


会場: L1

司会: 小町 守
"TORRES RODRIGUEZ,RAFAEL ANTONIO" D 音情報処理学 鹿野 清宏 松本 裕治 猿渡 洋 川波 弘道
発表題目:Stacked Generalization for Topic Classification of Spoken Inquiries
発表概要:In this work, we address the topic classification of spoken inquiries in Japanese received by a speech-oriented guidance system operating in a real environment. For this, we propose a stacked generalization scheme that uses predictions of support vector machine (SVM) with RBF kernel, prefixspan boosting (pboost) and maximum entropy (ME), as input for a second-level classification using SVM with RBF kernel. To deal with the shortness of the utterances, we also proposed to use characters n-grams as features instead of words. We are interested in evaluating the strength of the proposed scheme against automatic speech recognition (ASR) errors and the sparseness of the features present in spontaneous speech. Experimental results show that the proposed stacked generalization scheme improves the classification performance of spoken inquiries in comparison to the individual performance of the first-level classifiers.
 
羅 彦彦 D 自然言語処理学 松本 裕治 鹿野 清宏 新保 仁 浅原 正幸
発表題目:Dual decomposition method for Chinese semantic role labeling
発表概要:Dependency structure based pipeline semantic role labeling(SRL) systems performed better in previous work. Unfortunately, Chinese dependency parsing is still a certral bottleneck that serverly limits the perfomance of Chinese SRL. On the contray, Chinese shallow parsing has gained a promising development; on the other hand, pipeline SRL systems are inevitably influenced by the cascading errors that are introduced in early stages of the pipeline. In this research, two joint semantic parsing models are proposed dependently to overcome these problems. One model utilizes the shallow parsing information to joint implement dependency parsing and semantic role labeling. This model can alleviate the dependency parsing result influnce for semantic role labeling, but also make two parsings obtain each other's information for common improvements. Dependency structure contains the information that cannot exist in shallow parsing while these information is very helpful for semantic parsing. It's still necessary to build such a joint semantic role labeling system which can fully utilize dependency parsing result. Inference can be performed via dual decomposition which reuses the inference algorithms of the two joint models. Therefore, the dual decomposition method comprises the merits of both joint models without losing efficiency.
 
八木 浩介 M 音情報処理学 鹿野 清宏 池田和司 猿渡 洋
発表題目:拘束条件付き教師あり非負値行列因子分解による目的楽器音抽出とその評価
発表概要:本稿では,事前に教師情報を与えた非負値行列因子分解に,さらに抽出対象信号基底と非抽出対象信号基底が互いに無相関となるような制約をつけた非負値行列因子分解を提案する.教師情報のみを与えた非負値行列因子分解と,拘束条件を付与した非負値行列因子分解を用いて目的楽器音抽出実験を行った結果,拘束条件を付与した非負値行列因子分解の方が高い抽出精度を表し,さらに抽出対象信号となる楽器の種類によって抽出精度が異なることが明らかになった.