| 石橋 陽一 | D, 中間発表 | 知能コミュニケーション | 中村 哲, | 渡辺 太郎, | 須藤 克仁 | |
|
title: Quantum-Inspired Set Operations on a Word Embedding Space
abstract: Word embedding is a fundamental tool in natural language processing, and its quality evaluation is an important issue that affects the entire field. Most of the existing measures evaluate the relationship between individual words and their pairs. In contrast, we focus on word sets. In this study, we aim to evaluate the relations between word sets using set operations in the embedding space. We focus on quantum logic, which is a set operation corresponding to Euclidean space, and show experimentally that it can realize set operations in word embedding space. We also show that set-based evaluation is positively correlated with the results of downstream tasks. language of the presentation: Japanese 発表題目: 量子論理に基づく単語埋め込み空間上の集合演算 発表概要: 単語埋め込みは自然言語処理の基盤的な道具であり、その品質評価は分野全体に影響する重要な課題である。既存評価尺度は個々の単語やその組の関係を評価しているものほとんどであった。しかし単語の集合にも言語的な意味が存在する。そこで本研究では集合に焦点を当て、単語集合・単語集合間の関係を評価することを目指す。我々は量子力学で提案されたヒルベルト空間上の集合演算である「量子論理」に着目し、単語埋め込み空間で集合演算を実現する方法を提案する。本発表では埋め込み空間に量子論理を適用した実験と、集合に基づいた評価と後段タスクとの相関について紹介する。 | ||||||
| 髙橋 洸丞 | D, 中間発表 | 知能コミュニケーション | 中村 哲, | 渡辺 太郎, | 須藤 克仁 | |
|
title Autmatic evaluation of critical errors generated by machine translation
abstract Automatic evaluation metrics have been improved much better compared to the most frequentlly used metric named BLEU, however the recent and strongly performing metrics are tuned to better evaluate high-performing machine translation systems and their translations. The problem of not being able to correctly evalaute critical errors remains. In this presentation, we talk about the problem itself and our plan to approach this problem. language of the presentation Japanese | ||||||
| 田中 翔平 | D, 中間発表 | 知能コミュニケーション | 中村 哲, | 渡辺 太郎, | 須藤 克仁, | 吉野 幸一郎 |
|
title: Reflective Action Selection Based on Positive-Unlabeled Learning and Causality Detection Model
abstract: Task-oriented dialogue systems take appropriate actions depends on user requests. The systems need to take appropriate actions not only for clear user requests but also for ambiguous requests. To the ambiguous requests, taking reflective actions is one of the plausible choices for the systems. In order to make our system possible to select reflective actions for ambiguous user requests, we collected a corpus includes pairs of ambiguous user requests and reflective system actions on sightseeing navigation with a smartphone. Since annotating all possible combinations of user requests and system actions is impractical, in this study, we built a corpus that one of the multiple reflective actions is annotated to one ambiguous user request. In order to train the action selection model on such training data, we applied the positive/unlabeled (PU) learning method, which assumes that only a part of the data is labeled with positive examples. In addition, we tried to achieve more accurate selection by extracting and distilling knowledge corresponding to causality from the training data using a causality detection model. The experimental results show that the proposed method with the PU learning and the causality detection model achieved better performance than the model based on conventional positive/negative (PN) learning method to select the reflective actions. language of the presentation: Japanese | ||||||