コロキアムB発表

日時: 9月11日 (水) 2限目(11:00-12:30)


会場: L1

司会: Pham Hoai Luan
五藤 巧 D, 中間発表 自然言語処理学 渡辺 太郎 荒牧 英治 大内 啓樹
title: Evaluating grammaticality of the neural models through explaning of the prediction
abstract: The grammaticality of a model is important for language generation, which is a fundamental ability in language processing, and for application tasks that require grammatical knowledge, such as grammatical error correction. On the other hand, conventional evaluations are based on sentence-level evaluation or simple agreement in classification tasks, making detailed evaluation of the grammaticality of models difficult. In this presentation, I propose methods to evaluate grammaticality at a finer granularity based on explaining model predictions. Specifically, I describe a task-independent method for explaining predictions from k-nearest neighbor examples, and a method based on feature attribution for grammatical error correction task, by regarding each edit as a feature.
language of the presentation: Japanese
発表題目: ニューラルモデルにおける予測の説明に基づく文法性の評価
発表概要: ニューラルモデルに基づく自然言語処理モデルを評価する観点の一つに文法性がある. モデルの文法性の高さは,言語処理における基礎的な能力である言語生成能力や,文法誤り訂正などの文法知識を必要とする応用タスクにおいて重要である. 一方,従来の評価では文単位の評価や,分類タスクの単純な一致に基づく評価に終始しているため,モデルの文法性に関する詳細な評価が難しい. 本発表では,モデルの予測をより細かい粒度で説明することを通して,文法性を詳細に評価する方法を提案する. 具体的には,タスク非依存な手法としてk近傍事例から予測を説明する方法と,文法誤り訂正タスクのための編集を特徴量とみなした特徴量帰属に基づく方法について述べる.
 
帖佐 克己 D, 中間発表 自然言語処理学 渡辺 太郎 荒牧 英治 上垣外 英剛
title: Structure constrained neural machine translation
abstract: Lexically constrained machine translation involves controlling the translation model to ensure that the translation outputs include specific terms. Expanding the constraint unit from terms to text structures would enhance the controllability of translation, but machine translation with structural constraints has not yet been explored. In this presentation, we introduce a method for conducting machine translation using the substructures of the target language-side text as constraints, which builds upon traditional lexically constrained methods.
language of the presentation: Japanese
 
QU ZHI D, 中間発表 自然言語処理学 渡辺 太郎 荒牧 英治 上垣外 英剛
title: Languages Transferred Within the Encoder: On Representation Transfer in Zero-Shot Multilingual Translation
abstract: Understanding representation transfer in multilingual neural machine translation can reveal the representational issue causing the zero-shot translation deficiency. In this work, we introduce the identity pair, a sentence translated into itself, to address the lack of the base measure in multilingual investigations, as the identity pair represents the optimal state of representation among any language transfers. In our analysis, we demonstrate that the encoder transfers the source language to the representational subspace of the target language instead of the language-agnostic state. Thus, the zero-shot translation deficiency arises because representations are entangled with other languages and are not transferred effectively to the target language. Based on our findings, we propose two methods: 1) low-rank language-specific embedding at the encoder, and 2) language-specific contrastive learning of the representation at the decoder. The experimental results on Europarl-15, TED-19, and OPUS-100 datasets show that our methods substantially enhance the performance of zero-shot translations by improving language transfer capacity, thereby providing practical evidence to support our conclusions.
language of the presentation: English
 

会場: L2

司会: 笹田大翔
工藤 創大 D, 中間発表 計算システムズ生物学 金谷 重彦 安本 慶一 小野 直亮 MD.Altaf-Ul-Amin

title: *** Information extractor with variable compression level *** 

abstract: *** Information Bottleneck (IB) is a widely used machine learning framework that enables the extraction of information related to a target random variable from a source random variable. In the objective function, IB controls the trade-off between data compression and predictiveness through the Lagrange multiplier $\beta$. Traditionally, to find the trade-off to be learned, IB requires a search for $\beta$ through multiple training cycles, which is computationally expensive. In this study, we introduce Flexible Variational Information Bottleneck (FVIB), a framework for classification task that can obtain optimal models for all values of $\beta$ with single, computationally efficient training.  *** 

language of the presentation: *** Japanese ***