| VU HUY HIEN | D, 中間発表 | 自然言語処理学 | 渡辺 太郎, 中村 哲, 上垣外 英剛 |
|
title: Improving Coherence of Neural Machine Translation
abstract: Recently, Machine Translation (MT) has gained a big jump since the development of neural networks and the contribution of research communities in many aspects. However, most research in MT focuses on sentence-level translation, and only a few works aim to solve problems at the document-level, which often contains the coherence issue known as the connection between sentences, entities, and objects in a whole document. Therefore, in most current systems, there is no guarantee of discourse-coherent translation, which is an essential aspect of helping readers understand the original texts. In this research, we focus on document-level machine translation and aim to improve the coherence between sentences by using external information (the coreference information) and controlling the coherence in both training and inference steps. We first do a training step with the coreference probing task for maintaining the coherence of sentences based on provided external information, and then we use the coherence criteria as a supplying condition to generate the output of translation in the inference step. The initial results show that our approach can maintain the quality of translation at the document-level while keeping coherence in long sentence settings. language of the presentation: English | |||
| 出口 祥之 | D, 中間発表 | 自然言語処理学 | 渡辺 太郎, 中村 哲, 上垣外 英剛 |
|
title: k-Nearest-Neighbor Neural Machine Translation with Sentence Retrieval
abstract: Recently, k-nearest-neighbor neural machine translation (kNN-MT) has been improved the translation performance using translation examples. Using translation examples can significantly improve translation performance without additional training, especially in out-of-domains different from the training corpus. However, kNN-MT uses word vectors as queries and searches for neighbors from large parallel corpora, which reduces the translation speed by more than 100 times compared to non-retrieval models. Therefore, I propose a similar sentence retrieval method to speed up the retrieval process. The method improves the translation speed over the conventional kNN-MT while preventing the degradation of translation performance. language of the presentation: Japanese 発表題目: 類似文検索を用いたkNNニューラル機械翻訳 発表概要: 近年，ニューラル機械翻訳の翻訳精度を向上を目的として，翻訳用例を活用したkNNニューラル機械翻訳 (kNN-MT) の研究が盛んに行われている． 翻訳用例を用いることで，特に，翻訳モデルの訓練コーパスと異なるドメインの翻訳精度を追加学習なしに飛躍的に改善できる． しかし，kNN-MTでは翻訳用例の検索クエリには単語ベクトルを用い，大規模対訳データから近傍単語を検索するため，翻訳速度は従来手法の100倍以上低下する． 本研究では，入力文の類似文検索により，翻訳精度の低下を防ぎながら従来手法より翻訳速度を大きく改善する． | |||
| 星野 智紀 | M, 2回目発表 | 自然言語処理学 | 渡辺 太郎, 中村 哲, 上垣外 英剛 |
|
title: Abstractive summarization with controlling faithfulness and fluency
abstract: Finetuned language model can generate very fluent summaries. But there is a problem that these models can generate summaries not based on facts. The presenter tackles this problem to control summaries' extractiveness. Controlling summaries' extractiveness by Reinforcement Learning, It was suggested that changes occur in the appearance of words in the generated summary. The presenter is going to talk about the current situation and the plan of the research. language of the presentation: Japanese | |||