A study on compact and fast translation and language models for statistical machine translation

2017 Fiscal Year Final Research Report

• Project/Area Number: 15H02744
• Research Category: Grant-in-Aid for Scientific Research (B)
• Allocation Type: Single-year Grants
• Section: 一般
• Research Field: Intelligent informatics
• Research Institution: University of Tsukuba
• Principal Investigator: YAMAMOTO Mikio 筑波大学, システム情報系, 教授 (40210562)
• Co-Investigator(Kenkyū-buntansha): 乾 孝司 筑波大学, システム情報系, 准教授 (60397031)
• Research Collaborator: NORIMATSU Jun-ya ; TANIGUCHI Masanori ; HAGA Shumpei ; OSUMI Kenji ; TAKENAKA Kousuke ; ISHII Akihiko
• Project Period (FY): 2015-04-01 – 2018-03-31
• Keywords: 言語モデル / ダブル配列 / 部分転置ダブル配列 / ランダム配置

Outline of Final Research Achievements

Although DALM (Double-Array Language Model) is a fast and compact implementation of ngram language models, it fails to fully capitalize on quantization techniques for values of model parameters such as probabilities of ngrams, because of a structual limitation: it stores values and indexes in the common array. In this study, we developed some variants of DALM which have separate arrays for values and indexes and can exploit benefits of quantization. We investigated basic characteristics of DALM empirically and propose "partly transposed double-array" which is a key technique to educe the ability of DALMs with separate arrays.

Free Research Field

情報工学