Covering a sentence in form and meaning with fewer retrieved sentences

Covering a sentence in form and meaning with fewer retrieved sentences Yuan Liu Graduate School of IPS Waseda University Kitakyushu, Japan lyalltowell@ruri.waseda.jp Yves Lepage Graduate School of IPS Waseda University Kitakyushu, Japan yves.lepage@waseda.jp Abstract Retrieving similar sentences from a given col- lection of sentences is essential in a range of applications. In this work, we propose a novel method to retrieve several sentences that cover an input sentence in form and mean- ing with minimal redundancy, so as to en- hance the overall coverage quality of the out- put sentences. We focus on the hierarchical granularity levels of sentence pieces, match- ing from common or similar n-grams to finer- grained words o subwords, using techniques from similar sentence retrieval and monolin- gual phrase alignment. Our method shows promising source and target coverage evalu- ation results when applied to parallel corpora. This shows the potential of our approach if in- tegrated into an example-based machine trans- lation system. 1 Introduction Recent research has indicated that informative sen- tences retrieved from translation memories (TM) boost the performance of neural machine translation (NMT) systems (Xu et al., 2020; Bult´e and Tezcan, 2019). In particular, to better integrate TM in NMT, Tezcan et al. (2021) implemented a fuzzy match combination method to maximise the coverage of words in source sentence for data augmentation. In their work, Tezcan et al. (2021) argued that the idea of TM-NMT integration closely relates to the principle of example-based machine translation (EBMT) (Nagao, 1984). In EBMT, sentences which are similar to a given input sentence are retrieved A : B :: C : D ᄉ ᅡᄆ ᅡ ᆨᄋ ᅦᄉ ᅥ ᄋ ᅮ ᆯᄃ ᅡ : ᄁ ᅮ ᆷᄋ ᅦᄉ ᅥ ᄋ ᅮ ᆯᄃ ᅡ :: ᄉ ᅡᄆ ᅡ ᆨᄋ ᅦᄉ ᅥ ᄋ ᅮ ᆺᄃ ᅡ : ᄁ ᅮ ᆷᄋ ᅦᄉ ᅥ ᄋ ᅮ ᆺᄃ ᅡ 赤色は 明るい : 紫色は 暗い :: 黄色は 暖かい : 緑色は 寒い ‘red is bright’ : ‘purple is dark’ :: ‘yellow is warm’ : ‘green is cold’ Figure 1: General pattern for analogy on the 1st line, formal analogy on the 2nd line, and semantic analogy on the 3rd line (with its translation into En- glish). Analogies require formal or semantic cover- age of a given sentence. from a bilingual corpus. The machine builds a trans- lation of the input sentence from pieces of the trans- lations of these similar sentences. For the machine to translate properly, the retrieved sentences should cover the input sentence. One approach to EBMT is translation by anal- ogy (Lepage and Denoual, 2005). It relies on the preservation of proportional analogies between sen- tences across languages. A necessary condition for formal analogies and semantic analogies to hold is that sentence A has to be covered by sentences B and C, either in form or in meaning (Stroppa and Yvon, 2005; Langlais et al., 2009; Lepage, 2019), as illustrated in Figure 1. Further examples of sen- tences formally and semantically covering a sen- tence are shown in Figure 2. As mentioned …

Content truncated to 500 words