A Study of Re-generating Sentences Given Similar Sentences that Cover Them on the Level of Form and Meaning
28799Authors
nan
Abstract
In this paper, we define a sentence re- generation task: re-generate a sentence given a set of sentences that cover it. Due to the absence of a dataset to perform this task, we firstly build three language resources of a new type containing more than 4 million anno- tated sentences. They contain sentences an- notated with similar sentences from the same corpus, that cover them on the level of form or meaning. We then perform the sentence re-generation task on the newly produced lan- guage resources using two approaches. The first one is a na¨ıve approach where we rely on a language model to reorder the covering parts. The second one is a neural approach where we treat the sentence re-generation task as a translation task from a sequence of cov- ering parts to the respective original sentence. The performance of the systems is evaluated on the level of form and meaning according to the type of covering used to re-generate the sentence. On the level of form, experi- mental results show that the neural approach outperforms the baseline in edit distance with up to 40% lower scores. However, in BLEU scores, the neural approach is similar or worse than the baseline. On the level of meaning, the neural approach always performs better than the baseline with average scores of 89% BERTScore. 1
Keywords
Full Content
A Study of Re-generating Sentences Given Similar Sentences that Cover Them on the Level of Form and Meaning Hsuan-Wei Lo, Yifei Zhou, Rashel Fam, Yves Lepage Graduate School of Information, Production and Systems Waseda University 2-7 Hibikino, Wakamatsu-ku, Kitakyushu-shi, 808-0135 Fukuoka-ken, Japan hsuanweilo@akane.waseda.jp, yifei.zhou@ruri.waseda.jp, fam.rashel@fuji.waseda.jp, yves.lepage@waseda.jp Abstract In this paper, we define a sentence re- generation task: re-generate a sentence given a set of sentences that cover it. Due to the absence of a dataset to perform this task, we firstly build three language resources of a new type containing more than 4 million anno- tated sentences. They contain sentences an- notated with similar sentences from the same corpus, that cover them on the level of form or meaning. We then perform the sentence re-generation task on the newly produced lan- guage resources using two approaches. The first one is a na¨ıve approach where we rely on a language model to reorder the covering parts. The second one is a neural approach where we treat the sentence re-generation task as a translation task from a sequence of cov- ering parts to the respective original sentence. The performance of the systems is evaluated on the level of form and meaning according to the type of covering used to re-generate the sentence. On the level of form, experi- mental results show that the neural approach outperforms the baseline in edit distance with up to 40% lower scores. However, in BLEU scores, the neural approach is similar or worse than the baseline. On the level of meaning, the neural approach always performs better than the baseline with average scores of 89% BERTScore. 1 Introduction Translation memories (TM) are used by translators to retrieve close sentences to a given source sen- tence as hints for translations. It has been shown that similar sentences or informative sentences re- trieved from TM can significantly boost the perfor- mance of neural machine translation (NMT) systems (Xu et al., 2020; Bult´e and Tezcan, 2019a). The idea of a set of sentences covering a given input sen- tence can be introduced to combine TM with ma- chine translation. (Liu and Lepage, 2021) proposed a twofold-objective approach to retrieve similar sen- tences based on an input sentence. To assess the quality of sentence coverage, we de- fine a sentence re-generation task: Re-generate an input sentence back given a set of sentences that cover it on two lev- els, form or meaning. Yet, there is a lack of data containing for this spe- cific task. The needed data should contain sentences similar in form and meaning along with the covering parts annotated for each input sentence from some corpus. Therefore, we build and release such lan- guage resources to meet the requirement of this new task. We then test a na¨ıve baseline and a neural ap- proach on the newly built resources to perform the sentence re-generation task. The two above aspects are the contributions of our work. The paper is organized as follows. Section 2 ex- plains the notion of …
Content truncated to 500 words
Labeled Text
In this paper, we define a sentence regeneration task: re-generate a sentence given a set of sentences that cover it. Due to the absence of a dataset to perform this task, we firstly build three language resources of a new type containing more than 4 million annotated sentences. They contain sentences annotated with similar sentences from the same corpus, that cover them on the level of form or meaning. We then perform the sentence re-generation task on the newly produced language resources using two approaches. The first one is a na¨ıve approach where we rely on a language model to reorder the covering parts. The second one is a neural approach where we treat the sentence re-generation task as a translation task from a sequence of covering parts to the respective original sentence. The performance of the systems is evaluated on the level of form and meaning according to the type of covering used to re-generate the sentence. On the level of form, experimental results show that the neural approach outperforms the baseline in edit distance with up to 40% lower scores. However, in BLEU scores, the neural approach is similar or worse than the baseline. On the level of meaning, the neural approach always performs better than the baseline with average scores of 89% BERTScore. 1
Publication Details
- Paper ID:
- 28799
- Year:
- 2022
- Publication:
- TIPSTER
- Publication Type:
- conf
- DOI:
- nan