English-to-Japanese Diverse Translation by Combining Forward and Backward Outputs

Masahiro Kaneko, Aizhan Imankulova, Tosho Hirasawa, Mamoru Komachi
Tokyo Metropolitan University
6-6 Asahigaoka, Hino, Tokyo 191-0065, Japan
{kaneko-masahiro, imankulova-aizhan, hirasawa-tosho}@ed.tmu.ac.jp
komachi@tmu.ac.jp

Abstract

We introduce our TMU system that is submitted to The 4th Workshop on Neural Generation and Translation (WNGT2020) to English-to-Japanese (En→Ja) track on Simultaneous Translation And Paraphrase for Language Education (STAPLE) shared task. In most cases machine translation systems generate a single output from the input sentence, however, in order to assist language learners in their journey with better and more diverse feedback, it is helpful to create a machine translation system that is able to produce diverse translations of each input sentence. In particular, we combined the outputs from forward and backward neural translation models (NMT). Our system achieved third place in En→Ja track, despite adopting only a simple approach.

1 Introduction

WNGT2020 on STAPLE addresses generating high-coverage sets of plausible translations which can be useful in machine translation (MT), MT evaluation, multilingual paraphrase, and language education technology fields. In Duolingo, some learning takes place via translation-based exercises and assessment is done by comparing the learners’ responses to a large set of acceptable human-generated translations. Therefore, retaining richer paraphrases of the translation results would help to generate more accurate feedback to the learners.

Several studies have been conducted on the diversity of translation results (Vijayakumar et al., 2018; Xu et al., 2018; Shu et al., 2019; Ippolito et al., 2019). On the other hand, these methods rely on complex approaches. However, we were curious if we can produce diverse outputs only using a simple approach.

1.1 Motivation

• No need to modify the NMT model.
• It is known that L2R translates prefixes and R2L translates suffixes better (Liu et al., 2016). This indicates that L2R and R2L produce different translation results.

Our TMU system achieved third place using only the simple approach.

2 Related Work

Several models have been proposed to generate diverse decoding outputs for different tasks. For example, Xu et al. (2018) proposed diverse paraphrase generation by introducing rewriting patterns into the decoder of the encoder-decoder model. Vijayakumar et al. (2018) proposed diverse beam search algorithm for decoding diverse sequences. Shu et al. (2019) generated diverse translations by conditioning sentence generation with the sentence codes. Ippolito et al. (2019) proposed over-sampling followed by post-decoding clustering to remove similar sequences.

3 Experiments

We used the open-source fairseq for training NMT models. We adopt the Transformer as our translation model. We train two types of models, L2R and R2L for decoding.

In our preliminary experiments, we found that even the combination of L2R and R2L translation results can produce a sufficiently diverse set of translations.

4 Results

We used weighted macro F1 as the main scoring metric. Table 3 lists the F1 scores of participating systems in En→Ja track. Our TMU system was ranked the third.

In this paper, we introduced our system submitted to WNGT2020 shared task to En→Ja track on STAPLE. We have shown that even a simple method can generate a variety of translations while maintaining original meaning and grammaticality.