Neural Machine Translation: A Review

In this work, we provide a systematic and comprehensive empirical comparison of pretrained multilingual language models versus their monolingual counterparts with regard to their monolingual task performance. We study a set of nine typologically diverse languages with readily available pretrained monolingual models on a set of five diverse monolingual downstream tasks. We first aim to establish, via fair and controlled comparisons, if a gap between the multilingual and the corresponding monolingual representation of that language exists, and subsequently investigate the reason for any performance difference. To disentangle conflating factors, we train new monolingual models on the same data, with monolingually and multilingually trained tokenizers. We find that while the pretraining data size is an important factor, a designated monolingual tokenizer plays an equally important role in the downstream performance. Our results show that languages that are adequately represented in the multilingual model’s vocabulary exhibit negligible performance decreases over their monolingual counterparts. We further find that replacing the original multilingual tokenizer with the specialized monolingual tokenizer improves the downstream performance of the multilingual model for almost every task and language.

/pdf/how-good-is-your-tokenizer-on-the-monolingual-performance-of-56j7se1ztp.pdf

How Good is Your Tokenizer? On the Monolingual Performance of Multilingual Language Models

The success of pretrained transformer language models (LMs) in natural language processing has led to a wide range of pretraining setups. In particular, these models employ a variety of subword tokenization methods, most notably byte-pair encoding (BPE) (Sennrich et al., 2016; Gage, 1994), the WordPiece method (Schuster and Nakajima, 2012), and unigram language modeling (Kudo, 2018), to segment text. However, to the best of our knowledge, the literature does not contain a direct evaluation of the impact of tokenization on language model pretraining. We analyze differences between BPE and unigram LM tokenization, finding that the latter method recovers subword units that align more closely with morphology and avoids problems stemming from BPE's greedy construction procedure. We then compare the fine-tuned task performance of identical transformer masked language models pretrained with these tokenizations. Across downstream tasks and two languages (English and Japanese), we find that the unigram LM tokenization method matches or outperforms BPE. We hope that developers of future pretrained LMs will consider adopting the unigram LM method over the more prevalent BPE.

Byte Pair Encoding is Suboptimal for Language Model Pretraining

Byte-Pair Encoding (BPE) is an unsupervised sub-word tokenization technique, commonly used in neural machine translation and other NLP tasks. Its effectiveness makes it a de facto standard, but the reasons for this are not well understood. We link BPE to the broader family of dictionary-based compression algorithms and compare it with other members of this family. Our experiments across datasets, language pairs, translation models, and vocabulary size show that - given a fixed vocabulary size budget - the fewer tokens an algorithm needs to cover the test set, the better the translation (as measured by BLEU).

/pdf/investigating-the-effectiveness-of-bpe-the-power-of-shorter-1t4m230yex.pdf

Investigating the Effectiveness of BPE: The Power of Shorter Sequences.

In this work, we provide a systematic and comprehensive empirical comparison of pretrained multilingual language models versus their monolingual counterparts with regard to their monolingual task performance. We study a set of nine typologically diverse languages with readily available pretrained monolingual models on a set of five diverse monolingual downstream tasks. We first aim to establish, via fair and controlled comparisons, if a gap between the multilingual and the corresponding monolingual representation of that language exists, and subsequently investigate the reason for any performance difference. To disentangle conflating factors, we train new monolingual models on the same data, with monolingually and multilingually trained tokenizers. We find that while the pretraining data size is an important factor, a designated monolingual tokenizer plays an equally important role in the downstream performance. Our results show that languages that are adequately represented in the multilingual model's vocabulary exhibit negligible performance decreases over their monolingual counterparts. We further find that replacing the original multilingual tokenizer with the specialized monolingual tokenizer improves the downstream performance of the multilingual model for almost every task and language.

How Good is Your Tokenizer? On the Monolingual Performance of Multilingual Language Models.

Tokenization or segmentation is a wide concept that covers simple processes such as separating punctuation from words, or more sophisticated processes such as applying morphological knowledge. Neural Machine Translation (NMT) requires a limited-size vocabulary for computational cost and enough examples to estimate word embeddings. Separating punctuation and splitting tokens into words or subwords has proven to be helpful to reduce vocabulary and increase the number of examples of each word, improving the translation quality. Tokenization is more challenging when dealing with languages with no separator between words. In order to assess the impact of the tokenization in the quality of the final translation on NMT, we experimented on five tokenizers over ten language pairs. We reached the conclusion that the tokenization significantly affects the final translation quality and that the best tokenizer differs for different language pairs.

How Much Does Tokenization Affect Neural Machine Translation

A common use of machine translation in the industry is providing initial translation hypotheses, which are later supervised and post-edited by a human expert. During this revision process, new bilingual data are continuously generated. Machine translation systems can benefit from these new data, incrementally updating the underlying models under an online learning paradigm. We conducted a user study on this scenario, for a neural machine translation system. The experimentation was carried out by professional translators, with a vast experience in machine translation post-editing. The results showed a reduction in the required amount of human effort needed when post-editing the outputs of the system, improvements in the translation quality and a positive perception of the adaptive system by the users.

Incremental Adaptation of NMT for Professional Post-editors: A User Study

This paper reports the results of an indepth evaluation of 34 state-of-the-art domain-adapted machine translation (MT) systems that were built by four leading MT companies as part of the EU-funded iADAATPA project. These systems support a wide variety of languages for several domains. The evaluation combined automatic metrics and human methods, namely assessments of adequacy, ﬂuency, and comparative ranking. The paper also discusses the most effective techniques to build domain-adapted MT systems for the relevant language combinations and domains.

Large-scale machine translation evaluation of the iADAATPA project

We describe the MAPA project, funded under the Connecting Europe Facility programme, whose goal is the development of an open-source de-identification toolkit for all official European Union languages. It will be developed since January 2020 until December 2021.

https://hal.archives-ouvertes.fr/hal-03103205/document

Manuel Herranz

Papers

How Much Does Tokenization Affect Neural Machine Translation

Incremental Adaptation of NMT for Professional Post-editors: A User Study

Large-scale machine translation evaluation of the iADAATPA project

Incremental Adaptation of NMT for Professional Post-editors: A User Study

The Multilingual Anonymisation Toolkit for Public Administrations (MAPA) Project