spb/wp30_uqo

1{2  "document_id": "lb_qasper_f55efaaee938369fc4be8a430e531c164168be643b3cfe57",3  "family": "public_qasper",4  "split": "dev",5  "text": "Introduction\nNeural machine translation (NMT) BIBREF0 , BIBREF1 , BIBREF2 has enabled end-to-end training of a translation system without needing to deal with word alignments, translation rules, and complicated decoding algorithms, which are the characteristics of phrase-based statistical machine translation (PBSMT) BIBREF3 . Although NMT can be significantly better than PBSMT in resource-rich scenarios, PBSMT performs better in low-resource scenarios BIBREF4 . Only by exploiting cross-lingual transfer learning techniques BIBREF5 , BIBREF6 , BIBREF7 , can the NMT performance approach PBSMT performance in low-resource scenarios.\nHowever, such methods usually require an NMT model trained on a resource-rich language pair like French INLINEFORM0 English (parent), which is to be fine-tuned for a low-resource language pair like Uzbek INLINEFORM1 English (child). On the other hand, multilingual approaches BIBREF8 propose to train a single model to translate multiple language pairs. However, these approaches are effective only when the parent target or source language is relatively resource-rich like English (En). Furthermore, the parents and children models should be trained on similar domains; otherwise, one has to take into account an additional problem of domain adaptation BIBREF9 .\nIn this paper, we work on a linguistically distant and thus challenging language pair Japanese INLINEFORM0 Russian (Ja INLINEFORM1 Ru) which has only 12k lines of news domain parallel corpus and hence is extremely resource-poor. Furthermore, the amount of indirect in-domain parallel corpora, i.e., Ja INLINEFORM2 En and Ru INLINEFORM3 En, are also small. As we demonstrate in Section SECREF4 , this severely limits the performance of prominent low-resource techniques, such as multilingual modeling, back-translation, and pivot-based PBSMT. To remedy this, we propose a novel multistage fine-tuning method for NMT that combines multilingual modeling BIBREF8 and domain adaptation BIBREF9 .\nWe have addressed two important research questions (RQs) in the context of extremely low-resource machine translation (MT) and our explorations have derived rational contributions (CTs) as follows:\nTo the best of our knowledge, we are the first to perform such an extensive evaluation of extremely low-resource MT problem and propose a novel multilingual multistage fine-tuning approach involving multilingual modeling and domain adaptation to address it.\nOur Japanese–Russian Setting\nIn this paper, we deal with Ja INLINEFORM0 Ru news translation. This language pair is very challenging because the languages involved have completely different writing system, phonology, morphology, grammar, and syntax. Among various domains, we experimented with translations in the news domain, considering the importance of sharing news between different language speakers. Moreover, news domain is one of the most challenging tasks, due to large presence of out-of-vocabulary (OOV) tokens and long sentences. To establish and evaluate existing methods, we also involved English as the third language. As direct parallel corpora are scarce, involving a language such as English for pivoting is quite common BIBREF10 .\nThere has been no clean held-out parallel data for Ja INLINEFORM0 Ru and Ja INLINEFORM1 En news translation. Therefore, we manually compiled development and test sets using News Commentary data as a source. Since the given Ja INLINEFORM2 Ru and Ja INLINEFORM3 En data share many lines in the Japanese side, we first compiled tri-text data. Then, from each line, corresponding parts across languages were manually identified, and unaligned parts were split off into a new line. Note that we have never merged two or more lines. As a result, we obtained 1,654 lines of data comprising trilingual, bilingual, and monolingual segments (mainly sentences) as summarized in Table TABREF8 . Finally, for the sake of comparability, we randomly chose 600 trilingual sentences to create a test set, and concatenated the rest of them and bilingual sentences to form development sets.\nOur manually aligned development and test sets are publicly available.\nRelated Work\nkoehn-knowles:2017:NMT showed that NMT is unable to handle low-resource language pairs as opposed to PBSMT. Transfer learning approaches BIBREF5 , BIBREF6 , BIBREF7 work well when a large helping parallel corpus is available. This restricts one of the source or the target languages to be English which, in our case, is not possible. Approaches involving bi-directional NMT modeling is shown to drastically improve low-resource translation BIBREF11 . However, like most other, this work focuses on translation from and into English.\nRemaining options include (a) unsupervised MT BIBREF12 , BIBREF13 , BIBREF14 , (b) parallel sentence mining from non-parallel or comparable corpora BIBREF15 , BIBREF16 , (c) generating pseudo-parallel data BIBREF17 , and (d) MT based on pivot languages BIBREF10 . The linguistic distance between Japanese and Russian makes it extremely difficult to learn bilingual knowledge, such as bilingual lexicons and bilingual word embeddings. Unsupervised MT is thus not promising yet, due to its heavy reliance on accurate bilingual word embeddings. Neither does parallel sentence mining, due to the difficulty of obtaining accurate bilingual lexicons. Pseudo-parallel data can be used to augment existing parallel corpora for training, and previous work has reported that such data generated by so-called back-translation can substantially improve the quality of NMT. However, this approach requires base MT systems that can generate somewhat accurate translations. It is thus infeasible in our scenario, because we can obtain only a weak system which is the consequence of an extremely low-resource situation. MT based on pivot languages requires large in-domain parallel corpora involving the pivot languages. This technique is thus infeasible, because the in-domain parallel corpora for Ja INLINEFORM0 En and Ru INLINEFORM1 En pairs are also extremely limited, whereas there are large parallel corpora in other domains. Section SECREF4 empirically confirms the limit of these existing approaches.\nFortunately, there are two useful transfer learning solutions using NMT: (e) multilingual modeling to incorporate multiple language pairs into a single model BIBREF8 and (f) domain adaptation to incorporate out-of-domain data BIBREF9 . In this paper, we explore a novel method involving step-wise fine-tuning to combine these two methods. By improving the translation quality in this way, we can also increase the likelihood of pseudo-parallel data being useful to further improve translation quality.\nLimit of Using only In-domain Data\nThis section answers our first research question, [RQ1], about the translation quality that we can achieve using existing methods and in-domain parallel and monolingual data. We then use the strongest model to conduct experiments on generating and utilizing back-translated pseudo-parallel data for augmenting NMT. Our intention is to empirically identify the most effective practices as well as recognize the limitations of relying only on in-domain parallel corpora.\nData\nTo train MT systems among the three languages, i.e., Japanese, Russian, and English, we used all the parallel data provided by Global Voices, more specifically those available at OPUS. Table TABREF9 summarizes the size of train/development/test splits used in our experiments. The number of parallel sentences for Ja INLINEFORM0 Ru is 12k, for Ja INLINEFORM1 En is 47k, and for Ru INLINEFORM2 En is 82k. Note that the three corpora are not mutually exclusive: 9k out of 12k sentences in the Ja INLINEFORM3 Ru corpus were also included in the other two parallel corpora, associated with identical English translations. This puts a limit on the positive impact that the helping corpora can have on the translation quality.\nEven when one focuses on low-resource language pairs, we often have access to larger quantities of in-domain monolingual data of each language. Such monolingual data are useful to improve quality of MT, for example, as the source of pseudo-parallel data for augmenting training data for NMT BIBREF17 and as the training data for large and smoothed language models for PBSMT BIBREF4 . Table TABREF13 summarizes the statistics on our monolingual corpora for several domains including the news domain. Note that we removed from the Global Voices monolingual corpora those sentences that are already present in the parallel corpus.\nhttps://dumps.wikimedia.org/backup-index.html (20180501) http://www.statmt.org/wmt18/translation-task.html https://www.yomiuri.co.jp/database/glossary/ http://www.cs.jhu.edu/~kevinduh/a/multitarget-tedtalks/ http://opus.nlpl.eu/Tatoeba-v2.php\nWe tokenized English and Russian sentences using tokenizer.perl of Moses BIBREF3 . To tokenize Japanese sentences, we used MeCab with the IPA dictionary. After tokenization, we eliminated duplicated sentence pairs and sentences with more than 100 tokens for all the languages.\nMT Methods Examined\nWe began with evaluating standard MT paradigms, i.e., PBSMT BIBREF3 and NMT BIBREF1 . As for PBSMT, we also examined two advanced methods: pivot-based translation relying on a helping language BIBREF10 and induction of phrase tables from monolingual data BIBREF14 .\nAs for NMT, we compared two types of encoder-decoder architectures: attentional RNN-based model (RNMT) BIBREF2 and the Transformer model BIBREF18 . In addition to standard uni-directional modeling, to cope with the low-resource problem, we examined two multi-directional models: bi-directional model BIBREF11 and multi-to-multi (M2M) model BIBREF8 .\nAfter identifying the best model, we also examined the usefulness of a data augmentation method based on back-translation BIBREF17 .\nFirst, we built a PBSMT system for each of the six translation directions. We obtained phrase tables from parallel corpus using SyMGIZA++ with the grow-diag-final heuristics for word alignment, and Moses for phrase pair extraction. Then, we trained a bi-directional MSD (monotone, swap, and discontinuous) lexicalized reordering model. We also trained three 5-gram language models, using KenLM on the following monolingual data: (1) the target side of the parallel data, (2) the concatenation of (1) and the monolingual data from Global Voices, and (3) the concatenation of (1) and all monolingual data in the news domain in Table TABREF13 .\nSubsequently, using English as the pivot language, we examined the following three types of pivot-based PBSMT systems BIBREF10 , BIBREF19 for each of Ja INLINEFORM0 Ru and Ru INLINEFORM1 Ja.\n2-step decoding using the source-to-English and English-to-target systems.\nObtain a new phrase table from synthetic parallel data generated by translating English side of the target–English training parallel data to the source language with the English-to-source system.\nCompile a new phrase table combining those for the source-to-English and English-to-target systems.\nAmong these three, triangulation is the most computationally expensive method. Although we had filtered the component phrase tables using the statistical significance pruning method BIBREF20 , triangulation can generate an enormous number of phrase pairs. To reduce the computational cost during decoding and the negative effects of potentially noisy phrase pairs, we retained for each source phrase INLINEFORM0 only the INLINEFORM1 -best translations INLINEFORM2 according to the forward translation probability INLINEFORM3 calculated from the conditional probabilities in the component models as defined in utiyama:07. For each of the retained phrase pairs, we also calculated the backward translation probability, INLINEFORM4 , and lexical translation probabilities, INLINEFORM5 and INLINEFORM6 , in the same manner as INLINEFORM7 .\nWe also investigated the utility of recent advances in unsupervised MT. Even though we began with a publicly available implementation of unsupervised PBSMT BIBREF13 , it crashed due to unknown reasons. We therefore followed another method described in marie:usmt-unmt. Instead of short INLINEFORM0 -grams BIBREF12 , BIBREF13 , we collected a set of phrases in Japanese and Russian from respective monolingual data using the word2phrase algorithm BIBREF21 , as in marie:usmt-unmt. To reduce the complexity, we used randomly selected 10M monolingual sentences, and 300k most frequent phrases made of words among the 300k most frequent words. For each source phrase INLINEFORM1 , we selected 300-best target phrases INLINEFORM2 according to the translation probability as in D18-1549: INLINEFORM3 where INLINEFORM4 stands for a bilingual embedding of a given phrase, obtained through averaging bilingual embeddings of constituent words learned from the two monolingual data using fastText and vecmap. For each of the retained phrase pair, INLINEFORM5 was computed analogously. We also computed lexical translation probabilities relying on those learned from the given small parallel corpus.\nUp to four phrase tables were jointly exploited by the multiple decoding path ability of Moses. Weights for the features were tuned using KB-MIRA BIBREF22 on the development set; we took the best weights after 15 iterations. Two hyper-parameters, namely, INLINEFORM0 for the number of pivot-based phrase pairs per source phrase and INLINEFORM1 for distortion limit, were determined by a grid search on INLINEFORM2 and INLINEFORM3 . In contrast, we used predetermined hyper-parameters for phrase table induction from monolingual data, following the convention: 200 for the dimension of word and phrase embeddings and INLINEFORM4 .\nWe used the open-source implementation of the RNMT and the Transformer models in tensor2tensor. A uni-directional model for each of the six translation directions was trained on the corresponding parallel corpus. Bi-directional and M2M models were realized by adding an artificial token that specifies the target language to the beginning of each source sentence and shuffling the entire training data BIBREF8 .\nTable TABREF22 contains some specific hyper-parameters for our baseline NMT models. The hyper-parameters not mentioned in this table used the default values in tensor2tensor. For M2M systems, we over-sampled Ja INLINEFORM0 Ru and Ja INLINEFORM1 En training data so that their sizes match the largest Ru INLINEFORM2 En data. To reduce the number of unknown words, we used tensor2tensor's internal sub-word segmentation mechanism. Since we work in a low-resource setting, we used shared sub-word vocabularies of size 16k for the uni- and bi-directional models and 32k for the M2M models. The number of training iterations was determined by early-stopping: we evaluated our models on the development set every 1,000 updates, and stopped training if BLEU score for the development set was not improved for 10,000 updates (10 check-points). Note that the development set was created by concatenating those for the individual translation directions without any over-sampling.\nHaving trained the models, we averaged the last 10 check-points and decoded the test sets with a beam size of 4 and a length penalty which was tuned by a linear search on the BLEU score for the development set.\nSimilarly to PBSMT, we also evaluated “Cascade” and “Synthesize” methods with uni-directional NMT models.\nResults\nWe evaluated MT models using case-sensitive and tokenized BLEU BIBREF23 on test sets, using Moses's multi-bleu.perl. Statistical significance ( INLINEFORM0 ) on the difference of BLEU scores was tested by Moses's bootstrap-hypothesis-difference-significance.pl.\nTables TABREF27 and TABREF31 show BLEU scores of all the models, except the NMT systems augmented with back-translations. Whereas some models achieved reasonable BLEU scores for Ja INLINEFORM0 En and Ru INLINEFORM1 En translation, all the results for Ja INLINEFORM2 Ru, which is our main concern, were abysmal.\nAmong the NMT models, Transformer models (b INLINEFORM0 ) were proven to be better than RNMT models (a INLINEFORM1 ). RNMT models could not even outperform the uni-directional PBSMT models (c1). M2M models (a3) and (b3) outperformed their corresponding uni- and bi-directional models in most cases. It is worth noting that in this extremely low-resource scenario, BLEU scores of the M2M RNMT model for the largest language pair, i.e., Ru INLINEFORM2 En, were lower than those of the uni- and bi-directional RNMT models as in TACL1081. In contrast, with the M2M Transformer model, Ru INLINEFORM3 En also benefited from multilingualism.\nStandard PBSMT models (c1) achieved higher BLEU scores than uni-directional NMT models (a1) and (b1), as reported by koehn-knowles:2017:NMT, whereas they underperform the M2M Transformer NMT model (b3). As shown in Table TABREF31 , pivot-based PBSMT systems always achieved higher BLEU scores than (c1). The best model with three phrase tables, labeled “Synthesize / Triangulate / Gold,” brought visible BLEU gains with substantial reduction of OOV tokens (3047 INLINEFORM0 1180 for Ja INLINEFORM1 Ru, 4463 INLINEFORM2 1812 for Ru INLINEFORM3 Ja). However, further extension with phrase tables induced from monolingual data did not push the limit, despite their high coverage; only 336 and 677 OOV tokens were left for the two translation directions, respectively. This is due to the poor quality of the bilingual word embeddings used to extract the phrase table, as envisaged in Section SECREF3 .\nNone of pivot-based approaches with uni-directional NMT models could even remotely rival the M2M Transformer NMT model (b3).\nTable TABREF46 shows the results of our multistage fine-tuning, where the IDs of each row refer to those described in Section SECREF41 . First of all, the final models of our multistage fine-tuning, i.e., V and VII, achieved significantly higher BLEU scores than (b3) in Table TABREF27 , a weak baseline without using any monolingual data, and #10 in Table TABREF33 , a strong baseline established with monolingual data.\nThe performance of the initial model (I) depends on the language pair. For Ja INLINEFORM0 Ru pair, it cannot achieve minimum level of quality since the model has never seen parallel data for this pair. The performance on Ja INLINEFORM1 En pair was much lower than the two baseline models, reflecting the crucial mismatch between training and testing domains. In contrast, Ru INLINEFORM2 En pair benefited the most and achieved surprisingly high BLEU scores. The reason might be due to the proximity of out-of-domain training data and in-domain test data.\nThe first fine-tuning stage significantly pushed up the translation quality for Ja INLINEFORM0 En and Ru INLINEFORM1 En pairs, in both cases with fine-tuning (II) and mixed fine-tuning (III). At this stage, both models performed only poorly for Ja INLINEFORM2 Ru pair as they have not yet seen Ja INLINEFORM3 Ru parallel data. Either model had a consistent advantage to the other.\nWhen these models were further fine-tuned only on the in-domain Ja INLINEFORM0 Ru parallel data (IV and VI), we obtained translations of better quality than the two baselines for Ja INLINEFORM1 Ru pair. However, as a result of complete ignorance of Ja INLINEFORM2 En and Ru INLINEFORM3 En pairs, the models only produced translations of poor quality for these language pairs. In contrast, mixed fine-tuning for the second fine-tuning stage (V and VII) resulted in consistently better models than conventional fine-tuning (IV and VI), irrespective of the choice at the first stage, thanks to the gradual shift of parameters realized by in-domain Ja INLINEFORM4 En and Ru INLINEFORM5 En parallel data. Unfortunately, the translation quality for Ja INLINEFORM6 En and Ru INLINEFORM7 En pairs sometimes degraded from II and III. Nevertheless, the BLEU scores still retain the large margin against two baselines.\nThe last three rows in Table TABREF46 present BLEU scores obtained by the methods with fewer fine-tuning steps. The most naive model I', trained on the balanced mixture of whole five types of corpora from scratch, and the model II', obtained through a single-step conventional fine-tuning of I on all the in-domain data, achieved only BLEU scores consistently worse than VII. In contrast, when we merged our two fine-tuning steps into a single mixed fine-tuning on I, we obtained a model III' which is better for the Ja INLINEFORM0 Ru pair than VII. Nevertheless, they are still comparable to those of VII and the BLEU scores for the other two language pairs are much lower than VII. As such, we conclude that our multistage fine-tuning leads to a more robust in-domain multilingual model.\nAugmentation with Back-translation\nGiven that the M2M Transformer NMT model (b3) achieved best results for most of the translation directions and competitive results for the rest, we further explored it through back-translation.\nWe examined the utility of pseudo-parallel data for all the six translation directions, unlike the work of lakew2017improving and lakew2018comparison, which concentrate only on the zero-shot language pair, and the work of W18-2710, which compares only uni- or bi-directional models. We investigated whether each translation direction in M2M models will benefit from pseudo-parallel data and if so, what kind of improvement takes place.\nFirst, we selected sentences to be back-translated from in-domain monolingual data (Table TABREF13 ), relying on the score proposed by moore:intelligent via the following procedure.\nFor each language, train two 4-gram language models, using KenLM: an in-domain one on all the Global Voices data, i.e., both parallel and monolingual data, and a general-domain one on the concatenation of Global Voices, IWSLT, and Tatoeba data.\nFor each language, discard sentences containing OOVs according to the in-domain language model.\nFor each translation direction, select the INLINEFORM0 -best monolingual sentences in the news domain, according to the difference between cross-entropy scores given by the in-domain and general-domain language models.\nWhereas W18-2710 exploited monolingual data much larger than parallel data, we maintained a 1:1 ratio between them BIBREF8 , setting INLINEFORM0 to the number of lines of parallel data of given language pair.\nSelected monolingual sentences were then translated using the M2M Transformer NMT model (b3) to compose pseudo-parallel data. Then, the pseudo-parallel data were enlarged by over-sampling as summarized in Table TABREF32 . Finally, new NMT models were trained on the concatenation of the original parallel and pseudo-parallel data from scratch in the same manner as the previous NMT models with the same hyper-parameters.\nTable TABREF33 shows the BLEU scores achieved by several reasonable combinations of six-way pseudo-parallel data. We observed that the use of all six-way pseudo-parallel data (#10) significantly improved the base model for all the translation directions, except En INLINEFORM0 Ru. A translation direction often benefited when the pseudo-parallel data for that specific direction was used.\nSummary\nWe have evaluated an extensive variation of MT models that rely only on in-domain parallel and monolingual data. However, the resulting BLEU scores for Ja INLINEFORM2 Ru and Ru INLINEFORM3 Ja tasks do not exceed 10 BLEU points, implying the inherent limitation of the in-domain data as well as the difficulty of these translation directions.\nExploiting Large Out-of-Domain Data Involving a Helping Language\nThe limitation of relying only on in-domain data demonstrated in Section SECREF4 motivates us to explore other types of parallel data. As raised in our second research question, [RQ2], we considered the effective ways to exploit out-of-domain data.\nAccording to language pair and domain, parallel data can be classified into four categories in Table TABREF40 . Among all the categories, out-of-domain data for the language pair of interest have been exploited in the domain adaptation scenarios (C INLINEFORM0 A) BIBREF9 . However, for Ja INLINEFORM1 Ru, no out-of-domain data is available. To exploit out-of-domain parallel data for Ja INLINEFORM2 En and Ru INLINEFORM3 En pairs instead, we propose a multistage fine-tuning method, which combines two types of transfer learning, i.e., domain adaptation for Ja INLINEFORM4 En and Ru INLINEFORM5 En (D INLINEFORM6 B) and multilingual transfer (B INLINEFORM7 A), relying on the M2M model examined in Section SECREF4 . We also examined the utility of fine-tuning for iteratively generating and using pseudo-parallel data.\nMultistage Fine-tuning\nSimply using NMT systems trained on out-of-domain data for in-domain translation is known to perform badly. In order to effectively use large-scale out-of-domain data for our extremely low-resource task, we propose to perform domain adaptation through either (a) conventional fine-tuning, where an NMT system trained on out-of-domain data is fine-tuned only on in-domain data, or (b) mixed fine-tuning BIBREF9 , where pre-trained out-of-domain NMT system is fine-tuned using a mixture of in-domain and out-of-domain data. The same options are available for transferring from Ja INLINEFORM0 En and Ru INLINEFORM1 En to Ja INLINEFORM2 Ru.\nWe inevitably involve two types of transfer learning, i.e., domain adaptation for Ja INLINEFORM0 En and Ru INLINEFORM1 En and multilingual transfer for Ja INLINEFORM2 Ru pair. Among several conceivable options for managing these two problems, we examined the following multistage fine-tuning.\nPre-train a multilingual model only on the Ja INLINEFORM0 En and Ru INLINEFORM1 En out-of-domain parallel data (I), where the vocabulary of the model is determined on the basis of the in-domain parallel data in the same manner as the M2M NMT models examined in Section SECREF4 .\nFine-tune the pre-trained model (I) on the in-domain Ja INLINEFORM0 En and Ru INLINEFORM1 En parallel data (fine-tuning, II) or on the mixture of in-domain and out-of-domain Ja INLINEFORM2 En and Ru INLINEFORM3 En parallel data (mixed fine-tuning, III).\nFurther fine-tune the models (each of II and III) for Ja INLINEFORM0 Ru on in-domain parallel data for this language pair only (fine-tuning, IV and VI) or on all the in-domain parallel data (mixed fine-tuning, V and VII).\nWe chose this way due to the following two reasons. First, we need to take a balance between several different parallel corpora sizes. The other reason is division of labor; we assume that solving each sub-problem one by one should enable gradual shift of parameters.\nData Selection\nAs an additional large-scale out-of-domain parallel data for Ja INLINEFORM0 En, we used the cleanest 1.5M sentences from the Asian Scientific Paper Excerpt Corpus (ASPEC) BIBREF24 . As for Ru INLINEFORM1 En, we used the UN and Yandex corpora released for the WMT 2018 News Translation Task. We retained Ru INLINEFORM2 En sentence pairs that contain at least one OOV token in both sides, according to the in-domain language model trained in Section SECREF34 . Table TABREF45 summarizes the statistics on the remaining out-of-domain parallel data.\nFurther Augmentation with Back-translation\nHaving obtained a better model, we examined again the utility of back-translation. More precisely, we investigated (a) whether the pseudo-parallel data generated by an improved NMT model leads to a further improvement, and (b) whether one more stage of fine-tuning on the mixture of original parallel and pseudo-parallel data will result in a model better than training a new model from scratch as examined in Section SECREF34 .\nGiven an NMT model, we first generated six-way pseudo-parallel data by translating monolingual data. For the sake of comparability, we used the identical monolingual sentences sampled in Section SECREF34 . Then, we further fine-tuned the given model on the mixture of the generated pseudo-parallel data and the original parallel data, following the same over-sampling procedure in Section SECREF34 . We repeated these steps five times.\nTable TABREF51 shows the results. “new #10” in the second row indicates an M2M Transformer model trained from scratch on the mixture of six-way pseudo-parallel data generated by VII and the original parallel data. It achieved higher BLEU scores than #10 in Table TABREF33 thanks to the pseudo-parallel data of better quality, but underperformed the base NMT model VII. In contrast, our fine-tuned model VIII successfully surpassed VII, and one more iteration (IX) further improved BLEU scores for all translation directions, except Ru INLINEFORM0 En. Although further iterations did not necessarily gain BLEU scores, we came to a much higher plateau compared to the results in Section SECREF4 .\nConclusion\nIn this paper, we challenged the difficult task of Ja INLINEFORM0 Ru news domain translation in an extremely low-resource setting. We empirically confirmed the limited success of well-established solutions when restricted to in-domain data. Then, to incorporate out-of-domain data, we proposed a multilingual multistage fine-tuning approach and observed that it substantially improves Ja INLINEFORM1 Ru translation by over 3.7 BLEU points compared to a strong baseline, as summarized in Table TABREF53 . This paper contains an empirical comparison of several existing approaches and hence we hope that our paper can act as a guideline to researchers attempting to tackle extremely low-resource translation.\nIn the future, we plan to confirm further fine-tuning for each of specific translation directions. We will also explore the way to exploit out-of-domain pseudo-parallel data, better domain-adaptation approaches, and additional challenging language pairs.\nAcknowledgments\nThis work was carried out when Aizhan Imankulova was taking up an internship at NICT, Japan. We would like to thank the reviewers for their insightful comments. A part of this work was conducted under the program “Promotion of Global Communications Plan: Research, Development, and Social Demonstration of Multilingual Speech Translation Technology” of the Ministry of Internal Affairs and Communications (MIC), Japan.",6  "char_len": 30136,7  "document_hash": "7683714da7a1bc48d6f6ec306dc65e5b31dad6ce24f8026867f24a0e5978d1e1",8  "questions": [9    {10      "question_id": "q00",11      "qtype": "qasper",12      "question": "what was the baseline?"13    }14  ]15}