1{2 "document_id": "lb_qasper_d7fcca67f5487a55282f8217e4f05bc54ee67e2ddb578e73",3 "family": "public_qasper",4 "split": "test",5 "text": "Introduction\nDeep Learning approaches have achieved impressive results on various NLP tasks BIBREF0 , BIBREF1 , BIBREF2 and have become the de facto approach for any NLP task. However, these deep learning techniques have found to be less effective for low-resource languages when the available training data is very less BIBREF3 . Recently, several approaches like Multi-task learning BIBREF4 , multilingual learning BIBREF5 , semi-supervised learning BIBREF2 , BIBREF6 and transfer learning BIBREF7 , BIBREF3 have been explored by the deep learning community to overcome data sparsity in low-resource languages. Transfer learning trains a model for a parent task and fine-tunes the learned parent model weights (features) for a related child task BIBREF7 , BIBREF8 . This effectively reduces the requirement on training data for the child task as the model would have learned relevant features from the parent task data thereby, improving the performance on the child task.\nTransfer learning has also been explored in the multilingual Neural Machine Translation BIBREF3 , BIBREF9 , BIBREF10 . The goal is to improve the NMT performance on the source to target language pair (child task) using an assisting source language (assisting to target translation is the parent task). Here, the parent model is trained on the assisting and target language parallel corpus and the trained weights are used to initialize the child model. The child model can now be fine-tuned on the source-target language pairs, if parallel corpus is available. The divergence between the source and the assisting language can adversely impact the benefits obtained from transfer learning. Multiple studies have shown that transfer learning works best when the languages are related BIBREF3 , BIBREF10 , BIBREF9 . Several studies have tried to address lexical divergence between the source and the target languages BIBREF10 , BIBREF11 , BIBREF12 . However, the effect of word order divergence and its mitigation has not been explored. In a practical setting, it is not uncommon to have source and assisting languages with different word order. For instance, it is possible to find parallel corpora between English and some Indian languages, but very little parallel corpora between Indian languages. Hence, it is natural to use English as an assisting language for inter-Indian language translation.\nTo see how word order divergence can be detrimental, let us consider the case of the standard RNN (Bi-LSTM) encoder-attention-decoder architecture BIBREF13 . The encoder generates contextual representations (annotation vectors) for each source word, which are used by the attention network to match the source words to the current decoder state. The contextual representation is word-order dependent. Hence, if the assisting and the source languages do not have similar word order the generated contextual representations will not be consistent. The attention network (and hence the decoder) sees different contextual representations for similar words in parallel sentences across different languages. This makes it difficult to transfer knowledge learned from the assisting language to the source language.\nWe illustrate this by visualizing the contextual representations generated by the encoder of an English to Hindi NMT system for two versions of the English input: (a) original word order (SVO) (b) word order of the source language (SOV, for Bengali). Figure FIGREF1 shows that the encoder representations obtained are very different. The attention network and the decoder now have to work with very different representations. Note that the plot below does not take into account further lexical and other divergences between source and assisting languages, since we demonstrated word order divergence with the same language on the source side.\nTo address this word order divergence, we propose to pre-order the assisting language sentences to match the word order of the source language. We consider an extremely resource constrained scenario, where we do not have any parallel corpus for the child task. We are limited to a bilingual dictionary for transfer information from the assisting to the source language. From our experiments, we show that there is a significant increase in the translation accuracy for the unseen source-target language pair.\nAddressing Lexical Divergence\nBIBREF3 explored transfer learning for NMT on low-resource languages. They studied the influence of language divergence between languages chosen for training the parent and child model, and showed that choosing similar languages for training the parent and child model leads to better improvements from transfer learning. A limitation of BIBREF3 approach is that they ignore the lexical similarity between languages and also the source language embeddings are randomly initialized. BIBREF10 , BIBREF11 , BIBREF12 take advantage of lexical similarity between languages in their work. BIBREF10 proposed to use Byte-Pair Encoding (BPE) to represent the sentences in both the parent and the child language to overcome the above limitation. They show using BPE benefits transfer learning especially when the involved languages are closely-related agglutinative languages. Similarly, BIBREF11 utilize lexical similarity between the source and assisting languages by training a character-level NMT system. BIBREF12 address lexical divergence by using bilingual embeddings and mixture of universal token embeddings. One of the languages' vocabulary, usually English vocabulary is considered as universal tokens and every word in the other languages is represented as a mixture of universal tokens. They show results on extremely low-resource languages.\nAddressing Word Order Divergence\nTo the best of our knowledge, no work has addressed word order divergence in transfer learning for multilingual NMT. However, some work exists for other NLP tasks that could potentially address word order. For Named Entity Recognition (NER), BIBREF14 use a self-attention layer after the Bi-LSTM layer to address word-order divergence for Named Entity Recognition (NER) task. The approach does not show any significant improvements over multiple languages. A possible reason is that the divergence has to be addressed before/during construction of the contextual embeddings in the Bi-LSTM layer, and the subsequent self-attention layer does not address word-order divergence. BIBREF15 use adversarial training for cross-lingual question-question similarity ranking in community question answering. The adversarial training tries to force the encoder representations of similar sentences from different input languages to have similar representations.\nUse of Pre-ordering\nPre-ordering the source language sentences to match the target language word order has been useful in addressing word-order divergence for Phrase-Based SMT BIBREF16 , BIBREF17 , BIBREF18 , BIBREF19 . Recently, BIBREF20 proposed a way to measure and reduce the divergence between the source and target languages based on morphological and syntactic properties, also termed as anisomorphism. They demonstrated that by reducing the anisomorphism between the source and target languages, consistent improvements in NMT performance were obtained. The NMT system used additional features like word forms, POS tags and dependency relations in addition to parallel corpora. On the other hand, BIBREF21 observed a drop in performance due to pre-ordering for NMT. Unlike BIBREF20 , the NMT system was trained on pre-ordered sentences and no additional features were provided to the system. Note that all these works address source-target divergence, not divergence between source languages in multilingual NMT.\nProposed Solution\nConsider the task of translating from an extremely low-resource language (source) to a target language. The parallel corpus between the two languages if available may be too small to train a NMT model. Similar to existing works BIBREF3 , BIBREF10 , BIBREF12 , we use transfer learning to overcome data sparsity and train a NMT model between the source and the target languages. Specifically, the NMT model (parent model) is trained on the assisting language and target language pairs. We choose English as the assisting language in all our experiments. In our resource-scarce scenario, we have no parallel corpus for the child task. Hence, at test time, the source language sentence is translated using the parent model after performing a word-by-word translation into the assisting language.\nSince the source language and the assisting language (English) have different word order, we hypothesize that it leads to inconsistencies in the contextual representations generated by the encoder for the two languages. In this paper, we propose to pre-order English sentences (assisting language sentences) to match the word-order of the source language and train the parent model on this pre-ordered corpus. In our experiments, we look at scenarios where the assisting language has SVO word order and the source language has SOV word order.\nFor instance, consider the English sentence Anurag will meet Thakur. One of the pre-ordering rule swaps the position of the noun phrase followed by a transitive verb with the transitive verb. The original and the resulting re-ordered parse tree will be as shown in the Table TABREF5 . Applying this reordering rule to the above sentence Anurag will meet Thakur will yield the reordered sentence Anurag Thakur will meet. Additionally, the Table TABREF5 shows the parse trees for the above sentence with and without pre-ordering.\nPre-ordering should also be beneficial for other word order divergence scenarios (e.g., SOV to SVO), but we leave verification of these additional scenarios for future work.\nExperimental Setup\nIn this section, we describe the languages experimented with, datasets used, the network hyper-parameters used in our experiments.\nLanguages\nWe experimented with English INLINEFORM0 Hindi translation as the parent task. English is the assisting source language. Bengali, Gujarati, Marathi, Malayalam and Tamil are the primary source languages, and translation from these to Hindi constitute the child tasks. Hindi, Bengali, Gujarati and Marathi are Indo-Aryan languages, while Malayalam and Tamil are Dravidian languages. All these languages have a canonical SOV word order.\nDatasets\nFor training English-Hindi NMT systems, we use the IITB English-Hindi parallel corpus BIBREF22 ( INLINEFORM0 sentences from the training set) and the ILCI English-Hindi parallel corpus ( INLINEFORM1 sentences). The ILCI (Indian Language Corpora Initiative) multilingual parallel corpus BIBREF23 spans multiple Indian languages from the health and tourism domains. We use the 520-sentence dev-set of the IITB parallel corpus for validation. For each child task, we use INLINEFORM2 sentences from ILCI corpus as the test set.\nNetwork\nWe use OpenNMT-Torch BIBREF24 to train the NMT system. We use the standard sequence-to-sequence architecture with attention BIBREF13 . We use an encoder which contains two layers of bidirectional LSTMs with 500 neurons each. The decoder contains two LSTM layers with 500 neurons each. Input feeding approach BIBREF1 is used where the previous attention hidden state is fed as input to the decoder LSTM. We use a mini-batch of size 50 and use a dropout layer. We begin with an initial learning rate of INLINEFORM0 and decay the learning rate by a factor of INLINEFORM1 when the perplexity on validation set increases. The training is stopped when the learning rate falls below INLINEFORM2 or number of epochs=22. The English input is initialized with pre-trained embeddings trained using fastText BIBREF25 .\nEnglish vocabulary consists of INLINEFORM0 tokens appearing at least 2 times in the English training corpus. For constructing the Hindi vocabulary we considered only those tokens appearing at least 5 times in the training split resulting in a vocabulary size of INLINEFORM1 tokens. For representing English and other source languages into a common space, we translate each word in the source language into English using a bilingual dictionary (Google Translate word translation in our case). In an end-to-end solution, it would have been ideal to use bilingual embeddings or obtain word-by-word translations via bilingual embeddings BIBREF14 . But, the quality of publicly available bilingual embeddings for English-Indian languages is very low for obtaining good-quality, bilingual representations BIBREF26 , BIBREF27 . We also found that these embeddings were not useful for transfer learning.\nWe use the CFILT-preorder system for reordering English sentences to match the Indian language word order. It contains two re-ordering systems: (1) generic rules that apply to all Indian languages BIBREF17 , and (2) hindi-tuned rules which improve the generic rules by incorporating improvements found through an error analysis of English-Hindi reordering BIBREF28 . These Hindi-tuned rules have been found to improve reordering for many English to Indian language pairs BIBREF29 .\nResults\nIn this section, we describe the results from our experiments on NMT task. We report the results on X-Hindi pair, where X is one of Bengali, Gujarati, Marathi, Tamil, and Malayalam. The results are presented in the Table TABREF6 . We report BLEU scores and LeBLEU scores BIBREF30 . We observe that both the pre-ordering configurations significantly improve the BLEU scores over the baseline scores. We observe larger gains when generic pre-ordering rules are used compared to the Hindi-tuned pre-ordering rules.\nThese results support our hypothesis that word-order divergence can limit the benefits of multilingual translation. Reducing the word order divergence can improve translation in extremely low-resource scenarios.\nAn analysis of the outputs revealed that pre-ordering significantly reducing the number of UNK tokens (placeholder for unknown words) in the test output (Table TABREF14 ). We hypothesize that due to word order divergence between English and Indian languages, the encoder representation generated is not consistent leading to decoder generating unknown words. However, the pre-ordered models generate better contextual representations leading to less number of unknown tokens and better translation which is also reflected in the BLEU scores.\nConclusion\nIn this paper, we show that handling word-order divergence between source and assisting languages is crucial for the success of multilingual NMT in an extremely low-resource setting. We show that pre-ordering the assisting language to match the word order of the source language significantly improves translation quality in an extremely low-resource setting. While the current work focused on Indian languages, we would like to validate the hypothesis on a more diverse set of languages.",6 "char_len": 14863,7 "document_hash": "b1a7f83050f7b33dedccbbc2875d6413f0c64d7ed832ee7f5b0cf2f7cd10286e",8 "questions": [9 {10 "question_id": "q00",11 "qtype": "qasper",12 "question": "How do they match words before reordering them?"13 }14 ]15}