Please help us test our new pre-print finding feature by giving the pre-print link a rating. A 5 star rating indicates the linked pre-print has the exact same content as the published article.
Abstract: Abstract Unsupervised Neural Machine Translation (UNMT) approaches have gained widespread popularity in recent times. Though these approaches show impressive translation performance using only monolingual corpora of the languages involved, these approaches have mostly been tried on high-resource European language pairs viz. English–French, English–German, etc. In this paper, we explore UNMT for 6 Indic language pairs viz., Hindi–Bengali, Hindi–Gujarati, Hindi–Marathi, Hindi–Malayalam, Hindi–Tamil, and Hindi–Telugu which are low-resource language pairs. We additionally perform experiments on 4 European language pairs viz., English–Czech, English–Estonian, English–Lithuanian, and English–Finnish. We observe that the lexical divergence within these language pairs plays a big role in the success of UNMT. In this context, we explore three approaches viz., (i) script conversion, (ii) unsupervised bilingual embedding-based initialization to bring the vocabulary of the two languages closer, and (iii) dictionary word substitution using a bilingual dictionary. We found that the script conversion using a simple rule-based system benefits language pairs that have high cognate overlap but use different scripts. We observe that script conversion combined with word substitution using a dictionary further improves the UNMT performance. We use a ground truth bilingual dictionary in our dictionary word substitution experiments, and such dictionaries can also be obtained using unsupervised bilingual embeddings. We empirically demonstrate that minimizing lexical divergence using simple heuristics leads to significant improvements in the BLEU score for both related and distant language pairs. PubDate: 2022-01-05 DOI: 10.1007/s10590-021-09292-y
Please help us test our new pre-print finding feature by giving the pre-print link a rating. A 5 star rating indicates the linked pre-print has the exact same content as the published article.
Abstract: Abstract We investigated multiple pivot approaches for the Japanese and Indonesian (Ja–Id) language pair in phrase-based statistical machine translation (SMT). We used four languages as pivots: viz., English, Malaysian, Filipino, and the Myanmar language. Considering that each language pair between the source–pivot and pivot–target has a different word order, we conducted two experiments, namely, without reordering (WoR) and with reordering (WR) on the source language. Triangulation and linear interpolation (LI) approaches were used to combine multiple pivot phrase tables. The combination of phrase tables was employed without a source–target phrase table. In the WoR experiment, the use of multiple pivots improved the BLEU scores by 0.24 and 2.49 compared to the baseline and single pivot, respectively. However, the translation output of WoR was incomprehensible because it followed the Japanese word order. In the WR experiment, we reordered the Japanese word order, that is, subject–object–verb (SOV), into Indonesian word order, that is, subject–verb–object (SVO) using the Lader (Latent Derivation Reorderer). The multiple pivots of WR improved the BLEU scores by 0.47 compared with the baseline. Furthermore, by combining many pivot languages, the BLEU score was improved by more than 0.20. The translation output of WR is also more comprehensible than that of WoR. Finally, a comparison with neural machine translation (NMT) indicates that SMT obtained better results than NMT in the experiments, including a small dataset setup. PubDate: 2022-01-04 DOI: 10.1007/s10590-021-09288-8
Please help us test our new pre-print finding feature by giving the pre-print link a rating. A 5 star rating indicates the linked pre-print has the exact same content as the published article.
Abstract: Abstract Neural machine translation (NMT) has emerged as a preferred alternative to the previous mainstream statistical machine translation (SMT) approaches largely due to its ability to produce better translations. The NMT training is often characterized as data hungry since a lot of training data, in the order of a few million parallel sentences, is generally required. This is indeed a bottleneck for the under-resourced languages that lack the availability of such resources. The researchers in machine translation (MT) have tried to solve the problem of data sparsity by augmenting the training data using different strategies. In this paper, we propose a generalized linguistically motivated data augmentation approach for NMT taking low-resource translation into consideration. The proposed method operates by generating source—target phrasal segments from an authentic parallel corpus, whose target counterparts are linguistic phrases extracted from the syntactic parse trees of the target-side sentences. We augment the authentic training corpus with the parser generated phrasal-segments, and investigate the efficacy of our proposed strategy in low-resource scenarios. To this end, we carried out experiments with resource-poor language pairs, viz. Hindi-to-English, Malayalam-to-English, and Telugu-to-English, considering the three state-of-the-art NMT paradigms, viz. attention-based recurrent neural network (Bahdanau et al., 2015), Google Transformer (Vaswani et al. 2017) and convolution sequence-to-sequence (Gehring et al. 2017) neural network models. The MT systems built on the training data prepared with our data augmentation strategy significantly surpassed the state-of-the-art NMT systems with large margins in all three translation tasks. Further, we tested our approach along with back-translation (Sennrich et al. 2016a), and found these to be complementary to each other. This joint approach has turned out to be the best-performing one in our low-resource experimental settings. PubDate: 2021-12-21 DOI: 10.1007/s10590-021-09290-0
Please help us test our new pre-print finding feature by giving the pre-print link a rating. A 5 star rating indicates the linked pre-print has the exact same content as the published article.
Abstract: Abstract The pervasive attention model is a sequence-to-sequence model that addresses the issue of source–target interaction in encoder–decoder models by jointly encoding the two sequences with a two-dimensional convolutional neural network. We investigate different design choices for each building block of Pervasive Attention and study their impact to improve the predictive strength of the model. These include different types of layer connectivity, depth of the networks, the filter sizes, and source aggregation mechanisms. Machine translation experiments on the IWSLT’14 De \(\rightarrow\) En, IWSLT’15 En \(\rightarrow\) Vi, WMT’16 En \(\rightarrow\) Ro and WMT’15 De \(\rightarrow\) En datasets show results competitive with state-of-the-art encoder–decoder models, outperforming Transformer models on three of the four tested datasets. PubDate: 2021-12-10 DOI: 10.1007/s10590-021-09289-7
Please help us test our new pre-print finding feature by giving the pre-print link a rating. A 5 star rating indicates the linked pre-print has the exact same content as the published article.
Abstract: Abstract Parallel corpora are central to translation studies and contrastive linguistics. However, training machine translation (MT) systems by barely using the semantic aspects of a parallel corpus leads to unsatisfactory results, as then the trained MT systems are likely to generate target sentences that are semantically and pragmatically different from the source sentence. In the present work, we explore the improvement in the performance of an MT system when pragmatic features such as sentiment are introduced during its development. The language pair used for the experiments is English (source language) and Bengali (target language). The improvement in the MT output, before and after the introduction of sentiment features, is quantified by comparing various translation models, such as SMT, NMT and a newly developed translation model SeNA, with the help of automated (BLEU and TER) and manual evaluation metrics. In addition, the propagation of sentiment during the translation process is also studied extensively. We observe that the introduction of sentiment features during the system development process helps in elevating the translation quality. PubDate: 2021-12-09 DOI: 10.1007/s10590-021-09291-z
Please help us test our new pre-print finding feature by giving the pre-print link a rating. A 5 star rating indicates the linked pre-print has the exact same content as the published article.
Abstract: Abstract Bilingual word embeddings (BWEs) play a very important role in many natural language processing (NLP) tasks, especially cross-lingual tasks such as machine translation (MT) and cross-language information retrieval. Most existing methods to train BWEs are based on bilingual supervision. However, bilingual resources are not available for many low-resource language pairs. Although some studies addressed this issue with unsupervised methods, monolingual contextual data are not used to improve the performance of low-resource BWEs. To address these issues, we propose an unsupervised method to improve BWEs using optimized monolingual context information without any parallel corpora. In particular, we first build a bilingual word embeddings mapping model between two languages by aligning monolingual word embedding spaces based on unsupervised adversarial training. To further improve the performance of these mappings, we use monolingual context information to optimize them during the course. Experimental results show that our method outperforms other baseline systems significantly, including results for four low-resource language pairs. PubDate: 2021-12-01 DOI: 10.1007/s10590-021-09274-0
Please help us test our new pre-print finding feature by giving the pre-print link a rating. A 5 star rating indicates the linked pre-print has the exact same content as the published article.
Please help us test our new pre-print finding feature by giving the pre-print link a rating. A 5 star rating indicates the linked pre-print has the exact same content as the published article.
Please help us test our new pre-print finding feature by giving the pre-print link a rating. A 5 star rating indicates the linked pre-print has the exact same content as the published article.
Abstract: Abstract Non-autoregressive machine translation aims to speed up the decoding procedure by discarding the autoregressive model and generating the target words independently. Because non-autoregressive machine translation fails to exploit target-side information, the ability to accurately model source representations is critical. In this paper, we propose an approach to enhance the encoder’s modeling ability by using a pre-trained BERT model as an extra encoder. With a different tokenization method, the BERT encoder and the Raw encoder can model the source input from different aspects. Furthermore, having a gate mechanism, the decoder can dynamically determine which representations contribute to the decoding process. Experimental results on three translation tasks show that our method can significantly improve the performance of non-autoregressive MT, and surpass the baseline non-autoregressive models. On the WMT14 EN \(\rightarrow\) DE translation task, our method achieves 27.87 BLEU with a single decoding step. This is a comparable result with the baseline autoregressive Transformer model which obtains a score of 27.8 BLEU. PubDate: 2021-11-16 DOI: 10.1007/s10590-021-09285-x
Please help us test our new pre-print finding feature by giving the pre-print link a rating. A 5 star rating indicates the linked pre-print has the exact same content as the published article.
Abstract: Abstract An effective method to generate a large number of parallel sentences for training improved neural machine translation (NMT) systems is the use of the back-translations of the target-side monolingual data. The standard back-translation method has been shown to be unable to efficiently utilize huge amounts of existing monolingual data because of the inability of translation models to differentiate between authentic and synthetic parallel data during training. Tagging, or using gates, has been used to enable translation models to distinguish between synthetic and authentic data, improving standard back-translation and also enabling the use of iterative back-translation on language pairs that underperformed using standard back-translation. In this work, we approach back-translation as a domain adaptation problem, eliminating the need for explicit tagging. In our approach—tag-less back-translation—the synthetic and authentic parallel data are treated as out-of-domain and in-domain data, respectively, and through pre-training and fine-tuning, the translation model is shown to be able to learn more efficiently from them during training. Experimental results have shown that the approach outperforms the standard and tagged back-translation approaches on low resource English-Vietnamese and English-German NMT. PubDate: 2021-11-02 DOI: 10.1007/s10590-021-09284-y
Please help us test our new pre-print finding feature by giving the pre-print link a rating. A 5 star rating indicates the linked pre-print has the exact same content as the published article.
Abstract: Abstract Learning bilingual word embeddings can be much easier if the parallel corpora are available with their words well aligned explicitly. However, in most cases, the parallel corpora only provide a set of pairs that are semantically equivalent to each other at sentence level. While algorithms have been proposed to obtain word alignments, good alignments are still hard to achieve. In this study, we propose Bilingual word embeddings with soft alignment (BWESA) to learn bilingual word representations from the parallel corpora without explicit word-level alignment information. At the same time, this method learns to make ‘soft’ alignments between words by approximating a distribution for each word in a sentence to estimate how likely the word is aligned to the words in the parallel translation. Unlike previous methods that typically make use of a predetermined word alignment, our learning strategy makes similar words—properly chosen by the continuously improving word alignment—become closer in the shared vector space during the training process. This study is among the first to learn bilingual word alignments and embeddings in a joint manner. The proposed method was evaluated on two cross-lingual tasks (cross-lingual document classification and word translation) and achieved state-of-the-art or comparable results on all the tasks considered. PubDate: 2021-11-01 DOI: 10.1007/s10590-021-09283-z
Please help us test our new pre-print finding feature by giving the pre-print link a rating. A 5 star rating indicates the linked pre-print has the exact same content as the published article.
Abstract: Abstract In recent years, neural network-based machine translation (MT) approaches have steadily superseded the statistical MT (SMT) methods, and represents the current state-of-the-art in MT research. Neural MT (NMT) is a data-driven end-to-end learning protocol whose training routine usually requires a large amount of parallel data in order to build a reasonable-quality MT system. This is particularly problematic for those language pairs that do not have enough parallel text for training. In order to counter the data sparsity problem of the NMT training, MT researchers have proposed various strategies, e.g. augmenting training data, exploiting training data from other languages, alternative learning strategies that use only monolingual data. This paper presents a survey on recent advances of NMT research from the perspective of low-resource scenarios. PubDate: 2021-10-30 DOI: 10.1007/s10590-021-09281-1
Please help us test our new pre-print finding feature by giving the pre-print link a rating. A 5 star rating indicates the linked pre-print has the exact same content as the published article.
Abstract: Abstract This paper presents an overview of Apertium, a free and open-source rule-based machine translation platform. Translation in Apertium happens through a pipeline of modular tools, and the platform continues to be improved as more language pairs are added. Several advances have been implemented since the last publication, including some new optional modules: a module that allows rules to process recursive structures at the structural transfer stage, a module that deals with contiguous and discontiguous multi-word expressions, and a module that resolves anaphora to aid translation. Also highlighted is the hybridisation of Apertium through statistical modules that augment the pipeline, and statistical methods that augment existing modules. This includes morphological disambiguation, weighted structural transfer, and lexical selection modules that learn from limited data. The paper also discusses how a platform like Apertium can be a critical part of access to language technology for so-called low-resource languages, which might be ignored or deemed unapproachable by popular corpus-based translation technologies. Finally, the paper presents some of the released and unreleased language pairs, concluding with a brief look at some supplementary Apertium tools that prove valuable to users as well as language developers. All Apertium-related code, including language data, is free/open-source and available at https://github.com/apertium. PubDate: 2021-10-18 DOI: 10.1007/s10590-021-09260-6
Please help us test our new pre-print finding feature by giving the pre-print link a rating. A 5 star rating indicates the linked pre-print has the exact same content as the published article.
Abstract: Abstract Self-attention-based encoder-decoder frameworks have drawn increasing attention in recent years. The self-attention mechanism generates contextual representations by attending to all tokens in the sentence. Despite improvements in performance, recent research argues that the self-attention mechanism tends to concentrate more on the global context with less emphasis on the contextual information available within the local neighbourhood of tokens. This work presents the Dual Contextual (DC) module, an extension of the conventional self-attention unit, to effectively leverage both the local and global contextual information. The goal is to further improve the sentence representation ability of the encoder and decoder subnetworks, thus enhancing the overall performance of the translation model. Experimental results on WMT’14 English-German (En \(\rightarrow \) De) and eight IWSLT translation tasks show that the DC module can further improve the translation performance of the Transformer model. PubDate: 2021-10-12 DOI: 10.1007/s10590-021-09282-0
Please help us test our new pre-print finding feature by giving the pre-print link a rating. A 5 star rating indicates the linked pre-print has the exact same content as the published article.
Please help us test our new pre-print finding feature by giving the pre-print link a rating. A 5 star rating indicates the linked pre-print has the exact same content as the published article.
Please help us test our new pre-print finding feature by giving the pre-print link a rating. A 5 star rating indicates the linked pre-print has the exact same content as the published article.
Abstract: Abstract Existing work on the animation of signing avatars often relies on pure procedural techniques or on the playback of Motion Capture (MoCap) data. While the first solution results in robotic and unnatural motions, the second one is very limited in the number of signs that it can produce. In this paper, we propose to implement data-driven motion synthesis techniques to increase the variety of Sign Language (SL) motions that can be made from a limited database. In order to generate new signs and inflection mechanisms based on an annotated French Sign Language MoCap corpus, we rely on phonological recombination, i.e. on the motion retrieval and modular reconstruction of SL content at a phonological level with a particular focus on three phonological components of SL: hand placement, hand configuration and hand movement. We propose to modify the values taken by those components in different signs to create their inflected version or completely new signs by (i) applying motion retrieval at a phonological level to exchange the value of one component without any modification, (ii) editing the retrieved data with different operators, or, (iii) using conventional motion generation techniques such as interpolation or inverse kinematics, which are parameterized to comply to the kinematic properties of real motion observed in the data set. The quality of the synthesized motions is perceptually assessed through two distinct evaluations that involved 75 and 53 participants respectively. PubDate: 2021-09-01 DOI: 10.1007/s10590-021-09268-y
Please help us test our new pre-print finding feature by giving the pre-print link a rating. A 5 star rating indicates the linked pre-print has the exact same content as the published article.
Abstract: Abstract Fingerspelling is a process of communicating letters of a spoken language alphabet using a person’s hand or hands. Portraying animations of fingerspelling has proved surprisingly resistant to automation because of the collisions that arise from conventional interpolation of keyframes of individual manual letters. Previous methods have not been able to provide convincingly realistic fingerspelling due to the absence of effective collision avoidance in the underlying animation algorithms. This paper reports on the development and evaluation of a new collision avoidance algorithm that aids fingerspelling. Instead of analyzing letter transitions, the algorithm capitalizes on the transitions of individual fingers. The new strategy is efficient enough to support real-time fingerspelling while still maintaining a high level of predictive accuracy. Utilizing this strategy in signing avatars is expected to improve the current resources for deaf children, hearing teachers, hearing parents, and interpreting students who want to improve their fingerspelling comprehension. Future work will include testing the strategy’s generality when applying it to other one-handed manual alphabets. PubDate: 2021-09-01 DOI: 10.1007/s10590-021-09273-1
Please help us test our new pre-print finding feature by giving the pre-print link a rating. A 5 star rating indicates the linked pre-print has the exact same content as the published article.
Abstract: Abstract Professional Sign Language translators, unlike their text-to-text counterparts, are not equipped with computer-assisted translation (CAT) software. Those softwares are meant to ease the translators’ tasks. No prior study as been conducted on this topic, and we aim at specifying such a software. To do so, we based our study on the professional Sign Language translators’ practices and needs. The aim of this paper is to identify the necessary steps in the text-to-sign translation process. By filming and interviewing professionals for both objective and subjective data, we build a list of tasks and see if they are systematic and performed in a definite order. Finally, we reflect on how CAT tools could assist those tasks, how to adapt the existing tools to Sign Language and what is necessary to add in order to fit the needs of Sign Language translation. In the long term, we plan to develop a first prototype of CAT software for sign languages. PubDate: 2021-09-01 DOI: 10.1007/s10590-021-09278-w
Please help us test our new pre-print finding feature by giving the pre-print link a rating. A 5 star rating indicates the linked pre-print has the exact same content as the published article.
Abstract: Abstract Facial nonmanual signals and expressions convey critical linguistic and affective information in signed languages. However, the complexity of human facial anatomy has made the implementation of these movements a particular challenge in avatar research. Recent advances have improved the possible range of motion and expression. Because of this, we propose that an important next step is incorporating fine detail such as wrinkles to increase the visual clarity of these facial movements for the purposes of enhancing the legibility of avatar animation, particularly on small screens. This paper reviews research efforts to portray nonmanual signals via avatar technology and surveys extant illumination models for their suitability for this application. Based on this information, The American Sign Language Avatar Project at DePaul University has developed a new technique based on commercial visual effects paradigms for implementing realistic fine detail on the Paula avatar that functions within the complexity constraints of real-time sign language avatars. PubDate: 2021-09-01 DOI: 10.1007/s10590-021-09269-x
Hybrid journal (It can contain Open Access articles)
[2468 journals]
