No full-text available
To read the full-text of this research,
you can request a copy directly from the authors.
Automatic Machine Translation of Poetry and a Low-Resource Language Pair
... Ghazvininejad et al. [15] was among the irst studies to focus on translating prose into poetry. They introduced an initial model using the encoder-decoder architecture and two biased decoding models where they encouraged the decoding of certain words.Similarly, Dunder et al. [12] also addressed poetry translation, but among two diferent languages, Croatian and German. They examined and demonstrated the usability of statistical and neural machine translation techniques in a low-resource setting. ...
... Then we iterate couplets to extract unique word combinations of two and draw an edge between them at their earliest co-occurrence. Next, given v 1 and v 2 as two connected vertices and v i 1 and v i 2 as their indices (ranging from 1 to 20), we deine edge attribute E(v 1 , v 2 ) using Equation 12, which scores the two vertices co-occurrence by their spatial distance. ...
Persian Poetry has consistently expressed its philosophy, wisdom, speech, and rationale based on its couplets, making it an enigmatic language on its own to both native and non-native speakers. Nevertheless, the noticeable gap between Persian prose and poem has left the two pieces of literature medium-less. Having curated a parallel corpus of prose and their equivalent poems, we introduce a novel Neural Machine Translation (NMT) approach for translating prose to ancient Persian poetry using transformer-based language models in an exceptionally low-resource setting. Translating input prose into ancient Persian poetry presents two primary challenges: In addition to being reasonable in conveying the same context as the input prose, the translation must also satisfy poetic standards. Hence, we designed our method consisting of three stages. First, we trained a transformer model from scratch to obtain an initial translations of the input prose. Next, we designed a set of heuristics to leverage contextually-rich initial translations and produced a poetic masked template. In the last stage, we pretrained different variations of BERT on a poetry corpus to use the masked language modelling technique to obtain final translations. During the evaluation process, we considered both automatic and human assessment. The final results demonstrate the eligibility and creativity of our novel heuristically aided approach among Literature professionals and non-professionals in generating novel Persian poems.
... Stimulated by advances in neural machine translation, there has emerged a body of empirical research on literary machine translation. Most studies focus on the quality of raw machine translations of poetry (Greene et al., 2010;Genzel et al., 2010;Almahasees, 2017;Humblé, 2019;Dunđer et al., 2020Dunđer et al., , 2021Seljan et al., 2020) and prose (Voigt and Jurafsky, 2012;Jones and Irvine, 2013;Way, 2015a, 2015b;Kuzman et al., 2019;Matusov, 2019;Tezcan et al., 2019;Toral et al., 2020;Webster et al., 2020;Jiang and Niu, 2022). In comparison, the post-editing of literary machine translation remains underexplored. ...
Despite the advances in the automation of translation processes, several discourses insist that there is still a privileged status for the non-automated human, especially in idealized conceptualisations of literary translation. Empirical data from experiments with 141 students show general awareness of trade-offs between the advantages and disadvantages of post-editing texts by Agatha Christie. Although machine translation was assumed to suffer from excessive literalism, it was also criticized for being deceptively fluent. Comparison of work by post-editors and from-scratch translators shows that the former tends to use risk transfer, trusting machine translation in situations of uncertainty, while the latter are more prone to risk taking. The two strategies lead to errors of different kinds, with the from-scratch translations having significantly fewer errors than the results of post-editing, with the exception of the “deviations” resulting from excessive risk taking.
... With the continuous deepening of globalization and the rapid development of information technology [1], machine translation, as an important language communication tool, plays an increasingly important role in cross-cultural communication. However, although the development of deep learning technology has made significant progress in machine translation in recent years, there are still challenges in evaluating translation quality [2][3]. Traditional manual evaluation methods are not only time-consuming and laborintensive, but may also have subjective biases [4]. ...
Nowadays, machine translation has been a prevalent Internet application. But there still lacks mature intelligent algorithms to automatically evaluate quality of machine translation results. Considering the complexity inside machine intelligence-based semantic comprehension, we resort to pretraining language model (PLM) to deal with this challenge. Hence, this paper proposes a semantic context context-aware automatic quality scoring method for machine translation based on a specific PLM. The purpose of introducing the calculation of sentiment vectors in research is to consider emotional information in machine translation quality automatic scoring methods, in order to improve the accuracy and robustness of scoring. In particular, a novel PLM that combines multiple key features and tasks is established, which is utilized to make encoding towards largescale initial sentences and object sentences. It is finely tuned by integrating two typical pretraining structures. By applying the proposed PLM to complex semantic context and analysis tasks, we finally demonstrate its effectiveness through experiments on the News Crawl corpus and WMT dataset. The obtained results show that the proposal method has achieved significant improvements in various evaluation indicators, demonstrating its superiority in the quality evaluation of machine translation by perceiving semantic contexts. Through comparison experiments, efficiency of the proposal can be acknowledged.
... Machine translation research in low-resource and professional fields has also been a hot topic in recent years. Dungyer et al. [27] created a data set that includes the works of a Croatian-related contemporary poet and the translation of German poetry by two professional literary translators. The research results show the effectiveness of poetry machine translation in terms of special automatic quality indicators. ...
In practical applications, the accuracy of domain terminology translation is an important criterion for the performance evaluation of domain machine translation models. Aiming at the problem of phrase mismatch and improper translation caused by word-by-word translation of English terminology phrases, this paper constructs a dictionary of terminology phrases in the field of electrical engineering and proposes three schemes to integrate the dictionary knowledge into the translation model. Scheme 1 replaces the terminology phrases of the source language. Scheme 2 uses the residual connection at the encoder end after the terminology phrase is replaced. Scheme 3 uses a segmentation method of combining character segmentation and terminology segmentation for the target language and uses an additional loss module in the training process. The results show that all three schemes are superior to the baseline model in two aspects: BLEU value and correct translation rate of terminology words. In the test set, the highest accuracy of terminology words was 48.3% higher than that of the baseline model. The BLEU value is up to 3.6 higher than the baseline model. The phenomenon is also analyzed and discussed in this paper.
... In the context of the Croatian language and available tools and resources, machine learning has been especially explored with a special focus on machine translation (Dunđer, 2020;Dunđer et al., 2020) and document classification (Dunđer et al., 2015). ...
Machine learning, and more specifically its subfield known as deep learning, have been driving new disruptive technologies and recent accomplishments in various industries. This paper applies to some extent the same technology to develop a mobile application for educational purposes, that allows one to identify unknown flags with the help of artificial intelligence. For the application to be functional and effective, a custom dataset had to be created and processed in various ways in order to provide a collection of data that resembles data from the real world. The collection would then be used as the input data for constructing the model, which is developed within the framework TensorFlow. Once the model was developed, it was implemented as part of a mobile application programmed with Flutter. The functionality of the mobile application and the model’s accuracy are then put to the test against over a hundred new images the model has not seen previously or been trained on. The result of this evaluation could be considered an estimate of the readiness and usability of the application in a real-world scenario.
... In the context of the Croatian language and available tools and resources, machine learning has been especially explored with a special focus on machine translation (Dunđer, 2020;Dunđer et al., 2020) and document classification (Dunđer et al., 2015). ...
Strojno učenje te posebno pripadajuće potpodručje poznato kao duboko učenje pokreću novedisruptivne tehnologije i nedavna postignuća u različitim industrijama. Ovaj rad koristi do određenerazine istu tehnologiju pri izradi mobilne aplikacije za edukacijske svrhe, a koja omogućuje identifikacijunepoznatih zastava pomoću umjetne inteligencije. Kako bi aplikacija bila funkcionalna i efikasna,potrebno je bilo kreirati vlastiti podatkovni skup i obraditi ga na razne načine kako bi se osiguralazbirka podataka koja odražava podatke iz stvarnoga svijeta. Ova zbirka bi se potom koristila kaopodatkovni ulaz za razvijanje modela, koji se razvija unutar okruženja TensorFlow. Nakon što je modelrazvijen, implementiran je kao dio mobilne aplikacije programirane Flutterom. Funkcionalnost mobilneaplikacije i točnost modela ispitana je na brojci od preko 100 novih slika koje model prethodno nijevidio niti je nad njima bio treniran. Rezultat ovog ispitivanja mogao bi se smatrati odrazom okvirnespremnosti i primjenjivosti ove aplikacije u stvarnom svijetu.
... This study aims to illustrate the melodiousness of the Poem and set up a standard of how to compose a poem melodiously. We found some researcher works for poem domain such as classify the poem [3]- [6], extract features poem [7], poetry generation [8]- [10], evaluation poem [11], translated poetry [12]- [14], poem entity recognition [15], and analysis of the melodiousness of the Poem [16]- [18]. Nobody has researched extracting the melodious sound patterns before. ...
The melodious poems have been written from the distinctive features of poetry or based on each country's typical style. Especially, Thai poems which composed by the use of specific forming, such as Internal Rhyme to develop melodiousness. The most attractive and well-known poems were composed by a genius Thai poet named Sunthorn Phu. He is a role model for Thai poets. UNESCO honored him as the world’s great poet and the best role model in poetry works. In this article, we proposed extracting 15,796 sentences (Waks) of the beautiful sound patterns of Phra Aphai Mani’s tales by machine learning technology in conjunction with the rules of internal Rhyme Klon-Suphap by using the Apriori Algorithm. The extraction of vowel rhymes separated by a group of Waks including 1) Poem Wak No. 1; 2) Poem Wak No. 2; 3) Poem Wak No. 3; and 4) Poem Wak No. 4. In this article, “Wak” means sentence. The created tool can extract the internal rhyme patterns and the 25 popular pattern vowels. The popular pattern illustrates the melodiousness of the Poem and sets up a standard of how to melodiously compose a poem. Then, the evaluation of the experiments was done by using 144 Waks selected from the extraction of the beautiful patterns and evaluated by the consistency score from 3 experts. The average accuracy score resulted in 95.30%.
... One of the problems in machine translation is the lack of training data. This problem was reported by Seljan [17] and Dunder [18,19] for the problem of the automatic translation of poetry with a low-resource language pair. It was reported that the fluency and adequacy of the translation results were skewed to higher scores. ...
Featured Application
machine translation; information retrieval; text-to-speech.
Abstract
Neural machine translation (NMT) methods based on various artificial neural network models have shown remarkable performance in diverse tasks and have become mainstream for machine translation currently. Despite the recent successes of NMT applications, a predefined vocabulary is still required, meaning that it cannot cope with out-of-vocabulary (OOV) or rarely occurring words. In this paper, we propose a postprocessing method for correcting machine translation outputs using a named entity recognition (NER) model to overcome the problem of OOV words in NMT tasks. We use attention alignment mapping (AAM) between the named entities of input and output sentences, and mistranslated named entities are corrected using word look-up tables. The proposed method corrects named entities only, so it does not require retraining of existing NMT models. We carried out translation experiments on a Chinese-to-Korean translation task for Korean historical documents, and the evaluation results demonstrated that the proposed method improved the bilingual evaluation understudy (BLEU) score by 3.70 from the baseline.
... Whereas distributed representation of arbitrary language can be realized through the end-to-end training of the NMT system, the NMT model can prevent the problems during the SMT training process. NMT can produce more fluent results [8][9][10] but which are often not adequate, while SMT models generally obtain lower results for the criteria of fluency, especially for low-resource languages with relatively free word order [11]. Generally, MT evaluation metrics favor translations that follow a strict word order when compared to the reference translations, which could be the reason for lower BLEU scores. ...
Both the statistical machine translation (SMT) model and neural machine translation (NMT) model are the representative models in Uyghur–Chinese machine translation tasks with their own merits. Thus, it will be a promising direction to combine the advantages of them to further improve the translation performance. In this paper, we present a hybrid framework of developing a system combination for a Uyghur–Chinese machine translation task that works in three layers to achieve better translation results. In the first layer, we construct various machine translation systems including SMT and NMT. In the second layer, the outputs of multiple systems are combined to leverage the advantage of SMT and NMT models by using a multi-source-based system combination approach and the voting-based system combination approaches. Moreover, instead of selecting an individual system’s combined outputs as the final results, we transmit the outputs of the first layer and the second layer into the final layer to make a better prediction. Experiment results on the Uyghur–Chinese translation task show that the proposed framework can significantly outperform the baseline systems in terms of both the accuracy and fluency, which achieves a better performance by 1.75 BLEU points compared with the best individual system and by 0.66 BLEU points compared with the conventional system combination methods, respectively.
... The research included results of the human evaluation and automatic metrics (WER and PER). In the research presented by Dunđer et al. (2020) the authors conducted an automatic evaluation of machine-translated Croatian-German and German-Croatian datasets, using BLEU, METEOR, RIBES and CharacTER metrics to evaluate machine translation at the corpus level for two online tools, showing deficiencies concerning dataset type and size and domain coverage. ...
Increased use of computer-assisted translation (CAT) technology in business settings with augmented amounts of tasks, collaborative work, and short deadlines give rise to errors and the need for quality assurance (QA). The research has three operational aims: 1) methodological framework for QA analysis, 2) comparative evaluation of four QA tools, 3) to justify introduction of QA into CAT process. The research includes building of translation memory, terminology extraction, and creation of terminology base. Error categorization is conducted by multidimensional quality (MQM) framework. The level of mistake is calculated considering detected, false, and not detected errors. Weights are assigned to errors (minor, major, or critical), penalties are calculated, and quality estimation for translation memory is given. Results show that process is prone to errors due to differences in error detection, harmonization, and error counting. Data analysis of detected errors leads to further data-driven decisions related to the quality of output results and improved efficacy of translation business process.
Translation technology plays a crucial role in facilitating multilingual communication, requiring specialized tools and expertise. This study explores the application of computer-assisted translation (CAT) tool, which integrates automatic machine translation, translation memories, and terminology databases, to assess their effectiveness compared to traditional translation methods. The research evaluates translation performance across metrics such as time efficiency, error frequency, translation similarity, and terminological accuracy.Two distinct translation processes were conducted, leveraging different tools, resources, and methodologies. A translation memory was built by collecting and aligning parallel corpora, alongside the creation of a term base. Results revealed that the use of CAT tools reduced translation time by 49% on average. Error analysis showed a marked decrease in spelling and grammatical errors when using CAT tools. Additionally, translation similarity with the reference text increased from 79% in traditional translation to 95% with the CAT tool. Terminological analysis highlighted a substantial reduction in terminology errors, although some errors persisted due to limitations in automatic machine translation. The findings emphasize the critical role of a well-maintained term base in achieving consistent terminology.
Text-to-image generation is a rapidly growing field that aims to generate images from textual descriptions. This paper provides a comprehensive overview of the latest trends and developments, highlighting their importance and relevance in various domains, such as art, photography, marketing, and learning. The paper describes and compares various text-to-image models and discusses the challenges and limitations of this field. The findings of this paper demonstrate that recent advancements in deep learning and computer vision have led to significant progress in text-to-image models, enabling them to generate high-quality images from textual descriptions. However, challenges such as ensuring the legality and ethical implications of the final products generated by these models need to be addressed. This paper provides insights into these challenges and suggests future directions for this field. In addition, this study emphasises the need for a sustainability-oriented approach in the text-to-image domain. As text-to-image models advance, it is crucial to conscientiously assess their impact on ecological, cultural, and societal dimensions. Prioritising ethical model use while being mindful of their carbon footprint and potential effects on human creativity becomes crucial for sustainable progress.
Automatsko strojno prevođenje sve je popularnija istraživačka tema u znanosti i raznim znanstvenim disciplinama, kao što su informacijske i komunikacijske znanosti, računarstvo, računalna lingvistika i sl. Razlog tome je prvenstveno to što danas omogućuje nezaobilaznu komunikaciju i brz prijenos informacija između različitih prirodnih jezika. To je posebno bitno za manje govorene jezike poput hrvatskoga, za koji još uvijek ne postoji dovoljan broj softverskih alata i digitalnih resursa potrebnih za razvoj specijaliziranih i kvalitetnih sustava za strojno prevođenje koji bi bili optimizirani za upotrebu u jednom specifičnom području. Sve brži rast količine podataka i sve veća potreba raznih dionika u sektorima industrije, gospodarstvu, znanosti, ali i u svakodnevnom životu ljudi impliciraju motivaciju za sistematiziranim i organiziranim razvojem te naknadnom prilagodbom sustava za automatsko strojno prevođenje za različite jezične parove. Budući da strojni prijevodi nisu savršeni, važno je primijeniti metode za računalno generiranje prijevoda prihvatljive razine kvalitete koja ovisi o samom zadatku i području primjene sustava za strojno provođenje. U ovom radu analiziran je model sustava za automatsko statističko strojno prevođenje, njegove komponente te uloga i značaj pojedinih elemenata unutar modela.
This article investigates essential ethical issues that should be taken into consideration when adopting or tailoring technological tools for literary translation. The discussion on ethical issues draws on recent studies on translation technology and on the usage of machine(-assisted) translation for literary language. An overview of the consequences of the recent increase in technologization for both non-literary and literary translation is provided and an argument for sustainable development in literary translation is made, based on a holistic understanding of translation quality. The notion of voice is taken as an example of the special challenges related to the translation of literary language, which research on machine(-assisted) translation of literary language has not yet sufficiently taken into account. Lastly, avant-garde aesthetic views and usages of machine translation are presented through the notion of noise.
The poster presents 3 basic steps used in the system TMrepository used for collection of parallel data through crowdsourcing: - collecting and uploading parallel corpora - review and quality check - get top ranking (gamification)
This paper presents an idea to bring crowdsourcing to a higher level, for the purpose of acquiring valuable machine translation and natural language processing resources. In the proposed scenario, students are being educated in order to improve the quality and effectiveness of their natural language processing (NLP) related work. Their motivation is ensured by introducing an element of gamification-a ranking is kept, where the best contributing users are decorated with medals. The ranking is available at all times to all users and is always up-to-date, hence the effects of the contributions are immediately visible to the users. This scenario was applied to a group of students enrolled in Natural Language Processing course, who were presented with a task of collecting parallel corpora for less-resourced language pairs, in this case Croatian-English and English-Croatian. The whole experiment was supervised with the help of a custom-made open-source system named TMrepository, developed and maintained by the authors of this paper.
Automatic quality evaluation of machine translation systems has become an important issue in the field of natural language processing, due to raised interest and needs of industry and everyday users. Development of online machine translation systems is also important for less-resourced languages, as they enable basic information transfer and communication. Although the quality of free online automatic translation systems is not perfect, it is important to assure acceptable quality. As human evaluation is time-consuming, expensive and subjective, automatic quality evaluation metrics try to approach and approximate human evaluation as much as possible. In this paper, several automatic quality metrics will be utilised, in order to assess the quality of specific machine translated text. Namely, the research is performed on sociological-philosophical-spiritual domain, resulting from the digitisation process of a scientific publication written in Croatian and English. The quality evaluation results are discussed and further analysis is proposed.
In this research, a specific data set was machine translated by two publicly available machine translation services, Google Translate and Yandex.Translate. Machine translations were performed for two language pairs: English-Croatian and Russian-Croatian. Afterwards, automatic quality evaluation of the machine translated data set was carried out. Several automatic metrics were used: BLEU, NIST, METEOR and GTM, in order to evaluate machine translations relating to the domain of city description, for each language pair and for each machine translation service.
Automatic quality evaluation of machine translation systems has become an important issue in the field of natural language processing, due to raised interest and needs of industry and everyday users. Development of online machine translation systems is also important for less-resourced languages, as they enable basic information transfer and communication. Although the quality of free online automatic translation systems is not perfect, it is important to assure acceptable quality. As human evaluation is time-consuming, expensive and subjective, automatic quality evaluation metrics try to approach and approximate human evaluation as much as possible. In this paper, several automatic quality metrics will be utilised, in order to assess the quality of specific machine translated text. Namely, the research is performed on sociological-philosophical-spiritual domain, resulting from the digitisation process of a scientific publication written in Croatian and English. The quality evaluation results are discussed and further analysis is proposed.
This paper presents results of human evaluation of machine translated texts for one non closely-related language pair, English-Croatian, and for one closely-related language pair, Russian-Croatian. 400 sentences from the domain of tourist guides were analysed, i.e. 100 sentences for each language pair and for two online machine translation services, Google Translate and Yandex.Translate. Human evaluation is made with regard to the criteria of fluency and adequacy. In order to measure internal consistency, Cronbach’s alpha is calculated. Error analysis is made for several categories: untranslated/omitted words, surplus words, morphological errors/wrong word endings, lexical errors/wrong translations, syntactic errors/wrong word order and punctuation errors. At the end of this paper, conclusions and suggestions for further research are given.
The paper presents combined automatic speech recognition (ASR) of English and machine translation (MT) for English and Croatian and Croatian-English language pairs in the domain of business correspondence. The first part presents results of training the ASR commercial system on English data sets, enriched by error analysis. The second part presents results of machine translation performed by free online tool for English and Croatian and Croatian-English language pairs. Human evaluation in terms of usability is conducted and internal consistency calculated by Cronbach's alpha coefficient, enriched by error analysis. Automatic evaluation is performed by WER (Word Error Rate) and PER (Position-independent word Error Rate) metrics, followed by investigation of Pearson's correlation with human evaluation.
The poetry translation from original language to another is very different from general machine translator because the poem is written with prosody. Thai Poetry is composed with sets of syllables. Those rhymes, existing from stanzas, lines and the text in paragraph of the poetry, may not represent the complete syntax. This research is focus on Google and Bing machine translators and the tuning the prosody on syllable and rhyme. We have compared the errors (in percent), between the standard translators to those translators with tuning. The error rate of both translators before tune them, was at 97 % per rhyme. After tuning them, the percentage of errors decreased down to 60% per rhyme. To evaluate the meaning, concerning the gained results of both kinds of translators, we use BLEU (Bilingual Evaluation Understudy) metric to compare between reference and candidate. BLEU score of Google is 0.287 and Bing is 0.215. We can conclude that machine translators cannot provide good result for translate Thai poetry. This research should be the initial point for a new kind of innovative machine translators to Thai poetry. Furthermore, it is a way to encourage Thai art created language to the global public as well
This paper describes our submissions to the machine translation evaluation shared task in ACL WMT-08. Our primary submission is the Meteor metric tuned for optimizing correlation with human rankings of translation hypotheses. We show significant improvement in correlation as compared to the earlier version of metric which was tuned to optimized correlation with traditional adequacy and fluency judgments. We also describe m-bleu and m-ter, enhanced versions of two other widely used metrics bleu and ter respectively, which extend the exact word matching used in these metrics with the flexible matching based on stemming and Wordnet in Meteor.
Neural machine translation is a relatively new approach to statistical machine translation based purely on neural networks. The neural machine translation models often consist of an encoder and a decoder. The encoder extracts a fixed-length representation from a variable-length input sentence, and the decoder generates a correct translation from this representation. In this paper, we focus on analyzing the properties of the neural machine translation using two models; RNN Encoder--Decoder and a newly proposed gated recursive convolutional neural network. We show that the neural machine translation performs relatively well on short sentences without unknown words, but its performance degrades rapidly as the length of the sentence and the number of unknown words increase. Furthermore, we find that the proposed gated recursive convolutional network learns a grammatical structure of a sentence
automatically.
Monolingual and multilingual terminology and collocation bases represent valuable additional electronic resources, which can be used in further research, in written communication and in everyday communication. Building of such resources can be supported by terminology extraction tools relying on statistical or language approaches, or on hybrid model, but require considerable human expertise in evaluation and final compilation. The paper describes the whole process: from digitisation of printed material, OCR techniques, sentence alignment and creation of translation memories, up to terminology extraction and evaluation. The performance of tools and applied methodology is assessed through standard statistical measures of precision, recall and F-measure. Experimental results are produced, deficiencies of semi-automatic statistical and linguistic system highlighted and recommendations for further research suggested.
In the paper the evaluation of formant speech synthesis for Croatian is conducted in four domains: hotel reservation, insurance, automobile industry and weather forecast. Human evaluation is performed in order to evaluate quality of speech according to criteria of sentence comprehensibility, word intelligibility and correctness of word pronunciation. Automatic evaluation is conducted using word error rate metric across four domains and across specific lexical units (names, numbers, dates, general and special terminology). Correlation between human evaluation and word error rate metric is given by Pearson's correlation. The results are discussed and suggestions for further research mentioned.
Formant speech synthesis method mimics the time-varying formant frequencies of human speech and does not use prerecorded speech samples. In this paper, related work is discussed and an experiment is conducted using formant synthesis-based text-to-speech tool CroSS (Croatian Speech Synthesizer), in order to assess and evaluate the quality of synthesized Croatian speech, according to five criteria, then by domain suitability, affective attitudes and appropriateness of implementation in broader public use and in Computer-assisted Language Learning (CALL). The aim of integrating speech synthesis technology in Computer-assisted Language Learning resulted from the need to provide an interactive learning and teaching environment. This paper also addressed weaknesses and problems of Croatian language (e.g. input preprocessing of word classes) in the process of text-to-speech synthesis. The results are discussed and suggestions for further research mentioned.
We present two methods that merge ideas from statistical machine translation (SMT) and translation memories (TM). We use a TM to retrieve matches for source segments, and replace the mismatched parts with instructions to an SMT system to fill in the gap. We show that for fuzzy matches of over 70%, one method outperforms both SMT and TM base-lines.
Word order is essential in Japanese-to-English translation. But popular metrics do not regard global word order very much. Motivation Word order is essential in Japanese-to-English translation. But popular metrics do not regard global word order very much. Most SMT systems make mistakes like this. Japanese: (input) kare wa ame ni nureta node kaze wo hiitaMotivation Word order is essential in Japanese-to-English translation. But popular metrics do not regard global word order very much. Most SMT systems make mistakes like this. Japanese: (input) kare wa ame ni nureta node kaze wo hiita SMT output: (performs only local reordering) he got soaked in the rain because he caught a coldMotivation Word order is essential in Japanese-to-English translation.
We propose a new phrase-based translation model and decoding algorithm that enables us to evaluate and compare several, previously proposed phrase-based translation models. Within our framework, we carry out a large number of experiments to understand better and explain why phrase-based models out-perform word-based models. Our empirical results, which hold for all examined language pairs, suggest that the highest levels of performance can be obtained through relatively simple means: heuristic learning of phrase translations from word-based alignments and lexical weighting of phrase translations. Surprisingly, learning phrases longer than three words and learning phrases from high-accuracy word-level alignment models does not have a strong impact on performance. Learning only syntactically motivated phrases degrades the performance of our systems.
Human evaluations of machine translation are extensive but expensive. Human evaluations can take months to finish and involve human labor that can not be reused.
With the widespread adoption of deep learning, natural language processing (NLP),and speech applications in many areas (including Finance, Healthcare, and Government) there is a growing need for one comprehensive resource that maps deep learning techniques to NLP and speech and provides insights into using the tools and libraries for real-world applications. Deep Learning for NLP and Speech Recognition explains recent deep learning methods applicable to NLP and speech, provides state-of-the-art approaches, and offers real-world case studies with code to provide hands-on experience.
The book is organized into three parts, aligning to different groups of readers and their expertise. The three parts are:
Machine Learning, NLP, and Speech Introduction
The first part has three chapters that introduce readers to the fields of NLP, speech recognition, deep learning and machine learning with basic theory and hands-on case studies using Python-based tools and libraries.
Deep Learning Basics
The five chapters in the second part introduce deep learning and various topics that are crucial for speech and text processing, including word embeddings, convolutional neural networks, recurrent neural networks and speech recognition basics. Theory, practical tips, state-of-the-art methods, experimentations and analysis in using the methods discussed in theory on real-world tasks.
Advanced Deep Learning Techniques for Text and Speech
The third part has five chapters that discuss the latest and cutting-edge research in the areas of deep learning that intersect with NLP and speech. Topics including attention mechanisms, memory augmented networks, transfer learning, multi-task learning, domain adaptation, reinforcement learning, and end-to-end deep learning for speech recognition are covered using case studies.
Neural Machine Translation (NMT) is an end-to-end learning approach for automated translation, with the potential to overcome many of the weaknesses of conventional phrase-based translation systems. Unfortunately, NMT systems are known to be computationally expensive both in training and in translation inference. Also, most NMT systems have difficulty with rare words. These issues have hindered NMT's use in practical deployments and services, where both accuracy and speed are essential. In this work, we present GNMT, Google's Neural Machine Translation system, which attempts to address many of these issues. Our model consists of a deep LSTM network with 8 encoder and 8 decoder layers using attention and residual connections. To improve parallelism and therefore decrease training time, our attention mechanism connects the bottom layer of the decoder to the top layer of the encoder. To accelerate the final translation speed, we employ low-precision arithmetic during inference computations. To improve handling of rare words, we divide words into a limited set of common sub-word units ("wordpieces") for both input and output. This method provides a good balance between the flexibility of "character"-delimited models and the efficiency of "word"-delimited models, naturally handles translation of rare words, and ultimately improves the overall accuracy of the system. Our beam search technique employs a length-normalization procedure and uses a coverage penalty, which encourages generation of an output sentence that is most likely to cover all the words in the source sentence. On the WMT'14 English-to-French and English-to-German benchmarks, GNMT achieves competitive results to state-of-the-art. Using a human side-by-side evaluation on a set of isolated simple sentences, it reduces translation errors by an average of 60% compared to Google's phrase-based production system.
As a prerequisite to translation of poetry, we implement the ability to produce translations with meter and rhyme for phrase-based MT, examine whether the hypothesis space of such a system is flexible enough to accomodate such constraints, and investigate the impact of such constraints on translation quality.
Statistical machine translation and computational domain adaptation (Sustav za statističko strojno prevođenje i računalna adaptacija domene)
- I Dunder
I. Dunder, Statistical machine translation and computational
domain adaptation (Sustav za statističko strojno prevođenje i
računalna adaptacija domene) // doctoral dissertation. Zagreb:
University of Zagreb, 2015, p. 281.
Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation
- Y Wu
Y. Wu et al., "Google's Neural Machine Translation System:
Bridging the Gap between Human and Machine Translation,"
arXiv:1609.08144 [cs.CL], 2016, p. 23.
About machine translation
- Yandex Tech
- Com
Tech.yandex.com, "About machine translation," Available:
https://tech.yandex.com/translate/doc/dg/concepts/how-worksmachine-translation-docpage/ (20.11.2019)
An Awkward Disparity between BLEU / RIBES Scores and Human Judgements in Machine Translation
- L Tan
- J Dehdari
- J Van Genabith
L. Tan, J. Dehdari, and. J. van Genabith, "An Awkward Disparity
between BLEU / RIBES Scores and Human Judgements in
Machine Translation," Proc. of the 2nd Workshop on Asian
Translation (WAT2015), 2015, pp. 74-81.
CroSS: Croatian Speech Synthesizer - design and implementation
- I Dunđer
I. Dunđer, "CroSS: Croatian Speech Synthesizer -design and
implementation," Proc. 16th International Multiconference
INFORMATION SOCIETY -IS 2013 / Collaboration, Software
and Services in Information Society (CSS'2013), vol. A, 2013, pp.
257-260.
METEOR, M-BLEU and M-TER: Evaluation Metrics for High-Correlation with Human Rankings of Machine Translation Output
An Awkward Disparity between BLEU / RIBES Scores and Human Judgements in Machine Translation
- tan