Sequence-to-sequence learning led to significant improvements to machine translation (MT) and automatic speech recognition (ASR) systems. These advancements were first reflected in spoken language translation (SLT) when using a cascade of (at least) ASR and MT with the new "neural" models, then by using sequence-to-sequence learning to directly translate the input audio speech into text in the target language. In this thesis we cover both approaches to the SLT task. First, we show the limits of NMT in terms of robustness to input errors when compared to the previous phrase-based state of the art. We then focus on the NMT component to achieve better translation quality with higher computational efficiency by using a network based on weakly-recurrent units. Our last work involving a cascade explores the effects on the NMT robustness when adding automatic transcripts to the training data. In order to move to the direct speech-to-text approach, we introduce MuST-C, the largest multilingual SLT corpus for training direct translation systems. MuST-C increases significantly the size of publicly available data for this task as well as their language coverage. With such availability of data, we adapted the Transformer architecture to the SLT task for its computational efficiency . Our adaptation, which we call S-Transformer, is meant to better model the audio input, and with it we set a new state of the art for MuST-C. Building on these positive results, we finally use S-Transformer with different data applications: i) one-to-many multilingual translation by training it on MuST-C; ii participation to the IWSLT 19 shared task with data augmentation; and iii) instance-based adaptation for using the training data at test time. The results in this thesis show a steady quality improvement in direct SLT. Our hope is that the presented resources and technological solutions will increase its adoption in the near future, so to make multilingual information access easier in a globalized world.

This thesis aims to provide an in-depth overview of machine translation and machine interpreting, describing their history, development, and current state, as well as their place on the market and their potential use. The thesis describes machine translation and machine interpreting on theoretical, practical, and technological level, including the description of their basic principles, evaluation criteria, obstacles and challenges, and scope of their use. The thesis also includes a practical test of one of currently available machine interpreting programs - Skype Translator by Microsoft. The aim of the test is to determine whether the program can facilitate successful communication between two people who don't speak the same language, and to describe the user's experience. Keywords: machine translation, machine interpreting, machine translation limitations, neural translation, Skype Translator, STACL, speech recognition, speech production

Automatic translation has seen tremendous progress in recent years, mainly thanks to statistical methods applied to large parallel corpora. Transductions represent a principled approach to modeling translation, but existing transduction classes are either not expressive enough to capture structural regularities between natural languages or too complex to support efficient statistical induction on a large scale. A common approach is to severely prune search over a relatively unrestricted space of transduction grammars. These restrictions are often applied at different stages in a pipeline, with the obvious drawback of committing to irrevocable decisions that should not have been made. In this thesis we will instead restrict the space of transduction grammars to a space that is less expressive, but can be efficiently searched. First, the class of linear transductions is defined and characterized. They are generated by linear transduction grammars, which represent the natural bilingual case of linear grammars, as well as the natural linear case of inversion transduction grammars (and higher order syntax-directed transduction grammars). They are recognized by zipper finite-state transducers, which are equivalent to finite-state automata with four tapes. By allowing this extra dimensionality, linear transductions can represent alignments that finite-state transductions cannot, and by keeping the mechanism free of auxiliary storage, they become much more efficient than inversion transductions. Secondly, we present an algorithm for parsing with linear transduction grammars that allows pruning. The pruning scheme imposes no restrictions a priori, but guides the search to potentially interesting parts of the search space in an informed and dynamic way. Being able to parse efficiently allows learning of stochastic linear transduction grammars through expectation maximization. All the above work would be for naught if linear transductions were too poor a reflection of the actual transduction between natural languages. We test this empirically by building systems based on the alignments imposed by the learned grammars. The conclusion is that stochastic linear inversion transduction grammars learned from observed data stand up well to the state of the art.

Spelling normalization is the task to normalize non-standard words into standard words in texts, resulting in a decrease in out-of-vocabulary (OOV) words in texts for natural language processing (NLP) tasks such as information retrieval, machine translation, and opinion mining, improving the performance of various NLP applications on normalized texts. In this thesis, we explore different methods for spelling normalization of English student writings including traditional Levenshtein edit distance comparison, phonetic similarity comparison, character-based Statistical Machine Translation (SMT) and character-based Neural Machine Translation (NMT) methods. An important improvement of our implementation is that we develop an approach combining Levenshtein edit distance and phonetic similarity methods with added components of frequency count and compound splitting and it is evaluated as a best approach with 0.329% accuracy improvement and 63.63% error reduction on the original unnormalized test set.

Machine translation is a fundamental technology that is gaining more importance each day in our multilingual society. Companies and particulars are turning their attention to machine translation since it dramatically cuts down their expenses on translation and interpreting. However, the output of current machine translation systems is still far from the quality of translations generated by human experts. The overall goal of this thesis is to narrow down this quality gap by developing new methodologies and tools that improve the broader and more efficient deployment of machine translation technology. We start by proposing a new technique to improve the quality of the translations generated by fully-automatic machine translation systems. The key insight of our approach is that different translation systems, implementing different approaches and technologies, can exhibit different strengths and limitations. Therefore, a proper combination of the outputs of such different systems has the potential to produce translations of improved quality. We present minimum Bayes¿ risk system combination, an automatic approach that detects the best parts of the candidate translations and combines them to generate a consensus translation that is optimal with respect to a particular performance metric. We thoroughly describe the formalization of our approach as a weighted ensemble of probability distributions and provide efficient algorithms to obtain the optimal consensus translation according to the widespread BLEU score. Empirical results show that the proposed approach is indeed able to generate statistically better translations than the provided candidates. Compared to other state-of-the-art systems combination methods, our approach reports similar performance not requiring any additional data but the candidate translations. Then, we focus our attention on how to improve the utility of automatic translations for the end-user of the system. Since automatic translations are not perfect, a desirable feature of machine translation systems is the ability to predict at run-time the quality of the generated translations. Quality estimation is usually addressed as a regression problem where a quality score is predicted from a set of features that represents the translation. However, although the concept of translation quality is intuitively clear, there is no consensus on which are the features that actually account for it. As a consequence, quality estimation systems for machine translation have to utilize a large number of weak features to predict translation quality. This involves several learning problems related to feature collinearity and ambiguity, and due to the ¿curse¿ of dimensionality. We address these challenges by adopting a two-step training methodology. First, a dimensionality reduction method computes, from the original features, the reduced set of features that better explains translation quality. Then, a prediction model is built from this reduced set to finally predict the quality score. We study various reduction methods previously used in the literature and propose two new ones based on statistical multivariate analysis techniques. More specifically, the proposed dimensionality reduction methods are based on partial least squares regression. The results of a thorough experimentation show that the quality estimation systems estimated following the proposed two-step methodology obtain better prediction accuracy that systems estimated using all the original features. Moreover, one of the proposed dimensionality reduction methods obtained the best prediction accuracy with only a fraction of the original features. This feature reduction ratio is important because it implies a dramatic reduction of the operating times of the quality estimation system. An alternative use of current machine translation systems is to embed them within an interactive editing environment where the system and a human expert collaborate to generate error-free translations. This interactive machine translation approach have shown to reduce supervision effort of the user in comparison to the conventional decoupled post-edition approach. However, interactive machine translation considers the translation system as a passive agent in the interaction process. In other words, the system only suggests translations to the user, who then makes the necessary supervision decisions. As a result, the user is bound to exhaustively supervise every suggested translation. This passive approach ensures error-free translations but it also demands a large amount of supervision effort from the user. Finally, we study different techniques to improve the productivity of current interactive machine translation systems. Specifically, we focus on the development of alternative approaches where the system becomes an active agent in the interaction process. We propose two different active approaches. On the one hand, we describe an active interaction approach where the system informs the user about the reliability of the suggested translations. The hope is that this information may help the user to locate translation errors thus improving the overall translation productivity. We propose different scores to measure translation reliability at the word and sentence levels and study the influence of such information in the productivity of an interactive machine translation system. Empirical results show that the proposed active interaction protocol is able to achieve a large reduction in supervision effort while still generating translations of very high quality. On the other hand, we study an active learning framework for interactive machine translation. In this case, the system is not only able to inform the user of which suggested translations should be supervised, but it is also able to learn from the user-supervised translations to improve its future suggestions. We develop a value-of-information criterion to select which automatic translations undergo user supervision. However, given its high computational complexity, in practice we study different selection strategies that approximate this optimal criterion. Results of a large scale experimentation show that the proposed active learning framework is able to obtain better compromises between the quality of the generated translations and the human effort required to obtain them. Moreover, in comparison to a conventional interactive machine translation system, our proposal obtained translations of twice the quality with the same supervision effort. / González Rubio, J. (2014). On the effective deployment of current machine translation technology [Tesis doctoral no publicada]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/37888 / TESIS

Ce mémoire a pour but de déterminer la stratégie de traduction automatique des alertes météorologiques produites par Environnement Canada, qui nécessite le moins d’efforts de postédition de la part des correcteurs du bureau de la traduction. Nous commencerons par constituer un corpus bilingue d’alertes météorologiques représentatives de la tâche de traduction. Ensuite, ces données nous serviront à comparer les performances de différentes approches de traduction automatique, de configurations de mémoires de traduction et de systèmes hybrides. Nous comparerons les résultats de ces différents modèles avec le système WATT, développé par le RALI pour Environnement Canada, ainsi qu’avec les systèmes de l’industrie GoogleTranslate et DeepL. Nous étudierons enfin une approche de postédition automatique. / The purpose of this paper is to determine the strategy for the automatic translation of weather warnings produced by Environment Canada, which requires the least post-editing effort by the proofreaders of the Translation Bureau. We will begin by developing a bilingual corpus of weather warnings representative of this task. Then, this data will be used to compare the performance of different approaches of machine translation, translation memory configurations and hybrid systems. We will compare the results of these models with the system WATT, the latest system provided by RALI for Environment Canada, as well as with the industry systems GoogleTranslate and DeepL. Finaly, we will study an automatic post-edition system.

[ES] La Traducción Automática Estadística es un sup-campo de la lingüística computacional que investiga como emplear los ordenadores en el proceso de traducción de un texto de un lenguaje humano a otro. La traducción automática estadística es el enfoque más popular que se emplea para construir estos sistemas de traducción automáticos. La calidad de dichos sistemas depende en gran medida de los ejemplos de traducción que se emplean durante los procesos de entrenamiento y adaptación de los modelos. Los conjuntos de datos empleados son obtenidos a partir de una gran variedad de fuentes y en muchos casos puede que no tengamos a mano los datos más adecuados para un dominio específico. Dado este problema de carencia de datos, la idea principal para solucionarlo es encontrar aquellos conjuntos de datos más adecuados para entrenar o adaptar un sistema de traducción. En este sentido, esta tesis propone un conjunto de técnicas de selección de datos que identifican los datos bilingües más relevantes para una tarea extraídos de un gran conjunto de datos. Como primer paso en esta tesis, las técnicas de selección de datos son aplicadas para mejorar la calidad de la traducción de los sistemas de traducción bajo el paradigma basado en frases. Estas técnicas se basan en el concepto de representación continua de las palabras o las oraciones en un espacio vectorial. Los resultados experimentales demuestran que las técnicas utilizadas son efectivas para diferentes lenguajes y dominios. El paradigma de Traducción Automática Neuronal también fue aplicado en esta tesis. Dentro de este paradigma, investigamos la aplicación que pueden tener las técnicas de selección de datos anteriormente validadas en el paradigma basado en frases. El trabajo realizado se centró en la utilización de dos tareas diferentes de adaptación del sistema. Por un lado, investigamos cómo aumentar la calidad de traducción del sistema, aumentando el tamaño del conjunto de entrenamiento. Por otro lado, el método de selección de datos se empleó para crear un conjunto de datos sintéticos. Los experimentos se realizaron para diferentes dominios y los resultados de traducción obtenidos son convincentes para ambas tareas. Finalmente, cabe señalar que las técnicas desarrolladas y presentadas a lo largo de esta tesis pueden implementarse fácilmente dentro de un escenario de traducción real. / [CAT] La Traducció Automàtica Estadística és un sup-camp de la lingüística computacional que investiga com emprar els ordinadors en el procés de traducció d'un text d'un llenguatge humà a un altre. La traducció automàtica estadística és l'enfocament més popular que s'empra per a construir aquests sistemes de traducció automàtics. La qualitat d'aquests sistemes depèn en gran mesura dels exemples de traducció que s'empren durant els processos d'entrenament i adaptació dels models. Els conjunts de dades emprades són obtinguts a partir d'una gran varietat de fonts i en molts casos pot ser que no tinguem a mà les dades més adequades per a un domini específic. Donat aquest problema de manca de dades, la idea principal per a solucionar-ho és trobar aquells conjunts de dades més adequades per a entrenar o adaptar un sistema de traducció. En aquest sentit, aquesta tesi proposa un conjunt de tècniques de selecció de dades que identifiquen les dades bilingües més rellevants per a una tasca extrets d'un gran conjunt de dades. Com a primer pas en aquesta tesi, les tècniques de selecció de dades són aplicades per a millorar la qualitat de la traducció dels sistemes de traducció sota el paradigma basat en frases. Aquestes tècniques es basen en el concepte de representació contínua de les paraules o les oracions en un espai vectorial. Els resultats experimentals demostren que les tècniques utilitzades són efectives per a diferents llenguatges i dominis. El paradigma de Traducció Automàtica Neuronal també va ser aplicat en aquesta tesi. Dins d'aquest paradigma, investiguem l'aplicació que poden tenir les tècniques de selecció de dades anteriorment validades en el paradigma basat en frases. El treball realitzat es va centrar en la utilització de dues tasques diferents. D'una banda, investiguem com augmentar la qualitat de traducció del sistema, augmentant la grandària del conjunt d'entrenament. D'altra banda, el mètode de selecció de dades es va emprar per a crear un conjunt de dades sintètiques. Els experiments es van realitzar per a diferents dominis i els resultats de traducció obtinguts són convincents per a ambdues tasques. Finalment, cal assenyalar que les tècniques desenvolupades i presentades al llarg d'aquesta tesi poden implementar-se fàcilment dins d'un escenari de traducció real. / [EN] La Traducció Automàtica Estadística és un sup-camp de la lingüística computacional que investiga com emprar els ordinadors en el procés de traducció d'un text d'un llenguatge humà a un altre. La traducció automàtica estadística és l'enfocament més popular que s'empra per a construir aquests sistemes de traducció automàtics. La qualitat d'aquests sistemes depèn en gran mesura dels exemples de traducció que s'empren durant els processos d'entrenament i adaptació dels models. Els conjunts de dades emprades són obtinguts a partir d'una gran varietat de fonts i en molts casos pot ser que no tinguem a mà les dades més adequades per a un domini específic. Donat aquest problema de manca de dades, la idea principal per a solucionar-ho és trobar aquells conjunts de dades més adequades per a entrenar o adaptar un sistema de traducció. En aquest sentit, aquesta tesi proposa un conjunt de tècniques de selecció de dades que identifiquen les dades bilingües més rellevants per a una tasca extrets d'un gran conjunt de dades. Com a primer pas en aquesta tesi, les tècniques de selecció de dades són aplicades per a millorar la qualitat de la traducció dels sistemes de traducció sota el paradigma basat en frases. Aquestes tècniques es basen en el concepte de representació contínua de les paraules o les oracions en un espai vectorial. Els resultats experimentals demostren que les tècniques utilitzades són efectives per a diferents llenguatges i dominis. El paradigma de Traducció Automàtica Neuronal també va ser aplicat en aquesta tesi. Dins d'aquest paradigma, investiguem l'aplicació que poden tenir les tècniques de selecció de dades anteriorment validades en el paradigma basat en frases. El treball realitzat es va centrar en la utilització de dues tasques diferents d'adaptació del sistema. D'una banda, investiguem com augmentar la qualitat de traducció del sistema, augmentant la grandària del conjunt d'entrenament. D'altra banda, el mètode de selecció de dades es va emprar per a crear un conjunt de dades sintètiques. Els experiments es van realitzar per a diferents dominis i els resultats de traducció obtinguts són convincents per a ambdues tasques. Finalment, cal assenyalar que les tècniques desenvolupades i presentades al llarg d'aquesta tesi poden implementar-se fàcilment dins d'un escenari de traducció real. / Chinea Ríos, M. (2019). Advanced techniques for domain adaptation in Statistical Machine Translation [Tesis doctoral no publicada]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/117611 / TESIS

The goal of this thesis is to describe and build a system for neural machine translation. System is built with recurrent neural networks - encoder-decoder architecture in particular. The result is a nmt library used to conduct experiments with different model parameters. Results of the experiments are compared with system built with the statistical tool Moses.

In this project, we have investigated how to perform rule-based machine translation of sets of keywords between two languages. The goal was to translate an input set, which contains one or more keywords in a source language, to a corresponding set of keywords, with the same number of elements, in the target language. However, some words in the source language may have several senses and may be translated to several, or no, words in the target language. If ambiguous translations occur, the best translation of the keyword should be chosen with respect to the context. In traditional machine translation, a word's context is determined by a phrase or sentences where the word occurs. In this project, the set of keywords represents the context. By investigating traditional approaches to machine translation (MT), we designed and described models for the specific purpose of keyword- translation. We have proposed a solution, based on direct translation for translating keywords between English and Swedish. In the proposed solu- tion, we also introduced a simple graph-based model for solving ambigu- ous translations. / I detta projekt har vi undersökt hur man utför regelbaserad maskinöver- sättning av nyckelord mellan två språk. Målet var att översätta en given mängd med ett eller flera nyckelord på ett källspråk till en motsvarande, lika stor mängd nyckelord på målspråket. Vissa ord i källspråket kan dock ha flera betydelser och kan översättas till flera, eller inga, ord på målsprå- ket. Om tvetydiga översättningar uppstår ska nyckelordets bästa över- sättning väljas med hänsyn till sammanhanget. I traditionell maskinö- versättning bestäms ett ords sammanhang av frasen eller meningen som det befinner sig i. I det här projektet representerar den givna mängden nyckelord sammanhanget. Genom att undersöka traditionella tillvägagångssätt för maskinöversätt- ning har vi designat och beskrivit modeller specifikt för översättning av nyckelord. Vi har presenterat en direkt maskinöversättningslösning av nyckelord mellan engelska och svenska där vi introducerat en enkel graf- baserad modell för tvetydiga översättningar.

This thesis presents scientific contributions to the field of multimodal interac- tive structured prediction (MISP). The aim of MISP is to reduce the human effort required to supervise an automatic output, in an efficient and ergonomic way. Hence, this thesis focuses on the two aspects of MISP systems. The first aspect, which refers to the interactive part of MISP, is the study of strate- gies for efficient human¿computer collaboration to produce error-free outputs. Multimodality, the second aspect, deals with other more ergonomic modalities of communication with the computer rather than keyboard and mouse. To begin with, in sequential interaction the user is assumed to supervise the output from left-to-right so that errors are corrected in sequential order. We study the problem under the decision theory framework and define an optimum decoding algorithm. The optimum algorithm is compared to the usually ap- plied, standard approach. Experimental results on several tasks suggests that the optimum algorithm is slightly better than the standard algorithm. In contrast to sequential interaction, in active interaction it is the system that decides what should be given to the user for supervision. On the one hand, user supervision can be reduced if the user is required to supervise only the outputs that the system expects to be erroneous. In this respect, we define a strategy that retrieves first the outputs with highest expected error first. Moreover, we prove that this strategy is optimum under certain conditions, which is validated by experimental results. On the other hand, if the goal is to reduce the number of corrections, active interaction works by selecting elements, one by one, e.g., words of a given output to be supervised by the user. For this case, several strategies are compared. Unlike the previous case, the strategy that performs better is to choose the element with highest confidence, which coincides with the findings of the optimum algorithm for sequential interaction. However, this also suggests that minimizing effort and supervision are contradictory goals. With respect to the multimodality aspect, this thesis delves into techniques to make multimodal systems more robust. To achieve that, multimodal systems are improved by providing contextual information of the application at hand. First, we study how to integrate e-pen interaction in a machine translation task. We contribute to the state-of-the-art by leveraging the information from the source sentence. Several strategies are compared basically grouped into two approaches: inspired by word-based translation models and n-grams generated from a phrase-based system. The experiments show that the former outper- forms the latter for this task. Furthermore, the results present remarkable improvements against not using contextual information. Second, similar ex- periments are conducted on a speech-enabled interface for interactive machine translation. The improvements over the baseline are also noticeable. How- ever, in this case, phrase-based models perform much better than word-based models. We attribute that to the fact that acoustic models are poorer estima- tions than morphologic models and, thus, they benefit more from the language model. Finally, similar techniques are proposed for dictation of handwritten documents. The results show that speech and handwritten recognition can be combined in an effective way. Finally, an evaluation with real users is carried out to compare an interactive machine translation prototype with a post-editing prototype. The results of the study reveal that users are very sensitive to the usability aspects of the user interface. Therefore, usability is a crucial aspect to consider in an human evaluation that can hinder the real benefits of the technology being evaluated. Hopefully, once usability problems are fixed, the evaluation indicates that users are more favorable to work with the interactive machine translation system than to the post-editing system. / Alabau Gonzalvo, V. (2014). Multimodal interactive structured prediction [Tesis doctoral]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/35135 / TESIS / Premios Extraordinarios de tesis doctorales