En un mundo globalizado como el actual en el que, además, muchas sociedades son inherentemente multilingües, la traducción e interpretación entre diversas lenguas requiere de un esfuerzo notable debido a su volumen. Diversas tecnologías de asistencia existen para facilitar la realización de estas tareas de traducción, entre las que se encuentra la traducción automática interactiva (TAI), una modalidad en la que el traductor va escribiendo la traducción y el sistema ofrece sugerencias que predicen las próximas palabras que va a teclear. En el estado de la cuestión, las aproximaciones a la TAI siguen una aproximación de caja de cristal: están firmemente acopladas a un sistema de traducción automática (muchas veces estadístico) que utilizan para generar las sugerencias, por lo que tienen las mismas limitaciones que el sistema de traducción automática subyacente. Esta tesis desarrolla una nueva aproximación de caja negra, donde cualquier recurso bilingüe (no solo sistemas de traducción automática, sino también otros recursos como memorias de traducción, diccionarios, glosarios, etc.) puede ser utilizado para generar las sugerencias.

京都大学 / 0048 / 新制・課程博士 / 博士(情報学) / 甲第17921号 / 情博第503号 / 新制||情||89(附属図書館) / 30741 / 京都大学大学院情報学研究科社会情報学専攻 / (主査)教授 石田 亨, 教授 田中 克己, 教授 黒橋 禎夫 / 学位規則第4条第1項該当 / Doctor of Informatics / Kyoto University / DFAM

Given a sentence, a paraphrase generation system produces a sentence that says the same thing but usually in a different way. The paraphrase generation problem can be formulated in the machine translation paradigm; instead of translation of English to a foreign language, the system translates an English sentence (for example) to another English sentence. Quirk et al. (2004) demonstrated this approach to generate almost 90% acceptable paraphrases. However, most of the sentences had little variation from the original input sentence. Leveraging syntactic information, this thesis project presents an approach that successfully generated more varied paraphrase sentences than the approach of Quirk et al. while maintaining coverage of the proportion of acceptable paraphrases generated. The ParaMeTer system (Paraphrasing by MT) identifies syntactic chunks in paraphrase sentences and substitutes labels for those chunks. This enables the system to generalize movements that are more syntactically plausible, as syntactic chunks generally capture sets of words that can change order in the sentence without losing grammaticality. ParaMeTer then uses statistical phrase-based MT techniques to learn alignments for the words and chunk labels alike. The baseline system followed the same pattern as the Quirk et al. system - a statistical phrase-based MT system. Human judgments showed that the syntactic approach and baseline both achieve approximately the same ratio of fluent, acceptable paraphrase sentences per fluent sentences. These judgments also showed that the ParaMeTer system has more phrase rearrangement than the baseline system. Though the baseline has more within-phrase alteration, future modifications such as a chunk-only translation model should improve ParaMeTer's variation for phrase alteration as well.

This study focuses on and tries to identify areas where machine translation can be useful by examining translated fictional and non-fictional texts, and the extent to which these different text types are better or worse suited for machine translation.  It additionally evaluates the performance of the free online translation tool Google Translate (GT). The BLEU automatic evaluation metric for machine translation was used for this study, giving a score of 27.75 BLEU value for fictional texts and 32.16 for the non-fictional texts. The non-fictional texts are samples of law documents, (commercial) company reports, social science texts (religion, welfare, astronomy) and medicine. These texts were selected because of their degree of difficulty. The non-fictional sentences are longer than those of the fictional texts and in this regard MT systems have struggled. In spite of having longer sentences, the non-fictional texts got a higher BLUE score than the fictional ones. It is speculated that one reason for the higher score of non-fictional texts might be that more specific terminology is used in these texts, leaving less room for subjective interpretation than for the fictional texts. There are other levels of meaning at work in the fictional texts that the human translator needs to capture.

Natural languages display a great variety of different word orders, and one of the major challenges facing statistical machine translation is in modelling these differences. This thesis is motivated by a survey of 110 different language pairs drawn from the Europarl project, which shows that word order differences account for more variation in translation performance than any other factor. This wide ranging analysis provides compelling evidence for the importance of research into reordering. There has already been a great deal of research into improving the quality of the word order in machine translation output. However, there has been very little analysis of how best to evaluate this research. Current machine translation metrics are largely focused on evaluating the words used in translations, and their ability to measure the quality of word order has not been demonstrated. In this thesis we introduce novel metrics for quantitatively evaluating reordering. Our approach isolates the word order in translations by using word alignments. We reduce alignment information to permutations and apply standard distance metrics to compare the word order in the reference to that of the translation. We show that our metrics correlate more strongly with human judgements of word order quality than current machine translation metrics. We also show that a combined lexical and reordering metric, the LRscore, is useful for training translation model parameters. Humans prefer the output of models trained using the LRscore as the objective function, over those trained with the de facto standard translation metric, the BLEU score. The LRscore thus provides researchers with a reliable metric for evaluating the impact of their research on the quality of word order.

Combining Outputs from Machine Translation Systems By Fahim A. Salim Supervised by: Ing. Zdenek Zabokrtsky, Ph.D Institute of Formal and Applied Linguistics, Charles University in Prague 2010. Abstract: Due to the massive ongoing research there are many paradigms of Machine Translation systems with diverse characteristics. Even systems designed on the same paradigm might perform differently in different scenarios depending upon their training data used and other design decisions made. All Machine Translation Systems have their strengths and weaknesses and often weakness of one MT system is the strength of the other. No single approach or system seems to always perform best, therefore combining different approaches or systems i.e. creating systems of Hybrid nature, to capitalize on their strengths and minimizing their weaknesses in an ongoing trend in Machine Translation research. But even Systems of Hybrid nature has limitations and they also tend to perform differently in different scenarios. Thanks to the World Wide Web and open source, nowadays one can have access to many different and diverse Machine Translation systems therefore it is practical to have techniques which could combine the translation of different MT systems and produce a translation which is better than any of the individual systems....

In recent years, corpus based machine translation systems produce significant results for a number of language pairs. However, for low-resource languages like Urdu the purely statistical or purely example based methods are not performing well. On the other hand, the rule-based approaches require a huge amount of time and resources for the development of rules, which makes it difficult in most scenarios. Hybrid machine translation systems might be one of the solutions to overcome these problems, where we can combine the best of different approaches to achieve quality translation. The goal of the thesis is to explore different combinations of approaches and to evaluate their performance over the standard corpus based methods currently in use. This includes: 1. Use of syntax-based and dependency-based reordering rules with Statistical Machine Translation. 2. Automatic extraction of lexical and syntactic rules using statistical methods to facilitate the Transfer-Based Machine Translation. The novel element in the proposed work is to develop an algorithm to learn automatic reordering rules for English-to-Urdu statistical machine translation. Moreover, this approach can be extended to learn lexical and syntactic rules to build a rule-based machine translation system.

In recent years there has been an increased interest in domain adaptation techniques for statistical machine translation (SMT) to deal with the growing amount of data from different sources. Topic modelling techniques applied to SMT are closely related to the field of domain adaptation but more flexible in dealing with unstructured text. Topic models can capture latent structure in texts and are therefore particularly suitable for modelling structure in between and beyond corpus boundaries, which are often arbitrary. In this thesis, the main focus is on dynamic translation model adaptation to texts of unknown origin, which is a typical scenario for an online MT engine translating web documents. We introduce a new bilingual topic model for SMT that takes the entire document context into account and for the first time directly estimates topic-dependent phrase translation probabilities in a Bayesian fashion. We demonstrate our model’s ability to improve over several domain adaptation baselines and further provide evidence for the advantages of bilingual topic modelling for SMT over the more common monolingual topic modelling. We also show improved performance when deriving further adapted translation features from the same model which measure different aspects of topical relatedness. We introduce another new topic model for SMT which exploits the distributional nature of phrase pair meaning by modelling topic distributions over phrase pairs using their distributional profiles. Using this model, we explore combinations of local and global contextual information and demonstrate the usefulness of different levels of contextual information, which had not been previously examined for SMT. We also show that combining this model with a topic model trained at the document-level further improves performance. Our dynamic topic adaptation approach performs competitively in comparison with two supervised domain-adapted systems. Finally, we shed light on the relationship between domain adaptation and topic adaptation and propose to combine multi-domain adaptation and topic adaptation in a framework that entails automatic prediction of domain labels at the document level. We show that while each technique provides complementary benefits to the overall performance, there is an amount of overlap between domain and topic adaptation. This can be exploited to build systems that require less adaptation effort at runtime.

In the past few years, evaluating on adversarial examples has become a standard procedure to measure robustness of deep learning models. Literature on adversarial examples for neural nets has largely focused on image data, which are represented as points in continuous space. However, a vast proportion of machine learning models operate on discrete input, and thus demand a similar rigor in understanding their vulnerabilities and robustness. We study robustness of neural network architectures for textual and graph inputs, through the lens of adversarial input perturbations. We will cover methods for both attacks and defense; we will focus on 1) addressing challenges in optimization for creating adversarial perturbations for discrete data; 2) evaluating and contrasting white-box and black-box adversarial examples; and 3) proposing efficient methods to make the models robust against adversarial attacks.

This thesis introduces the concept of an encoder-decoder neural network and develops architectures for the construction of such networks. Encoder-decoder neural networks are probabilistic conditional generative models of high-dimensional structured items such as natural language utterances and natural images. Encoder-decoder neural networks estimate a probability distribution over structured items belonging to a target set conditioned on structured items belonging to a source set. The distribution over structured items is factorized into a product of tractable conditional distributions over individual elements that compose the items. The networks estimate these conditional factors explicitly. We develop encoder-decoder neural networks for core tasks in natural language processing and natural image and video modelling. In Part I, we tackle the problem of sentence modelling and develop deep convolutional encoders to classify sentences; we extend these encoders to models of discourse. In Part II, we go beyond encoders to study the longstanding problem of translating from one human language to another. We lay the foundations of neural machine translation, a novel approach that views the entire translation process as a single encoder-decoder neural network. We propose a beam search procedure to search over the outputs of the decoder to produce a likely translation in the target language. Besides known recurrent decoders, we also propose a decoder architecture based solely on convolutional layers. Since the publication of these new foundations for machine translation in 2013, encoder-decoder translation models have been richly developed and have displaced traditional translation systems both in academic research and in large-scale industrial deployment. In services such as Google Translate these models process in the order of a billion translation queries a day. In Part III, we shift from the linguistic domain to the visual one to study distributions over natural images and videos. We describe two- and three- dimensional recurrent and convolutional decoder architectures and address the longstanding problem of learning a tractable distribution over high-dimensional natural images and videos, where the likely samples from the distribution are visually coherent. The empirical validation of encoder-decoder neural networks as state-of- the-art models of tasks ranging from machine translation to video prediction has a two-fold significance. On the one hand, it validates the notions of assigning probabilities to sentences or images and of learning a distribution over a natural language or a domain of natural images; it shows that a probabilistic principle of compositionality, whereby a high- dimensional item is composed from individual elements at the encoder side and whereby a corresponding item is decomposed into conditional factors over individual elements at the decoder side, is a general method for modelling cognition involving high-dimensional items; and it suggests that the relations between the elements are best learnt in an end-to-end fashion as non-linear functions in distributed space. On the other hand, the empirical success of the networks on the tasks characterizes the underlying cognitive processes themselves: a cognitive process as complex as translating from one language to another that takes a human a few seconds to perform correctly can be accurately modelled via a learnt non-linear deterministic function of distributed vectors in high-dimensional space.