In this thesis, we focus on detecting fine-grained semantic divergences—subtle meaning differences in sentences that overlap in content—to improve machine and human translation understanding.

In an evolving landscape of crisis communication, the need for robust and adaptable Machine Translation (MT) systems is more pressing than ever, particularly for low-resource languages. This study presents a comprehensive exploration of leveraging Large Language Models (LLMs) and Multilingual LLMs (MLLMs) to enhance MT capabilities in such scenarios. By focusing on the unique challenges posed by crisis situations where speed, accuracy, and the ability to handle a wide range of languages are paramount, this research outlines a novel approach that combines the cutting-edge capabilities of LLMs with fine-tuning techniques and community-driven corpus development strategies. At the core of this study is the development and empirical evaluation of MT systems tailored for two low-resource language pairs, illustrating the process from initial model selection and fine-tuning through to deployment. Bespoke systems are developed and modelled on the recent Covid-19 pandemic. The research highlights the importance of community involvement in creating highly specialised, crisis-specific datasets and compares custom GPTs with NLLB-adapted MLLM models. It identifies fine-tuned MLLM models as offering superior performance compared with their LLM counterparts. A scalable and replicable model for rapid MT system development in crisis scenarios is outlined. Our approach enhances the field of humanitarian technology by offering a blueprint for developing multilingual communication systems during emergencies.

While most current work in multimodal machine translation (MMT) uses the Multi30k dataset for training and evaluation, we find that the resulting models overfit to the Multi30k dataset to an extreme degree. Consequently, these models perform very badly when evaluated against typical text-only testing sets such as the newstest datasets. In order to perform well on both Multi30k and typical text-only datasets, we use a performant text-only machine translation (MT) model as the starting point of our MMT model. We add vision-text adapter layers connected via gating mechanisms to the MT model, and incrementally transform the MT model into an MMT model by 1) pre-training using vision-based masking of the source text and 2) fine-tuning on Multi30k. We achieve a state-of-the-art performance on the Multi30k 2016 en-de test set of 46.5 BLEU4 score and 0.61 CoMMuTE score via this approach while retaining the performance of the original text-only MT model against the newstest dataset.

The challenge of visual grounding and masking in multimodal machine translation (MMT) systems has encouraged varying approaches to the detection and selection of visually-grounded text tokens for masking. We introduce new methods for detection of visually and contextually relevant (concrete) tokens from source sentences, including detection with natural language processing (NLP), detection with object detection, and a joint detection-verification technique. We also introduce new methods for selection of detected tokens, including shortest n tokens, longest n tokens, and all detected concrete tokens. We utilize the GRAM MMT architecture to train models against synthetically collated multimodal datasets of source images with masked sentences, showing performance improvements and improved usage of visual context during translation tasks over the baseline model.

Translation memories (TMs) are the backbone for professional translation tools called computer-aided translation (CAT) tools. In order to perform a translation using a CAT tool, a translator uses the TM to gather translations similar to the desired segment to translate (s’). Many CAT tools offer a fuzzy-match algorithm to locate segments (s) in the TM that are close in distance to s’. After locating two similar segments, the CAT tool will present parallel segments (s, t) that contain one segment in the source language along with its translation in the target language. Additionally, CAT tools contain fuzzy-match repair (FMR) techniques that will automatically use the parallel segments from the TM to create new TM entries containing a modified version of the original with the idea in mind that it will be the translation of s’. Most FMR techniques use machine translation as a way of ‘repairing’ those words that have to be modified. In this article, we show that for a large part of those words which are anchored, we can use other techniques that are based on machine learning approaches such as Word2Vec. BERT, and even ChatGPT. Specifically, we show that for anchored words that follow the continuous bag-of-words (CBOW) paradigm, Word2Vec, BERT, and GPT-4 can be used to achieve similar and, for some cases, better results than neural machine translation for translating anchored words from French to English.

The typical workflow for a professional translator to translate a document from its source language (SL) to a target language (TL) is not always focused on what many language models in natural language processing (NLP) do - predict the next word in a series of words. While high-resource languages like English and French are reported to achieve near human parity using common metrics for measurement such as BLEU and COMET, we find that an important step is being missed: the translation of technical terms, specifically acronyms. Some state-of-the art machine translation systems like Google Translate which are publicly available can be erroneous when dealing with acronyms - as much as 50% in our findings. This article addresses acronym disambiguation for MT systems by proposing an additional step to the SL-TL (FR-EN) translation workflow where we first offer a new acronym corpus for public consumption and then experiment with a search-based thresholding algorithm that achieves nearly 10% increase when compared to Google Translate and OpusMT.

Humans outperform large language models (LLMs) on sophisticated tasks because human cognition involves a range of cognitive functions and their dynamic interactions. This study explores how integrating human cognition through concept-guided instruction and few-shot teaching in the prompt can guide LLMs to improve translation outcomes. We first demonstrate that for simple and widely used concepts, concept-guided prompting approaches offer significant benefits. We then test prompt engineering with Chinese-to-English translation examples, using hypothetical spaces—generated by GPT4—to estimate the complexity of various concepts and Likert scores—generated by human experts—to evaluate the translation performance. Our findings show that LLM translation performance declines as concept complexity increases. We also identify additional challenges: LLMs struggle with continuity in explaining and practicing sophisticated concepts due to the lack of human-like cognitive functions, such as cognitive dissonance. Additionally, LLMs lack a graceful speed-accuracy tradeoff because they do not possess the dynamic information processing, response strategies, and performance assessment that humans do. However, LLMs can mitigate some of these challenges by using Chain-of-Thought (CoT) reasoning, which is especially effective for problems requiring consistent, well-structured reasoning steps. Despite this, LLMs can only represent the effects of complex human cognitive functions through (often) fragmented linguistic descriptions, whereas humans excel at understanding critical and broader contexts and the interconnections between cognitive aspects.

Recent works have shown that prompting large language models (LLMs) is effective for translation with markup where LLMs can simultaneously transfer markup tags while ensuring that the content, both inside and outside tag pairs is correctly translated. However, these works make a rather unrealistic assumption of the existence of high-quality parallel sentences with markup for prompting. Furthermore, the impact of instruction fine-tuning (IFT) in this setting is unknown. In this paper, we provide a study, the first of its kind, focusing on the effectiveness of synthetically created markup data and IFT for translation with markup using LLMs. We focus on translation from English to five European languages, German, French, Dutch, Finnish and Russian, where we show that regardless of few-shot prompting or IFT, synthetic data created via word alignments, while leading to inferior markup transfer compared to using original data with markups, does not negatively impact the translation quality. Furthermore, IFT mainly impacts the translation quality compared to few-shot prompting and has slightly better markup transfer capabilities than the latter. We hope our work will help practitioners make effective decisions on modeling choices for LLM based translation with markup.

The quality of output from large language models (LLMs), particularly in machine translation (MT), is closely tied to the quality of in-context examples (ICEs) provided along with the query, i.e., the text to translate. The effectiveness of these ICEs is influenced by various factors, such as the domain of the source text, the order in which the ICEs are presented, the number of these examples, and the prompt templates used. Naturally, selecting the most impactful ICEs depends on understanding how these affect the resulting translation quality, which ultimately relies on translation references or human judgment. This paper presents a novel methodology for in-context learning (ICL) that relies on a search algorithm guided by domain-specific quality estimation (QE). Leveraging the XGLM model, our methodology estimates the resulting translation quality without the need for translation references, selecting effective ICEs for MT to maximize translation quality. Our results demonstrate significant improvements over existing ICL methods and higher translation performance compared to fine-tuning a pre-trained language model (PLM), specifically mBART-50.

In long-term translation projects, like Parliamentary text, there is a desire to build machine translation systems that can adapt to changes over time. We implement and examine a simple approach to continual learning for neural machine translation, exploring tradeoffs between consistency, the model’s ability to learn from incoming data, and the time a client would need to wait to obtain a newly trained translation system.

In September 2023, the Government of Canada issued a ‘Guide on the Use of Generative AI’ with recommendations for Canadian government institutions and their employees. As other similar documents published by various organizations in recent years, this document makes recommendations regarding transparency, stating that whenever generative AI is used to produce content, the reader should be informed that “messages addressed to them are generated by AI”. While this guide does not address specifically the case of machine translation, it does mention translation as a potential application of generative AI. Therefore, one question that naturally arises is: Should machine-translated texts be explicitly labelled as AI-generated content wherever they are used? In this position paper, we examine this question in detail, with the goal of proposing clear guidelines specifically regarding MT, not only for government institutions, but for anyone using MT technology to produce new versions of a text. Our main conclusion is that machine-translated text is indeed AI-generated content. As such, it should be explicitly marked everywhere it is used. We make recommendations as to what form this labelling might take. We also examine under what conditions labelling can be removed or omitted.

Hyperparameter optimization (HPO) enhances neural machine translation (NMT) models but demands substantial computational resources. Successive halving, a multi-fidelity HPO method, mitigates this by early stopping unpromising models and allocating more resources to promising ones. This method is particularly relevant for NMT and large language models, which are computationally intensive. However, successive halving relies on a noisy estimation of model performance and assumes that early performance is highly correlated with final performance. We introduce a table lookup benchmark dataset to study the reliability of successive halving and propose best practices for its application in NMT and large language models.

Transformer-based models in Neural Machine Translation (NMT) rely heavily on multi-head attention for capturing dependencies within and across source and target sequences. In Transformers, attention mechanisms dynamically determine which parts of the sentence to focus on in the encoder and decoder through self-attention and cross-attention. Our experiments show that high-resource NMT systems often exhibit a specific peaked attention distribution, indicating a focus on key elements. However, in low-resource NMT, attention tends to be dispersed throughout the sentence, lacking the focus demonstrated by high-resource models. To tackle this issue, we present EaDRA (Entropy– and Distance-Regularized Attention), which introduces an inductive bias to prioritize essential elements and guide the attention mechanism accordingly. Extensive experiments using EaDRA on diverse low-resource language pairs demonstrate significant improvements in translation quality, while incurring negligible computational cost.

Despite advancements in neural machine translation, word sense disambiguation remains challenging, particularly with limited textual context. Multimodal Machine Translation enhances text-only models by integrating visual information, but its impact varies across translations. This study focuses on ambiguous sentences to investigate the effectiveness of utilizing visual information. By prioritizing these sentences, which benefit from visual cues, we aim to enhance hybrid multimodal and text-only translation approaches. We utilize Latent Semantic Analysis and Sentence-BERT to extract context vectors from the British National Corpus, enabling the assessment of semantic diversity. Our approach enhances translation quality for English-German and English-French on Multi30k, assessed through metrics including BLEU, chrF2, and TER.

Conversational speech translation is an important technology that fosters communication among people of different language backgrounds. Three-way parallel data in the form of source speech, source transcript, and target translation is usually required to train end-to-end systems. However, such datasets are not readily available and are expensive to create as this involves multiple annotation stages. In this paper, we investigate the use of synthetic data from generative models, namely machine translation and text-to-speech synthesis, for training conversational speech translation systems. We show that adding synthetic data to the training recipe increasingly improves end-to-end training performance, especially when limited real data is available. However, when no real data is available, no amount of synthetic data helps.

We explore the potential of LLMs to enhance the translation process of rhymed and non-rhymed poetry. We examine LLMs’ performance in terms of lexical variety, lexical density, and sentence length compared to human translations (HT). We also examine the models’ abilities to translate sonnets while preserving the rhyme scheme of the source text. Our findings suggest that LLMs can serve as valuable tools for literary translators, assisting with the creative process and suggesting solutions to problems that may not otherwise have been considered. However, if the paradigm is flipped, such that instead of the systems being as tools by human translators, humans are used to post-edit the outputs to a standard comparable to the published translations, the amount of work required to complete the post-editing stage may outweigh any benefits assocaiated with using machine translation in the first place.

In this position paper, we examine ways in which researchers in machine translation and translation studies have approached the problem of evaluating the output of machine translation systems and, more broadly, the questions of what it means to define translation quality. We explore their similarities and differences, highlighting the role that the purpose and context of translation plays in translation studies approaches. We argue that evaluation of machine translation (e.g., in shared tasks) would benefit from additional insights from translation studies, and we suggest the introduction of an ‘evaluation brief” (analogous to the ‘translation brief’) which could help set out useful context for annotators tasked with evaluating machine translation.

Word-level translation quality estimation (TQE) is the task of identifying erroneous words in a translation with respect to the source. State-of-the-art methods for TQE exploit large quantities of synthetic training data generated from bilingual parallel corpora, where pseudo-quality labels are determined by comparing two independent translations for the same source text, i.e., an output from a machine translation (MT) system and a reference translation in the parallel corpora. However, this process is sorely reliant on the surface forms of words, with acceptable synonyms and interchangeable word orderings regarded as erroneous. This can potentially mislead the pre-training of models. In this paper, we describe a method that integrates a degree of uncertainty in labeling the words in synthetic training data for TQE. To estimate the extent to which each word in the MT output is likely to be correct or erroneous with respect to the reference translation, we propose to use the concept of optimal transport (OT), which exploits contextual word embeddings. Empirical experiments using a public benchmarking dataset for word-level TQE demonstrate that pre-training TQE models with the pseudo-quality labels determined by OT produces better predictions of the word-level quality labels determined by manual post-editing than doing so with surface-based pseudo-quality labels.

In machine translation, rare words continue to be a problem for the dominant encoder-decoder architecture, especially in low-resource and out-of-domain translation settings. Human translators solve this problem with monolingual or bilingual dictionaries. In this paper, we propose appending definitions from a bilingual dictionary to source sentences and using attention masking to link together rare words with their definitions. We find that including definitions for rare words improves performance by up to 1.0 BLEU and 1.6 MacroF1.

In this study, we explore the effectiveness of fine-tuning Large Language Models (LLMs), particularly Llama 3 8B Instruct, using translation memories (TMs) for hyper-specific machine translation (MT) tasks. Decoder-only LLMs have shown impressive performance in MT due to their ability to learn from extensive datasets and generate high quality translations. However, LLMs often struggle with the nuances and style required for organisation-specific translation so we leverage TMs, which store human translated segments, as a valuable resource to enhance translation accuracy and efficiency. We investigate the impact of fine-tuning the Llama 3 model using TMs from a specific organisation in the software sector. Our experiments cover five translation directions across languages of varying resource levels (English to Brazilian Portuguese, Czech, German, Finnish, and Korean). We analyse diverse sizes of training datasets (1k to 100k+ segments) to evaluate their influence on translation quality. We fine-tune separate models for each training set and evaluate their performance based on automatic metrics, BLEU, chrF++, TER, and COMET. Our findings reveal improvement in translation performance with larger datasets across all metrics. On average, BLEU and COMET scores increase by 13 and 25 points respectively on the largest training set against the baseline model. Notably, there is a performance deterioration in comparison with the baseline model when fine-tuning on only 1k and 2k examples; however, we observe a substantial improvement as the training dataset size increases. The study highlights the potential of integrating TMs with LLMs to create bespoke translation models tailored to the specific needs of businesses, therefore enhancing translation quality and reducing turn-around times. This approach offers a valuable insight for organisations seeking to leverage TMs and LLMs for optimal translation outcomes, specially in narrower domains.

This paper aims to investigate the presence of cognitive biases, more specifically of Availability heuristics, Representativeness heuristics and Framing, in OpenAI’s ChatGPT 3.5 and ChatGPT 4, as well as the linguistic dependency of their occurrences in the Large Language Models’ (LLMs) outputs. The innovative aspect of this research is conveyed by rephrasing three tasks proposed in Kahneman and Tversky’s works and determining whether the LLMs’ answers to the tasks are correct or incorrect and human-like or non-human-like. The latter classification is made possible by interviewing a total of 56 native speakers of Italian, English and Spanish, thus introducing a new linguistic comparison of results and forming a “human standard’. Our study indicates that GPTs 3.5 and 4 are very frequently subject to the cognitive biases under discussion and their answers are mostly non-human-like. There is minimal but significant discrepancy in the performance of GPT 3.5 and 4, slightly favouring ChatGPT 4 in avoiding biased responses, specifically for Availability heuristics. We also reveal that, while the results for ChatGPT 4 are not significantly language dependent, meaning that the performances in avoiding biases are not affected by the prompting language, their difference with ChatGPT 3.5 is statistically significant.