Byte-Pair Encoding (BPE) is a widely used method for subword tokenization, with origins in grammar-based text compression. It is employed in a variety of language processing tasks such as machine translation or large language model (LLM) pretraining, to create a token dictionary of a prescribed size. Most evaluations of BPE to date are empirical, and the reasons for its good practical performance are not well understood. In this paper we focus on the optimization problem underlying BPE: finding a pair encoding that achieves optimal compression utility. We show that this problem is APX-complete, indicating that it is unlikely to admit a polynomial-time approximation scheme. This answers, in a stronger form, a question recently raised by Zouhar et al. On the positive side, we show that BPE approximates the compression utility of the optimal pair encoding to a worst-case factor between $1/3$ and $(1-e^{-2})/2 \approx 0.4324$. Our results aim to explain the ongoing success of BPE and are, to our knowledge, the first rigorous guarantees on its compression utility that hold for all inputs.
Translating wordplay across languages presents unique challenges that have long confounded both professional human translators and machine translation systems. This research proposes a novel approach for translating puns from English to French by combining state-of-the-art large language models with specialized techniques for wordplay generation. Our methodology employs a three-stage approach. First, we establish a baseline using multiple frontier large language models with feedback based on a new contrastive learning dataset. Second, we implement a guided chain-of-thought pipeline with combined phonetic-semantic embeddings. Third, we implement a multi-agent generator-discriminator framework for evaluating and regenerating puns with feedback. Moving beyond the limitations of literal translation, our methodology's primary objective is to capture the linguistic creativity and humor of the source text wordplay, rather than simply duplicating its vocabulary. Our best runs earned first and second place in the CLEF JOKER 2025 Task 2 competition where they were evaluated manually by expert native French speakers. This research addresses a gap between translation studies and computational linguistics by implementing linguistically-informed techniques for wordplay translation, advancing our understanding of how language models can be leveraged to handle the complex interplay between semantic ambiguity, phonetic similarity, and the implicit cultural and linguistic awareness needed for successful humor.
Document level Machine Translation (DocMT) approaches often struggle with effectively capturing discourse level phenomena. Existing approaches rely on heuristic rules to segment documents into discourse units, which rarely align with the true discourse structure required for accurate translation. Otherwise, they fail to maintain consistency throughout the document during translation. To address these challenges, we propose Graph Augmented Agentic Framework for Document Level Translation (GRAFT), a novel graph based DocMT system that leverages Large Language Model (LLM) agents for document translation. Our approach integrates segmentation, directed acyclic graph (DAG) based dependency modelling, and discourse aware translation into a cohesive framework. Experiments conducted across eight translation directions and six diverse domains demonstrate that GRAFT achieves significant performance gains over state of the art DocMT systems. Specifically, GRAFT delivers an average improvement of 2.8 d BLEU on the TED test sets from IWSLT2017 over strong baselines and 2.3 d BLEU for domain specific translation from English to Chinese. Moreover, our analyses highlight the consistent ability of GRAFT to address discourse level phenomena, yielding coherent and contextually accurate translations.
Multi-agent systems empowered by large language models (LLMs) have demonstrated remarkable capabilities in a wide range of downstream applications, including machine translation. However, the potential of LLMs in translating Hong Kong legal judgments remains uncertain due to challenges such as intricate legal terminology, culturally embedded nuances, and strict linguistic structures. In this work, we introduce TransLaw, a novel multi-agent framework implemented for real-world Hong Kong case law translation. It employs three specialized agents, namely, Translator, Annotator, and Proofreader, to collaboratively produce translations for high accuracy in legal meaning, appropriateness in style, and adequate coherence and cohesion in structure. This framework supports customizable LLM configurations and achieves tremendous cost reduction compared to professional human translation services. We evaluated its performance using 13 open-source and commercial LLMs as agents and obtained interesting findings, including that it surpasses GPT-4o in legal semantic accuracy, structural coherence, and stylistic fidelity, yet trails human experts in contextualizing complex terminology and stylistic naturalness. Our platform website is available at CityUHK, and our bilingual judgment corpus used for the evaluation is available at Hugging Face.
Helping deaf and hard-of-hearing people communicate more easily is the main goal of Automatic Sign Language Translation. Although most past research has focused on turning sign language into text, doing the reverse, turning spoken English into sign language animations, has been largely overlooked. That's because it involves multiple steps, such as understanding speech, translating it into sign-friendly grammar, and generating natural human motion. In this work, we introduce a complete pipeline that converts English speech into smooth, realistic 3D sign language animations. Our system starts with Whisper to translate spoken English into text. Then, we use a MarianMT machine translation model to translate that text into American Sign Language (ASL) gloss, a simplified version of sign language that captures meaning without grammar. This model performs well, reaching BLEU scores of 0.7714 and 0.8923. To make the gloss translation more accurate, we also use word embeddings such as Word2Vec and FastText to understand word meanings. Finally, we animate the translated gloss using a 3D keypoint-based motion system trained on Sign3D-WLASL, a dataset we created by extracting body, hand, and face key points from real ASL videos in the WLASL dataset. To support the gloss translation stage, we also built a new dataset called BookGlossCorpus-CG, which turns everyday English sentences from the BookCorpus dataset into ASL gloss using grammar rules. Our system stitches everything together by smoothly interpolating between signs to create natural, continuous animations. Unlike previous works like How2Sign and Phoenix-2014T that focus on recognition or use only one type of data, our pipeline brings together audio, text, and motion in a single framework that goes all the way from spoken English to lifelike 3D sign language animation.
In Machine Translation (MT) evaluation, metric performance is assessed based on agreement with human judgments. In recent years, automatic metrics have demonstrated increasingly high levels of agreement with humans. To gain a clearer understanding of metric performance and establish an upper bound, we incorporate human baselines in the MT meta-evaluation, that is, the assessment of MT metrics' capabilities. Our results show that human annotators are not consistently superior to automatic metrics, with state-of-the-art metrics often ranking on par with or higher than human baselines. Despite these findings suggesting human parity, we discuss several reasons for caution. Finally, we explore the broader implications of our results for the research field, asking: Can we still reliably measure improvements in MT evaluation? With this work, we aim to shed light on the limits of our ability to measure progress in the field, fostering discussion on an issue that we believe is crucial to the entire MT evaluation community.
Automatic Term Extraction (ATE) identifies domain-specific expressions that are crucial for downstream tasks such as machine translation and information retrieval. Although large language models (LLMs) have significantly advanced various NLP tasks, their potential for ATE has scarcely been examined. We propose a retrieval-based prompting strategy that, in the few-shot setting, selects demonstrations according to \emph{syntactic} rather than semantic similarity. This syntactic retrieval method is domain-agnostic and provides more reliable guidance for capturing term boundaries. We evaluate the approach in both in-domain and cross-domain settings, analyzing how lexical overlap between the query sentence and its retrieved examples affects performance. Experiments on three specialized ATE benchmarks show that syntactic retrieval improves F1-score. These findings highlight the importance of syntactic cues when adapting LLMs to terminology-extraction tasks.
Estimation of semantic similarity is an important research problem both in natural language processing and the natural language understanding, and that has tremendous application on various downstream tasks such as question answering, semantic search, information retrieval, document clustering, word-sense disambiguation and machine translation. In this work, we carry out the estimation of semantic similarity using different state-of-the-art techniques including the USE (Universal Sentence Encoder), InferSent and the most recent BERT, or Bidirectional Encoder Representations from Transformers, models. We use two question pairs datasets for the analysis, one is a domain specific in-house dataset and the other is a public dataset which is the Quora's question pairs dataset. We observe that the BERT model gave much superior performance as compared to the other methods. This should be because of the fine-tuning procedure that is involved in its training process, allowing it to learn patterns based on the training data that is used. This works demonstrates the applicability of BERT on domain specific datasets. We infer from the analysis that BERT is the best technique to use in the case of domain specific data.
Past work has long recognized the important role of context in guiding how humans search their memory. While context-based memory models can explain many memory phenomena, it remains unclear why humans develop such architectures over possible alternatives in the first place. In this work, we demonstrate that foundational architectures in neural machine translation -- specifically, recurrent neural network (RNN)-based sequence-to-sequence models with attention -- exhibit mechanisms that directly correspond to those specified in the Context Maintenance and Retrieval (CMR) model of human memory. Since neural machine translation models have evolved to optimize task performance, their convergence with human memory models provides a deeper understanding of the functional role of context in human memory, as well as presenting new ways to model human memory. Leveraging this convergence, we implement a neural machine translation model as a cognitive model of human memory search that is both interpretable and capable of capturing complex dynamics of learning. We show that our model accounts for both averaged and optimal human behavioral patterns as effectively as context-based memory models. Further, we demonstrate additional strengths of the proposed model by evaluating how memory search performance emerges from the interaction of different model components.
Human evaluators provide necessary contributions in evaluating large language models. In the context of Machine Translation (MT) systems for low-resource languages (LRLs), this is made even more apparent since popular automated metrics tend to be string-based, and therefore do not provide a full picture of the nuances of the behavior of the system. Human evaluators, when equipped with the necessary expertise of the language, will be able to test for adequacy, fluency, and other important metrics. However, the low resource nature of the language means that both datasets and evaluators are in short supply. This presents the following conundrum: How can developers of MT systems for these LRLs find adequate human evaluators and datasets? This paper first presents a comprehensive review of existing evaluation procedures, with the objective of producing a design proposal for a platform that addresses the resource gap in terms of datasets and evaluators in developing MT systems. The result is a design for a recruitment and gamified evaluation platform for developers of MT systems. Challenges are also discussed in terms of evaluating this platform, as well as its possible applications in the wider scope of Natural Language Processing (NLP) research.
In this paper, we compare Czech-specific and multilingual sentence embedding models through intrinsic and extrinsic evaluation paradigms. For intrinsic evaluation, we employ Costra, a complex sentence transformation dataset, and several Semantic Textual Similarity (STS) benchmarks to assess the ability of the embeddings to capture linguistic phenomena such as semantic similarity, temporal aspects, and stylistic variations. In the extrinsic evaluation, we fine-tune each embedding model using COMET-based metrics for machine translation evaluation. Our experiments reveal an interesting disconnect: models that excel in intrinsic semantic similarity tests do not consistently yield superior performance on downstream translation evaluation tasks. Conversely, models with seemingly over-smoothed embedding spaces can, through fine-tuning, achieve excellent results. These findings highlight the complex relationship between semantic property probes and downstream task, emphasizing the need for more research into 'operationalizable semantics' in sentence embeddings, or more in-depth downstream tasks datasets (here translation evaluation)
Large language models (LLMs), despite their ability to perform few-shot machine translation (MT), often lag behind dedicated MT systems trained on parallel corpora, which are crucial for high quality machine translation (MT). However, parallel corpora are often scarce or non-existent for low-resource languages. In this paper, we propose CycleDistill, a bootstrapping approach leveraging LLMs and few-shot translation to obtain high-quality MT systems. CycleDistill involves iteratively generating synthetic parallel corpora from monolingual corpora via zero- or few-shot MT, which is then used to fine-tune the model that was used for generating said data for MT. CycleDistill does not need parallel corpora beyond 1 to 4 few-shot examples, and in our experiments focusing on three Indian languages, by relying solely on monolingual corpora, it can achieve high-quality machine translation, improving upon a few-shot baseline model by over 20-30 chrF points on average in the first iteration. We also study the effect of leveraging softmax activations during the distillation process and observe mild improvements in translation quality.
Evaluating consistency in large language models (LLMs) is crucial for ensuring reliability, particularly in complex, multi-step interactions between humans and LLMs. Traditional self-consistency methods often miss subtle semantic changes in natural language and functional shifts in code or equations, which can accumulate over multiple transformations. To address this, we propose ConsistencyChecker, a tree-based evaluation framework designed to measure consistency through sequences of reversible transformations, including machine translation tasks and AI-assisted programming tasks. In our framework, nodes represent distinct text states, while edges correspond to pairs of inverse operations. Dynamic and LLM-generated benchmarks ensure a fair assessment of the model's generalization ability and eliminate benchmark leakage. Consistency is quantified based on similarity across different depths of the transformation tree. Experiments on eight models from various families and sizes show that ConsistencyChecker can distinguish the performance of different models. Notably, our consistency scores-computed entirely without using WMT paired data-correlate strongly (r > 0.7) with WMT 2024 auto-ranking, demonstrating the validity of our benchmark-free approach. Our implementation is available at: https://github.com/ulab-uiuc/consistencychecker.
Deep autoencoders have become a fundamental tool in various machine learning applications, ranging from dimensionality reduction and reduced order modeling of partial differential equations to anomaly detection and neural machine translation. Despite their empirical success, a solid theoretical foundation for their expressiveness remains elusive, particularly when compared to classical projection-based techniques. In this work, we aim to take a step forward in this direction by presenting a comprehensive analysis of what we refer to as symmetric autoencoders, a broad class of deep learning architectures ubiquitous in the literature. Specifically, we introduce a formal distinction between different classes of symmetric architectures, analyzing their strengths and limitations from a mathematical perspective. For instance, we show that the reconstruction error of symmetric autoencoders with orthonormality constraints can be understood by leveraging the well-renowned Eckart-Young-Schmidt (EYS) theorem. As a byproduct of our analysis, we end up developing the EYS initialization strategy for symmetric autoencoders, which is based on an iterated application of the Singular Value Decomposition (SVD). To validate our findings, we conduct a series of numerical experiments where we benchmark our proposal against conventional deep autoencoders, discussing the importance of model design and initialization.
As the demand for inclusive language increases, concern has grown over the susceptibility of machine translation (MT) systems to reinforce gender stereotypes. This study investigates gender bias in two commercial MT systems, Google Translate and DeepL, focusing on the understudied English-to-Greek language pair. We address three aspects of gender bias: i) male bias, ii) occupational stereotyping, and iii) errors in anti-stereotypical translations. Additionally, we explore the potential of prompted GPT-4o as a bias mitigation tool that provides both gender-explicit and gender-neutral alternatives when necessary. To achieve this, we introduce GendEL, a manually crafted bilingual dataset of 240 gender-ambiguous and unambiguous sentences that feature stereotypical occupational nouns and adjectives. We find persistent gender bias in translations by both MT systems; while they perform well in cases where gender is explicitly defined, with DeepL outperforming both Google Translate and GPT-4o in feminine gender-unambiguous sentences, they are far from producing gender-inclusive or neutral translations when the gender is unspecified. GPT-4o shows promise, generating appropriate gendered and neutral alternatives for most ambiguous cases, though residual biases remain evident.
Machine translation has long been a central task in natural language processing. With the rapid advancement of large language models (LLMs), there has been remarkable progress in translation quality. However, fully realizing the translation potential of LLMs remains an open challenge. Recent studies have explored multi-agent systems to decompose complex translation tasks into collaborative subtasks, showing initial promise in enhancing translation quality through agent cooperation and specialization. Nevertheless, existing multi-agent translation frameworks largely neglect foundational insights from cognitive translation studies. These insights emphasize how human translators employ different cognitive strategies, such as balancing literal and free translation, refining expressions based on context, and iteratively evaluating outputs. To address this limitation, we propose a cognitively informed multi-agent framework called TACTIC, which stands for T ranslation A gents with Cognitive- T heoretic Interactive Collaboration. The framework comprises six functionally distinct agents that mirror key cognitive processes observed in human translation behavior. These include agents for drafting, refinement, evaluation, scoring, context reasoning, and external knowledge gathering. By simulating an interactive and theory-grounded translation workflow, TACTIC effectively leverages the full capacity of LLMs for high-quality translation. Experimental results on diverse language pairs from the FLORES-200 and WMT24 benchmarks show that our method consistently achieves state-of-the-art performance. Using DeepSeek-V3 as the base model, TACTIC surpasses GPT-4.1 by an average of +0.6 XCOMET and +1.18 COMETKIWI-23. Compared to DeepSeek-R1, it further improves by +0.84 XCOMET and +2.99 COMETKIWI-23. Code is available at https://github.com/weiyali126/TACTIC.
This report presents EuroLLM-9B, a large language model trained from scratch to support the needs of European citizens by covering all 24 official European Union languages and 11 additional languages. EuroLLM addresses the issue of European languages being underrepresented and underserved in existing open large language models. We provide a comprehensive overview of EuroLLM-9B's development, including tokenizer design, architectural specifications, data filtering, and training procedures. We describe the pre-training data collection and filtering pipeline, including the creation of EuroFilter, an AI-based multilingual filter, as well as the design of EuroBlocks-Synthetic, a novel synthetic dataset for post-training that enhances language coverage for European languages. Evaluation results demonstrate EuroLLM-9B's competitive performance on multilingual benchmarks and machine translation tasks, establishing it as the leading open European-made LLM of its size. To support open research and adoption, we release all major components of this work, including the base and instruction-tuned models, the EuroFilter classifier, and the synthetic post-training dataset.
Despite doubts on data quality, instruction synthesis has been widely applied into instruction tuning (IT) of LLMs as an economic and rapid alternative. Recent endeavors focus on improving data quality for synthesized instruction pairs in English and have facilitated IT of English-centric LLMs. However, data quality issues in multilingual synthesized instruction pairs are even more severe, since the common synthesizing practice is to translate English synthesized data into other languages using machine translation (MT). Besides the known content errors in these English synthesized data, multilingual synthesized instruction data are further exposed to defects introduced by MT and face insufficient localization of the target languages. In this paper, we propose MIDB, a Multilingual Instruction Data Booster to automatically address the quality issues in multilingual synthesized data. MIDB is trained on around 36.8k revision examples across 16 languages by human linguistic experts, thereby can boost the low-quality data by addressing content errors and MT defects, and improving localization in these synthesized data. Both automatic and human evaluation indicate that not only MIDB steadily improved instruction data quality in 16 languages, but also the instruction-following and cultural-understanding abilities of multilingual LLMs fine-tuned on MIDB-boosted data were significantly enhanced.
Large Language Models (LLMs) have shown to be effective evaluators across various domains such as machine translations or the scientific domain. Current LLM-as-a-Judge approaches rely mostly on individual assessments or a single round of pairwise assessments, preventing the judge LLM from developing a global ranking perspective. To address this, we present Knockout Assessment, an LLM-asa Judge method using a knockout tournament system with iterative pairwise comparisons. Experiments across three LLMs on two datasets show that knockout assessment improves scoring accuracy, increasing Pearson correlation with expert evaluations by 0.07 on average for university-level exam scoring and machine translation evaluations, aligning LLM assessments more closely with human scoring.
Large language models (LLMs) possess strong multilingual capabilities, and combining Reinforcement Learning from Human Feedback (RLHF) with translation tasks has shown great potential. However, we observe that this paradigm performs unexpectedly poorly when applied to colloquial subtitle translation tasks. In this work, we investigate this issue and find that the offline reward model (RM) gradually diverges from the online LLM due to distributional shift, ultimately leading to undesirable training outcomes. To address this, we propose RIVAL, an adversarial training framework that formulates the process as a min-max game between the RM and the LLM. RIVAL iteratively updates the both models, with the RM trained to distinguish strong from weak translations (qualitative preference reward), and the LLM trained to enhance its translation for closing this gap. To stabilize training and improve generalizability, we also incorporate quantitative preference reward (e.g., BLEU) into the RM, enabling reference-free quality modeling aligned with human evaluation. Through extensive experiments, we demonstrate that the proposed adversarial training framework significantly improves upon translation baselines.