Serge Gladkoff

Lifeng Han, Gleb Erofeev, Irina Sorokina et al.

Pre-trained language models (PLMs) often take advantage of the monolingual and multilingual dataset that is freely available online to acquire general or mixed domain knowledge before deployment into specific tasks. Extra-large PLMs (xLPLMs) are proposed very recently to claim supreme performances over smaller-sized PLMs such as in machine translation (MT) tasks. These xLPLMs include Meta-AI's wmt21-dense-24-wide-en-X (2021) and NLLB (2022). In this work, we examine if xLPLMs are absolutely superior to smaller-sized PLMs in fine-tuning toward domain-specific MTs. We use two different in-domain data of different sizes: commercial automotive in-house data and clinical shared task data from the ClinSpEn2022 challenge at WMT2022. We choose popular Marian Helsinki as smaller sized PLM and two massive-sized Mega-Transformers from Meta-AI as xLPLMs. Our experimental investigation shows that 1) on smaller-sized in-domain commercial automotive data, xLPLM wmt21-dense-24-wide-en-X indeed shows much better evaluation scores using SacreBLEU and hLEPOR metrics than smaller-sized Marian, even though its score increase rate is lower than Marian after fine-tuning; 2) on relatively larger-size well prepared clinical data fine-tuning, the xLPLM NLLB tends to lose its advantage over smaller-sized Marian on two sub-tasks (clinical terms and ontology concepts) using ClinSpEn offered metrics METEOR, COMET, and ROUGE-L, and totally lost to Marian on Task-1 (clinical cases) on all official metrics including SacreBLEU and BLEU; 3) metrics do not always agree with each other on the same tasks using the same model outputs; 4) clinic-Marian ranked No.2 on Task- 1 (via SACREBLEU/BLEU) and Task-3 (via METEOR and ROUGE) among all submissions.

Lifeng Han, Gleb Erofeev, Irina Sorokina et al.

Massively multilingual pre-trained language models (MMPLMs) are developed in recent years demonstrating superpowers and the pre-knowledge they acquire for downstream tasks. This work investigates whether MMPLMs can be applied to clinical domain machine translation (MT) towards entirely unseen languages via transfer learning. We carry out an experimental investigation using Meta-AI's MMPLMs ``wmt21-dense-24-wide-en-X and X-en (WMT21fb)'' which were pre-trained on 7 language pairs and 14 translation directions including English to Czech, German, Hausa, Icelandic, Japanese, Russian, and Chinese, and the opposite direction. We fine-tune these MMPLMs towards English-\textit{Spanish} language pair which \textit{did not exist at all} in their original pre-trained corpora both implicitly and explicitly. We prepare carefully aligned \textit{clinical} domain data for this fine-tuning, which is different from their original mixed domain knowledge. Our experimental result shows that the fine-tuning is very successful using just 250k well-aligned in-domain EN-ES segments for three sub-task translation testings: clinical cases, clinical terms, and ontology concepts. It achieves very close evaluation scores to another MMPLM NLLB from Meta-AI, which included Spanish as a high-resource setting in the pre-training. To the best of our knowledge, this is the first work on using MMPLMs towards \textit{clinical domain transfer-learning NMT} successfully for totally unseen languages during pre-training.

Serge Gladkoff, Lifeng Han, Goran Nenadic

In natural language processing (NLP) we always rely on human judgement as the golden quality evaluation method. However, there has been an ongoing debate on how to better evaluate inter-rater reliability (IRR) levels for certain evaluation tasks, such as translation quality evaluation (TQE), especially when the data samples (observations) are very scarce. In this work, we first introduce the study on how to estimate the confidence interval for the measurement value when only one data (evaluation) point is available. Then, this leads to our example with two human-generated observational scores, for which, we introduce ``Student's \textit{t}-Distribution'' method and explain how to use it to measure the IRR score using only these two data points, as well as the confidence intervals (CIs) of the quality evaluation. We give quantitative analysis on how the evaluation confidence can be greatly improved by introducing more observations, even if only one extra observation. We encourage researchers to report their IRR scores in all possible means, e.g. using Student's \textit{t}-Distribution method whenever possible; thus making the NLP evaluation more meaningful, transparent, and trustworthy. This \textit{t}-Distribution method can be also used outside of NLP fields to measure IRR level for trustworthy evaluation of experimental investigations, whenever the observational data is scarce. Keywords: Inter-Rater Reliability (IRR); Scarce Observations; Confidence Intervals (CIs); Natural Language Processing (NLP); Translation Quality Evaluation (TQE); Student's \textit{t}-Distribution

Lifeng Han, Serge Gladkoff, Gleb Erofeev et al.

We conduct investigations on clinical text machine translation by examining multilingual neural network models using deep learning such as Transformer based structures. Furthermore, to address the language resource imbalance issue, we also carry out experiments using a transfer learning methodology based on massive multilingual pre-trained language models (MMPLMs). The experimental results on three subtasks including 1) clinical case (CC), 2) clinical terminology (CT), and 3) ontological concept (OC) show that our models achieved top-level performances in the ClinSpEn-2022 shared task on English-Spanish clinical domain data. Furthermore, our expert-based human evaluations demonstrate that the small-sized pre-trained language model (PLM) won over the other two extra-large language models by a large margin, in the clinical domain fine-tuning, which finding was never reported in the field. Finally, the transfer learning method works well in our experimental setting using the WMT21fb model to accommodate a new language space Spanish that was not seen at the pre-training stage within WMT21fb itself, which deserves more exploitation for clinical knowledge transformation, e.g. to investigate into more languages. These research findings can shed some light on domain-specific machine translation development, especially in clinical and healthcare fields. Further research projects can be carried out based on our work to improve healthcare text analytics and knowledge transformation. Our data will be openly available for research purposes at https://github.com/HECTA-UoM/ClinicalNMT

Serge Gladkoff, Lifeng Han, Gleb Erofeev et al.

Translation Quality Evaluation (TQE) is an essential step of the modern translation production process. TQE is critical in assessing both machine translation (MT) and human translation (HT) quality without reference translations. The ability to evaluate or even simply estimate the quality of translation automatically may open significant efficiency gains through process optimisation. This work examines whether the state-of-the-art large language models (LLMs) can be used for this purpose. We take OpenAI models as the best state-of-the-art technology and approach TQE as a binary classification task. On eight language pairs including English to Italian, German, French, Japanese, Dutch, Portuguese, Turkish, and Chinese, our experimental results show that fine-tuned gpt3.5 can demonstrate good performance on translation quality prediction tasks, i.e. whether the translation needs to be edited. Another finding is that simply increasing the sizes of LLMs does not lead to apparent better performances on this task by comparing the performance of three different versions of OpenAI models: curie, davinci, and gpt3.5 with 13B, 175B, and 175B parameters, respectively.

Serge Gladkoff, Lifeng Han

Traditional automatic evaluation metrics for machine translation have been widely criticized by linguists due to their low accuracy, lack of transparency, focus on language mechanics rather than semantics, and low agreement with human quality evaluation. Human evaluations in the form of MQM-like scorecards have always been carried out in real industry setting by both clients and translation service providers (TSPs). However, traditional human translation quality evaluations are costly to perform and go into great linguistic detail, raise issues as to inter-rater reliability (IRR) and are not designed to measure quality of worse than premium quality translations. In this work, we introduce HOPE, a task-oriented and human-centric evaluation framework for machine translation output based on professional post-editing annotations. It contains only a limited number of commonly occurring error types, and use a scoring model with geometric progression of error penalty points (EPPs) reflecting error severity level to each translation unit. The initial experimental work carried out on English-Russian language pair MT outputs on marketing content type of text from highly technical domain reveals that our evaluation framework is quite effective in reflecting the MT output quality regarding both overall system-level performance and segment-level transparency, and it increases the IRR for error type interpretation. The approach has several key advantages, such as ability to measure and compare less than perfect MT output from different systems, ability to indicate human perception of quality, immediate estimation of the labor effort required to bring MT output to premium quality, low-cost and faster application, as well as higher IRR. Our experimental data is available at \url{https://github.com/lHan87/HOPE}.

Lifeng Han, Irina Sorokina, Gleb Erofeev et al.

Human evaluation has always been expensive while researchers struggle to trust the automatic metrics. To address this, we propose to customise traditional metrics by taking advantages of the pre-trained language models (PLMs) and the limited available human labelled scores. We first re-introduce the hLEPOR metric factors, followed by the Python version we developed (ported) which achieved the automatic tuning of the weighting parameters in hLEPOR metric. Then we present the customised hLEPOR (cushLEPOR) which uses Optuna hyper-parameter optimisation framework to fine-tune hLEPOR weighting parameters towards better agreement to pre-trained language models (using LaBSE) regarding the exact MT language pairs that cushLEPOR is deployed to. We also optimise cushLEPOR towards professional human evaluation data based on MQM and pSQM framework on English-German and Chinese-English language pairs. The experimental investigations show cushLEPOR boosts hLEPOR performances towards better agreements to PLMs like LaBSE with much lower cost, and better agreements to human evaluations including MQM and pSQM scores, and yields much better performances than BLEU (data available at \url{https://github.com/poethan/cushLEPOR}). Official results show that our submissions win three language pairs including \textbf{English-German} and \textbf{Chinese-English} on \textit{News} domain via cushLEPOR(LM) and \textbf{English-Russian} on \textit{TED} domain via hLEPOR.

Lifeng Han, Serge Gladkoff

Starting from the 1950s, Machine Translation (MT) was challenged by different scientific solutions, which included rule-based methods, example-based and statistical models (SMT), to hybrid models, and very recent years the neural models (NMT). While NMT has achieved a huge quality improvement in comparison to conventional methodologies, by taking advantage of a huge amount of parallel corpora available from the internet and the recently developed super computational power support with an acceptable cost, it struggles to achieve real human parity in many domains and most language pairs, if not all of them. Alongside the long road of MT research and development, quality evaluation metrics played very important roles in MT advancement and evolution. In this tutorial, we overview the traditional human judgement criteria, automatic evaluation metrics, unsupervised quality estimation models, as well as the meta-evaluation of the evaluation methods. Among these, we will also cover the very recent work in the MT evaluation (MTE) fields, taking advantage of the large size of pre-trained language models for automatic metric customisation towards exactly deployed language pairs and domains. In addition, we also introduce the statistical confidence estimation regarding the sample size needed for human evaluation in real practice simulation. Full tutorial material is \textbf{available} to download at https://github.com/poethan/LREC22_MetaEval_Tutorial.

Serge Gladkoff, Lifeng Han, Katerina Gasova

Analytic Translation Quality Evaluation (TQE), based on Multidimensional Quality Metrics (MQM), traditionally uses a linear error-to-penalty scale calibrated to a reference sample of 1000-2000 words. However, linear extrapolation biases judgment on samples of different sizes, over-penalizing short samples and under-penalizing long ones, producing misalignment with expert intuition. Building on the Multi-Range framework, this paper presents a calibrated, non-linear scoring model that better reflects how human content consumers perceive translation quality across samples of varying length. Empirical data from three large-scale enterprise environments shows that acceptable error counts grow logarithmically, not linearly, with sample size. Psychophysical and cognitive evidence, including the Weber-Fechner law and Cognitive Load Theory, supports this premise by explaining why the perceptual impact of additional errors diminishes while the cognitive burden grows with scale. We propose a two-parameter model E(x) = a * ln(1 + b * x), a, b > 0, anchored to a reference tolerance and calibrated from two tolerance points using a one-dimensional root-finding step. The model yields an explicit interval within which the linear approximation stays within +/-20 percent relative error and integrates into existing evaluation workflows with only a dynamic tolerance function added. The approach improves interpretability, fairness, and inter-rater reliability across both human and AI-generated translations. By operationalizing a perceptually valid scoring paradigm, it advances translation quality evaluation toward more accurate and scalable assessment. The model also provides a stronger basis for AI-based document-level evaluation aligned with human judgment. Implementation considerations for CAT/LQA systems and implications for human and AI-generated text evaluation are discussed.

Serge Gladkoff, Irina Sorokina, Lifeng Han et al.

From both human translators (HT) and machine translation (MT) researchers' point of view, translation quality evaluation (TQE) is an essential task. Translation service providers (TSPs) have to deliver large volumes of translations which meet customer specifications with harsh constraints of required quality level in tight time-frames and costs. MT researchers strive to make their models better, which also requires reliable quality evaluation. While automatic machine translation evaluation (MTE) metrics and quality estimation (QE) tools are widely available and easy to access, existing automated tools are not good enough, and human assessment from professional translators (HAP) are often chosen as the golden standard \cite{han-etal-2021-TQA}. Human evaluations, however, are often accused of having low reliability and agreement. Is this caused by subjectivity or statistics is at play? How to avoid the entire text to be checked and be more efficient with TQE from cost and efficiency perspectives, and what is the optimal sample size of the translated text, so as to reliably estimate the translation quality of the entire material? This work carries out such motivated research to correctly estimate the confidence intervals \cite{Brown_etal2001Interval} depending on the sample size of the translated text, e.g. the amount of words or sentences, that needs to be processed on TQE workflow step for confident and reliable evaluation of overall translation quality. The methodology we applied for this work is from Bernoulli Statistical Distribution Modelling (BSDM) and Monte Carlo Sampling Analysis (MCSA).