Mark Fishel

Tom Kocmi, Eleftherios Avramidis, Rachel Bawden et al. · eth-zurich, microsoft-research

This is the preliminary ranking of WMT24 General MT systems based on automatic metrics. The official ranking will be a human evaluation, which is superior to the automatic ranking and supersedes it. The purpose of this report is not to interpret any findings but only provide preliminary results to the participants of the General MT task that may be useful during the writing of the system submission.

Maksym Del, Mark Fishel

Related works used indexes like CKA and variants of CCA to measure the similarity of cross-lingual representations in multilingual language models. In this paper, we argue that assumptions of CKA/CCA align poorly with one of the motivating goals of cross-lingual learning analysis, i.e., explaining zero-shot cross-lingual transfer. We highlight what valuable aspects of cross-lingual similarity these indexes fail to capture and provide a motivating case study \textit{demonstrating the problem empirically}. Then, we introduce \textit{Average Neuron-Wise Correlation (ANC)} as a straightforward alternative that is exempt from the difficulties of CKA/CCA and is good specifically in a cross-lingual context. Finally, we use ANC to construct evidence that the previously introduced ``first align, then predict'' pattern takes place not only in masked language models (MLMs) but also in multilingual models with \textit{causal language modeling} objectives (CLMs). Moreover, we show that the pattern extends to the \textit{scaled versions} of the MLMs and CLMs (up to 85x original mBERT).\footnote{Our code is publicly available at \url{https://github.com/TartuNLP/xsim}}

Taido Purason, Pavel Chizhov, Ivan P. Yamshchikov et al.

Tokenizer adaptation plays an important role in transferring pre-trained language models to new domains or languages. In this work, we address two complementary aspects of this process: vocabulary extension and pruning. The common approach to extension trains a new tokenizer on domain-specific text and appends the tokens that do not overlap with the existing vocabulary, which often results in many tokens that are unreachable or never used. We propose continued BPE training, which adapts a pre-trained tokenizer by continuing the BPE merge learning process on new data. Experiments across multiple languages and model families show that this approach improves tokenization efficiency and leads to better utilization of added vocabulary. We also introduce leaf-based vocabulary pruning, which removes redundant tokens while preserving model quality. Together, these methods provide practical tools for controlled vocabulary modification, which we release as an open-source package.

Maksym Del, Mark Fishel

Large language models (LLMs) have demonstrated solid zero-shot reasoning capabilities, which is reflected in their performance on the current test tasks. This calls for a more challenging benchmark requiring highly advanced reasoning ability to be solved. In this paper, we introduce such a benchmark, consisting of 191 long-form (1200 words on average) mystery narratives constructed as detective puzzles. Puzzles are sourced from the "5 Minute Mystery" platform and include a multiple-choice question for evaluation. Only 47% of humans solve a puzzle successfully on average, while the best human solvers achieve over 80% success rate. We show that GPT-3 models barely outperform random on this benchmark (with 28% accuracy) while state-of-the-art GPT-4 solves only 38% of puzzles. This indicates that there is still a significant gap in the deep reasoning abilities of LLMs and humans and highlights the need for further research in this area. Our work introduces a challenging benchmark for future studies on reasoning in language models and contributes to a better understanding of the limits of LLMs' abilities.

Maksym Del, Markus Kängsepp, Marharyta Domnich et al.

Uncertainty estimation is critical for deploying reasoning language models, yet remains poorly understood under extended chain-of-thought reasoning. We study parallel sampling as a fully black-box approach using verbalized confidence and self-consistency. Across three reasoning models and 17 tasks spanning mathematics, STEM, and humanities, we characterize how these signals scale. Both self-consistency and verbalized confidence scale in reasoning models, but self-consistency exhibits lower initial discrimination and lags behind verbalized confidence under moderate sampling. Most uncertainty gains, however, arise from signal combination: with just two samples, a hybrid estimator improves AUROC by up to $+12$ on average and already outperforms either signal alone even when scaled to much larger budgets, after which returns diminish. These effects are domain-dependent: in mathematics, the native domain of RLVR-style post-training, reasoning models achieve higher uncertainty quality and exhibit both stronger complementarity and faster scaling than in STEM or humanities.

Hele-Andra Kuulmets, Taido Purason, Agnes Luhtaru et al.

This paper explores cost-efficient methods to adapt pretrained Large Language Models (LLMs) to new lower-resource languages, with a specific focus on Estonian. Leveraging the Llama 2 model, we investigate the impact of combining cross-lingual instruction-tuning with additional monolingual pretraining. Our results demonstrate that even a relatively small amount of additional monolingual pretraining followed by cross-lingual instruction-tuning significantly enhances results on Estonian. Furthermore, we showcase cross-lingual knowledge transfer from high-quality English instructions to Estonian, resulting in improvements in commonsense reasoning and multi-turn conversation capabilities. Our best model, named \textsc{Llammas}, represents the first open-source instruction-following LLM for Estonian. Additionally, we publish Alpaca-est, the first general task instruction dataset for Estonia. These contributions mark the initial progress in the direction of developing open-source LLMs for Estonian.

Agnes Luhtaru, Martin Vainikko, Krista Liin et al.

The project was funded in 2021-2023 by the National Programme of Estonian Language Technology. Its main aim was to develop spelling and grammar correction tools for the Estonian language. The main challenge was the very small amount of available error correction data needed for such development. To mitigate this, (1) we annotated more correction data for model training and testing, (2) we tested transfer-learning, i.e. retraining machine learning models created for other tasks, so as not to depend solely on correction data, (3) we compared the developed method and model with alternatives, including large language models. We also developed automatic evaluation, which can calculate the accuracy and yield of corrections by error category, so that the effectiveness of different methods can be compared in detail. There has been a breakthrough in large language models during the project: GPT4, a commercial language model with Estonian-language support, has been created. We took into account the existence of the model when adjusting plans and in the report we present a comparison with the ability of GPT4 to improve the Estonian language text. The final results show that the approach we have developed provides better scores than GPT4 and the result is usable but not entirely reliable yet. The report also contains ideas on how GPT4 and other major language models can be implemented in the future, focusing on open-source solutions. All results of this project are open-data/open-source, with licenses that allow them to be used for purposes including commercial ones.

Maksym Del, Mark Fishel

Low-resource languages, such as Baltic languages, benefit from Large Multilingual Models (LMs) that possess remarkable cross-lingual transfer performance capabilities. This work is an interpretation and analysis study into cross-lingual representations of Multilingual LMs. Previous works hypothesized that these LMs internally project representations of different languages into a shared cross-lingual space. However, the literature produced contradictory results. In this paper, we revisit the prior work claiming that "BERT is not an Interlingua" and show that different languages do converge to a shared space in such language models with another choice of pooling strategy or similarity index. Then, we perform cross-lingual representational analysis for the two most popular multilingual LMs employing 378 pairwise language comparisons. We discover that while most languages share joint cross-lingual space, some do not. However, we observe that Baltic languages do belong to that shared space. The code is available at https://github.com/TartuNLP/xsim.

Agnes Luhtaru, Taido Purason, Martin Vainikko et al.

This study explores enhancing grammatical error correction (GEC) through artificial error generation (AEG) using language models (LMs). Specifically, we fine-tune Llama 2-based LMs for error generation and find that this approach yields synthetic errors akin to human errors. Next, we train GEC Llama models with the help of these artificial errors and outperform previous state-of-the-art error correction models, with gains ranging between 0.8 and 6 F0.5 points across all tested languages (German, Ukrainian, and Estonian). Moreover, we demonstrate that generating errors by fine-tuning smaller sequence-to-sequence models and prompting large commercial LMs (GPT-3.5 and GPT-4) also results in synthetic errors beneficially affecting error generation models.

Taido Purason, Hele-Andra Kuulmets, Mark Fishel

The advancement of large language models (LLMs) has predominantly focused on high-resource languages, leaving low-resource languages, such as those in the Finno-Ugric family, significantly underrepresented. This paper addresses this gap by focusing on Võro, Livonian, and Komi. We cover almost the entire cycle of LLM creation, from data collection to instruction tuning and evaluation. Our contributions include developing multilingual base and instruction-tuned models; creating evaluation benchmarks, including the smugri-MT-bench multi-turn conversational benchmark; and conducting human evaluation. We intend for this work to promote linguistic diversity, ensuring that lesser-resourced languages can benefit from advancements in NLP.

Tom Kocmi, Eleftherios Avramidis, Rachel Bawden et al. · eth-zurich, microsoft-research

We present the preliminary rankings of machine translation (MT) systems submitted to the WMT25 General Machine Translation Shared Task, as determined by automatic evaluation metrics. Because these rankings are derived from automatic evaluation, they may exhibit a bias toward systems that employ re-ranking techniques, such as Quality Estimation or Minimum Bayes Risk decoding. The official WMT25 ranking will be based on human evaluation, which is more reliable and will supersede these results. The official WMT25 ranking will be based on human evaluation, which is more reliable and will supersede these results. The purpose of releasing these findings now is to assist task participants with their system description papers; not to provide final findings.

Maksym Del, Elizaveta Korotkova, Mark Fishel

Many works proposed methods to improve the performance of Neural Machine Translation (NMT) models in a domain/multi-domain adaptation scenario. However, an understanding of how NMT baselines represent text domain information internally is still lacking. Here we analyze the sentence representations learned by NMT Transformers and show that these explicitly include the information on text domains, even after only seeing the input sentences without domains labels. Furthermore, we show that this internal information is enough to cluster sentences by their underlying domains without supervision. We show that NMT models produce clusters better aligned to the actual domains compared to pre-trained language models (LMs). Notably, when computed on document-level, NMT cluster-to-domain correspondence nears 100%. We use these findings together with an approach to NMT domain adaptation using automatically extracted domains. Whereas previous work relied on external LMs for text clustering, we propose re-using the NMT model as a source of unsupervised clusters. We perform an extensive experimental study comparing two approaches across two data scenarios, three language pairs, and both sentence-level and document-level clustering, showing equal or significantly superior performance compared to LMs.

Marina Fomicheva, Shuo Sun, Lisa Yankovskaya et al.

Quality Estimation (QE) is an important component in making Machine Translation (MT) useful in real-world applications, as it is aimed to inform the user on the quality of the MT output at test time. Existing approaches require large amounts of expert annotated data, computation and time for training. As an alternative, we devise an unsupervised approach to QE where no training or access to additional resources besides the MT system itself is required. Different from most of the current work that treats the MT system as a black box, we explore useful information that can be extracted from the MT system as a by-product of translation. By employing methods for uncertainty quantification, we achieve very good correlation with human judgments of quality, rivalling state-of-the-art supervised QE models. To evaluate our approach we collect the first dataset that enables work on both black-box and glass-box approaches to QE.

Elizaveta Korotkova, Agnes Luhtaru, Maksym Del et al.

Both grammatical error correction and text style transfer can be viewed as monolingual sequence-to-sequence transformation tasks, but the scarcity of directly annotated data for either task makes them unfeasible for most languages. We present an approach that does both tasks within the same trained model, and only uses regular language parallel data, without requiring error-corrected or style-adapted texts. We apply our model to three languages and present a thorough evaluation on both tasks, showing that the model is reliable for a number of error types and style transfer aspects.

Elizaveta Korotkova, Maksym Del, Mark Fishel

We introduce the task of zero-shot style transfer between different languages. Our training data includes multilingual parallel corpora, but does not contain any parallel sentences between styles, similarly to the recent previous work. We propose a unified multilingual multi-style machine translation system design, that allows to perform zero-shot style conversions during inference; moreover, it does so both monolingually and cross-lingually. Our model allows to increase the presence of dissimilar styles in corpus by up to 3 times, easily learns to operate with various contractions, and provides reasonable lexicon swaps as we see from manual evaluation.

Hasan Sait Arslan, Mark Fishel, Gholamreza Anbarjafari

In this paper a doubly attentive transformer machine translation model (DATNMT) is presented in which a doubly-attentive transformer decoder normally joins spatial visual features obtained via pretrained convolutional neural networks, conquering any gap between image captioning and translation. In this framework, the transformer decoder figures out how to take care of source-language words and parts of an image freely by methods for two separate attention components in an Enhanced Multi-Head Attention Layer of doubly attentive transformer, as it generates words in the target language. We find that the proposed model can effectively exploit not just the scarce multimodal machine translation data, but also large general-domain text-only machine translation corpora, or image-text image captioning corpora. The experimental results show that the proposed doubly-attentive transformer-decoder performs better than a single-decoder transformer model, and gives the state-of-the-art results in the English-German multimodal machine translation task.

Sander Tars, Mark Fishel

We present an approach to neural machine translation (NMT) that supports multiple domains in a single model and allows switching between the domains when translating. The core idea is to treat text domains as distinct languages and use multilingual NMT methods to create multi-domain translation systems, we show that this approach results in significant translation quality gains over fine-tuning. We also explore whether the knowledge of pre-specified text domains is necessary, turns out that it is after all, but also that when it is not known quite high translation quality can be reached.

Matīss Rikters, Mark Fishel

Attention distributions of the generated translations are a useful bi-product of attention-based recurrent neural network translation models and can be treated as soft alignments between the input and output tokens. In this work, we use attention distributions as a confidence metric for output translations. We present two strategies of using the attention distributions: filtering out bad translations from a large back-translated corpus, and selecting the best translation in a hybrid setup of two different translation systems. While manual evaluation indicated only a weak correlation between our confidence score and human judgments, the use-cases showed improvements of up to 2.22 BLEU points for filtering and 0.99 points for hybrid translation, tested on English<->German and English<->Latvian translation.