NLP in the age of monolithic large language models is approaching its limits in terms of size and information that can be handled. The trend goes to modularization, a necessary step into the direction of designing smaller sub-networks and components with specialized functionality. In this paper, we present the MAMMOTH toolkit: a framework designed for training massively multilingual modular machine translation systems at scale, initially derived from OpenNMT-py and then adapted to ensure efficient training across computation clusters.We showcase its efficiency across clusters of A100 and V100 NVIDIA GPUs, and discuss our design philosophy and plans for future information.The toolkit is publicly available online at https://github.com/Helsinki-NLP/mammoth.

The fine-tuning paradigm has been widely adopted to train neural models tailored for specific tasks. However, the recent upsurge of Large Language Models (LLMs), characterized by billions of parameters, has introduced profound computational challenges to the fine-tuning process. This has fueled intensive research on Parameter-Efficient Fine-Tuning (PEFT) techniques, usually involving the training of a selective subset of the original model parameters. One of the most used approaches is Adapters, which add trainable lightweight layers to the existing pretrained weights. Within this context, we propose AdaKron, an Adapter-based fine-tuning with the Kronecker product. In particular, we leverage the Kronecker product to combine the output of two small networks, resulting in a final vector whose dimension is the product of the dimensions of the individual outputs, allowing us to train only 0.55% of the model’s original parameters. We evaluate AdaKron performing a series of experiments on the General Language Understanding Evaluation (GLUE) benchmark, achieving results in the same ballpark as recent state-of-the-art PEFT methods, despite training fewer parameters.

This paper presents the results of the SHROOM, a shared task focused on detecting hallucinations: outputs from natural language generation (NLG) systems that are fluent, yet inaccurate. Such cases of overgeneration put in jeopardy many NLG applications, where correctness is often mission-critical. The shared task was conducted with a newly constructed dataset of 4000 model outputs labeled by 5 annotators each, spanning 3 NLP tasks: machine translation, paraphrase generation and definition modeling.The shared task was tackled by a total of 58 different users grouped in 42 teams, out of which 26 elected to write a system description paper; collectively, they submitted over 300 prediction sets on both tracks of the shared task. We observe a number of key trends in how this approach was tackled—many participants rely on a handful of model, and often rely either on synthetic data for fine-tuning or zero-shot prompting strategies. While a majority of the teams did outperform our proposed baseline system, the performances of top-scoring systems are still consistent with a random handling of the more challenging items.

This paper presents the Visual Word Sense Disambiguation (Visual-WSD) task. The objective of Visual-WSD is to identify among a set of ten images the one that corresponds to the intended meaning of a given ambiguous word which is accompanied with minimal context. The task provides datasets for three different languages: English, Italian, and Farsi.We received a total of 96 different submissions. Out of these, 40 systems outperformed a strong zero-shot CLIP-based baseline. Participating systems proposed different zero- and few-shot approaches, often involving generative models and data augmentation. More information can be found on the task’s website: \url{https://raganato.github.io/vwsd/}.

There are several ways of implementing multilingual NLP systems but little consensus as to whether different approaches exhibit similar effects. Are the trends that we observe when adding more languages the same as those we observe when sharing more parameters? We focus on encoder representations drawn from modular multilingual machine translation systems in an English-centric scenario, and study their quality from multiple aspects: how adequate they are for machine translation, how independent of the source language they are, and what semantic information they convey. Adding translation directions in English-centric scenarios does not conclusively lead to an increase in translation quality. Shared layers increase performance on zero-shot translation pairs and lead to more language-independent representations, but these improvements do not systematically align with more semantically accurate representations, from a monolingual standpoint.

We describe the enhancement of a multilingual NMT toolkit developed as part of the FoTran project. We devise our modular attention-bridge model, which connects language-specific components through a shared network layer. The system now supports distributed training over many nodes and GPUs in order to substantially scale up the number of languages that can be included in a modern neural translation architecture. The model enables the study of emerging language-agnostic representations and also provides a modular toolkit for efficient machine translation.

Masked language models have quickly become the de facto standard when processing text. Recently, several approaches have been proposed to further enrich word representations with external knowledge sources such as knowledge graphs. However, these models are devised and evaluated in a monolingual setting only. In this work, we propose a language-independent entity prediction task as an intermediate training procedure to ground word representations on entity semantics and bridge the gap across different languages by means of a shared vocabulary of entities. We show that our approach effectively injects new lexical-semantic knowledge into neural models, improving their performance on different semantic tasks in the zero-shot crosslingual setting. As an additional advantage, our intermediate training does not require any supplementary input, allowing our models to be applied to new datasets right away. In our experiments, we use Wikipedia articles in up to 100 languages and already observe consistent gains compared to strong baselines when predicting entities using only the English Wikipedia. Further adding extra languages lead to improvements in most tasks up to a certain point, but overall we found it non-trivial to scale improvements in model transferability by training on ever increasing amounts of Wikipedia languages.

Zero-shot translations is a fascinating feature of Multilingual Neural Machine Translation (MNMT) systems. These MNMT models are usually trained on English-centric data, i.e. English either as the source or target language, and with a language label prepended to the input indicating the target language. However, recent work has highlighted several flaws of these models in zero-shot scenarios where language labels are ignored and the wrong language is generated or different runs show highly unstable results. In this paper, we investigate the benefits of an explicit alignment to language labels in Transformer-based MNMT models in the zero-shot context, by jointly training one cross attention head with word alignment supervision to stress the focus on the target language label. We compare and evaluate several MNMT systems on three multilingual MT benchmarks of different sizes, showing that simply supervising one cross attention head to focus both on word alignments and language labels reduces the bias towards translating into the wrong language, improving the zero-shot performance overall. Moreover, as an additional advantage, we find that our alignment supervision leads to more stable results across different training runs.

This paper reports on our participation with the MUCOW test suite at the WMT 2020 news translation task. We introduced MUCOW at WMT 2019 to measure the ability of MT systems to perform word sense disambiguation (WSD), i.e., to translate an ambiguous word with its correct sense. MUCOW is created automatically using existing resources, and the evaluation process is also entirely automated. We evaluate all participating systems of the language pairs English -> Czech, English -> German, and English -> Russian and compare the results with those obtained at WMT 2019. While current NMT systems are fairly good at handling ambiguous source words, we could not identify any substantial progress - at least to the extent that it is measurable by the MUCOW method - in that area over the last year.

Lexical ambiguity is one of the many challenging linguistic phenomena involved in translation, i.e., translating an ambiguous word with its correct sense. In this respect, previous work has shown that the translation quality of neural machine translation systems can be improved by explicitly modeling the senses of ambiguous words. Recently, several evaluation test sets have been proposed to measure the word sense disambiguation (WSD) capability of machine translation systems. However, to date, these evaluation test sets do not include any training data that would provide a fair setup measuring the sense distributions present within the training data itself. In this paper, we present an evaluation benchmark on WSD for machine translation for 10 language pairs, comprising training data with known sense distributions. Our approach for the construction of the benchmark builds upon the wide-coverage multilingual sense inventory of BabelNet, the multilingual neural parsing pipeline TurkuNLP, and the OPUS collection of translated texts from the web. The test suite is available at http://github.com/Helsinki-NLP/MuCoW.

Neural machine translation has considerably improved the quality of automatic translations by learning good representations of input sentences. In this article, we explore a multilingual translation model capable of producing fixed-size sentence representations by incorporating an intermediate crosslingual shared layer, which we refer to as attention bridge. This layer exploits the semantics from each language and develops into a language-agnostic meaning representation that can be efficiently used for transfer learning. We systematically study the impact of the size of the attention bridge and the effect of including additional languages in the model. In contrast to related previous work, we demonstrate that there is no conflict between translation performance and the use of sentence representations in downstream tasks. In particular, we show that larger intermediate layers not only improve translation quality, especially for long sentences, but also push the accuracy of trainable classification tasks. Nevertheless, shorter representations lead to increased compression that is beneficial in non-trainable similarity tasks. Similarly, we show that trainable downstream tasks benefit from multilingual models, whereas additional language signals do not improve performance in non-trainable benchmarks. This is an important insight that helps to properly design models for specific applications. Finally, we also include an in-depth analysis of the proposed attention bridge and its ability to encode linguistic properties. We carefully analyze the information that is captured by individual attention heads and identify interesting patterns that explain the performance of specific settings in linguistic probing tasks.

Transformer-based models have brought a radical change to neural machine translation. A key feature of the Transformer architecture is the so-called multi-head attention mechanism, which allows the model to focus simultaneously on different parts of the input. However, recent works have shown that most attention heads learn simple, and often redundant, positional patterns. In this paper, we propose to replace all but one attention head of each encoder layer with simple fixed – non-learnable – attentive patterns that are solely based on position and do not require any external knowledge. Our experiments with different data sizes and multiple language pairs show that fixing the attention heads on the encoder side of the Transformer at training time does not impact the translation quality and even increases BLEU scores by up to 3 points in low-resource scenarios.

The ability to correctly model distinct meanings of a word is crucial for the effectiveness of semantic representation techniques. However, most existing evaluation benchmarks for assessing this criterion are tied to sense inventories (usually WordNet), restricting their usage to a small subset of knowledge-based representation techniques. The Word-in-Context dataset (WiC) addresses the dependence on sense inventories by reformulating the standard disambiguation task as a binary classification problem; but, it is limited to the English language. We put forward a large multilingual benchmark, XL-WiC, featuring gold standards in 12 new languages from varied language families and with different degrees of resource availability, opening room for evaluation scenarios such as zero-shot cross-lingual transfer. We perform a series of experiments to determine the reliability of the datasets and to set performance baselines for several recent contextualized multilingual models. Experimental results show that even when no tagged instances are available for a target language, models trained solely on the English data can attain competitive performance in the task of distinguishing different meanings of a word, even for distant languages. XL-WiC is available at https://pilehvar.github.io/xlwic/.

In this paper, we explore a multilingual translation model with a cross-lingually shared layer that can be used as fixed-size sentence representation in different downstream tasks. We systematically study the impact of the size of the shared layer and the effect of including additional languages in the model. In contrast to related previous work, we demonstrate that the performance in translation does correlate with trainable downstream tasks. In particular, we show that larger intermediate layers not only improve translation quality, especially for long sentences, but also push the accuracy of trainable classification tasks. On the other hand, shorter representations lead to increased compression that is beneficial in non-trainable similarity tasks. We hypothesize that the training procedure on the downstream task enables the model to identify the encoded information that is useful for the specific task whereas non-trainable benchmarks can be confused by other types of information also encoded in the representation of a sentence.

In this paper, we propose an architecture for machine translation (MT) capable of obtaining multilingual sentence representations by incorporating an intermediate attention bridge that is shared across all languages. We train the model with language-specific encoders and decoders that are connected through an inner-attention layer on the encoder side. The attention bridge exploits the semantics from each language for translation and develops into a language-agnostic meaning representation that can efficiently be used for transfer learning. We present a new framework for the efficient development of multilingual neural machine translation (NMT) using this model and scheduled training. We have tested the approach in a systematic way with a multi-parallel data set. The model achieves substantial improvements over strong bilingual models and performs well for zero-shot translation, which demonstrates its ability of abstraction and transfer learning.

In this paper we present the University of Helsinki submissions to the WMT 2019 shared news translation task in three language pairs: English-German, English-Finnish and Finnish-English. This year we focused first on cleaning and filtering the training data using multiple data-filtering approaches, resulting in much smaller and cleaner training sets. For English-German we trained both sentence-level transformer models as well as compared different document-level translation approaches. For Finnish-English and English-Finnish we focused on different segmentation approaches and we also included a rule-based system for English-Finnish.

Supervised Neural Machine Translation (NMT) systems currently achieve impressive translation quality for many language pairs. One of the key features of a correct translation is the ability to perform word sense disambiguation (WSD), i.e., to translate an ambiguous word with its correct sense. Existing evaluation benchmarks on WSD capabilities of translation systems rely heavily on manual work and cover only few language pairs and few word types. We present MuCoW, a multilingual contrastive test suite that covers 16 language pairs with more than 200 thousand contrastive sentence pairs, automatically built from word-aligned parallel corpora and the wide-coverage multilingual sense inventory of BabelNet. We evaluate the quality of the ambiguity lexicons and of the resulting test suite on all submissions from 9 language pairs presented in the WMT19 news shared translation task, plus on other 5 language pairs using NMT pretrained models. The MuCoW test suite is available at http://github.com/Helsinki-NLP/MuCoW.

The attention mechanism is a successful technique in modern NLP, especially in tasks like machine translation. The recently proposed network architecture of the Transformer is based entirely on attention mechanisms and achieves new state of the art results in neural machine translation, outperforming other sequence-to-sequence models. However, so far not much is known about the internal properties of the model and the representations it learns to achieve that performance. To study this question, we investigate the information that is learned by the attention mechanism in Transformer models with different translation quality. We assess the representations of the encoder by extracting dependency relations based on self-attention weights, we perform four probing tasks to study the amount of syntactic and semantic captured information and we also test attention in a transfer learning scenario. Our analysis sheds light on the relative strengths and weaknesses of the various encoder representations. We observe that specific attention heads mark syntactic dependency relations and we can also confirm that lower layers tend to learn more about syntax while higher layers tend to encode more semantics.

This paper describes the University of Helsinki’s submissions to the WMT18 shared news translation task for English-Finnish and English-Estonian, in both directions. This year, our main submissions employ a novel neural architecture, the Transformer, using the open-source OpenNMT framework. Our experiments couple domain labeling and fine tuned multilingual models with shared vocabularies between the source and target language, using the provided parallel data of the shared task and additional back-translations. Finally, we compare, for the English-to-Finnish case, the effectiveness of different machine translation architectures, starting from a rule-based approach to our best neural model, analyzing the output and highlighting future research.

This paper describes Sew-Embed, our language-independent approach to multilingual and cross-lingual semantic word similarity as part of the SemEval-2017 Task 2. We leverage the Wikipedia-based concept representations developed by Raganato et al. (2016), and propose an embedded augmentation of their explicit high-dimensional vectors, which we obtain by plugging in an arbitrary word (or sense) embedding representation, and computing a weighted average in the continuous vector space. We evaluate Sew-Embed with two different off-the-shelf embedding representations, and report their performances across all monolingual and cross-lingual benchmarks available for the task. Despite its simplicity, especially compared with supervised or overly tuned approaches, Sew-Embed achieves competitive results in the cross-lingual setting (3rd best result in the global ranking of subtask 2, score 0.56).

Word Sense Disambiguation models exist in many flavors. Even though supervised ones tend to perform best in terms of accuracy, they often lose ground to more flexible knowledge-based solutions, which do not require training by a word expert for every disambiguation target. To bridge this gap we adopt a different perspective and rely on sequence learning to frame the disambiguation problem: we propose and study in depth a series of end-to-end neural architectures directly tailored to the task, from bidirectional Long Short-Term Memory to encoder-decoder models. Our extensive evaluation over standard benchmarks and in multiple languages shows that sequence learning enables more versatile all-words models that consistently lead to state-of-the-art results, even against word experts with engineered features.

In this demonstration we present SupWSD, a Java API for supervised Word Sense Disambiguation (WSD). This toolkit includes the implementation of a state-of-the-art supervised WSD system, together with a Natural Language Processing pipeline for preprocessing and feature extraction. Our aim is to provide an easy-to-use tool for the research community, designed to be modular, fast and scalable for training and testing on large datasets. The source code of SupWSD is available at http://github.com/SI3P/SupWSD.

Word Sense Disambiguation is a long-standing task in Natural Language Processing, lying at the core of human language understanding. However, the evaluation of automatic systems has been problematic, mainly due to the lack of a reliable evaluation framework. In this paper we develop a unified evaluation framework and analyze the performance of various Word Sense Disambiguation systems in a fair setup. The results show that supervised systems clearly outperform knowledge-based models. Among the supervised systems, a linear classifier trained on conventional local features still proves to be a hard baseline to beat. Nonetheless, recent approaches exploiting neural networks on unlabeled corpora achieve promising results, surpassing this hard baseline in most test sets.

Parallel corpora are widely used in a variety of Natural Language Processing tasks, from Machine Translation to cross-lingual Word Sense Disambiguation, where parallel sentences can be exploited to automatically generate high-quality sense annotations on a large scale. In this paper we present EuroSense, a multilingual sense-annotated resource based on the joint disambiguation of the Europarl parallel corpus, with almost 123 million sense annotations for over 155 thousand distinct concepts and entities from a language-independent unified sense inventory. We evaluate the quality of our sense annotations intrinsically and extrinsically, showing their effectiveness as training data for Word Sense Disambiguation.

The increasing amount of multilingual text collections available in different domains makes its automatic processing essential for the development of a given field. However, standard processing techniques based on statistical clues and keyword searches have clear limitations. Instead, we propose a knowledge-based processing pipeline which overcomes most of the limitations of these techniques. This, in turn, enables direct comparison across texts in different languages without the need of translation. In this paper we show the potential of this approach for semantically indexing multilingual text collections in the history domain. In our experiments we used a version of the Bible translated in four different languages, evaluating the precision of our semantic indexing pipeline and showing its reliability on the cross-lingual text retrieval task.

Linking concepts and named entities to knowledge bases has become a crucial Natural Language Understanding task. In this respect, recent works have shown the key advantage of exploiting textual definitions in various Natural Language Processing applications. However, to date there are no reliable large-scale corpora of sense-annotated textual definitions available to the research community. In this paper we present a large-scale high-quality corpus of disambiguated glosses in multiple languages, comprising sense annotations of both concepts and named entities from a unified sense inventory. Our approach for the construction and disambiguation of the corpus builds upon the structure of a large multilingual semantic network and a state-of-the-art disambiguation system; first, we gather complementary information of equivalent definitions across different languages to provide context for disambiguation, and then we combine it with a semantic similarity-based refinement. As a result we obtain a multilingual corpus of textual definitions featuring over 38 million definitions in 263 languages, and we make it freely available at http://lcl.uniroma1.it/disambiguated-glosses. Experiments on Open Information Extraction and Sense Clustering show how two state-of-the-art approaches improve their performance by integrating our disambiguated corpus into their pipeline.

Entity Linking (EL) and Word Sense Disambiguation (WSD) both address the lexical ambiguity of language. But while the two tasks are pretty similar, they differ in a fundamental respect: in EL the textual mention can be linked to a named entity which may or may not contain the exact mention, while in WSD there is a perfect match between the word form (better, its lemma) and a suitable word sense. In this paper we present Babelfy, a unified graph-based approach to EL and WSD based on a loose identification of candidate meanings coupled with a densest subgraph heuristic which selects high-coherence semantic interpretations. Our experiments show state-of-the-art performances on both tasks on 6 different datasets, including a multilingual setting. Babelfy is online at http://babelfy.org