Few-shot learning aims to learn representations that can tackle novel tasks given a small number of examples. Recent studies show that cross-modal learning can improve representations for few-shot classification. More specifically, language is a rich modality that can be used to guide visual learning. In this work, we experiment with a multi-modal architecture for few-shot learning that consists of three components: a classifier, an auxiliary network, and a bridge network. While the classifier performs the main classification task, the auxiliary network learns to predict language representations from the same input, and the bridge network transforms high-level features of the auxiliary network into modulation parameters for layers of the few-shot classifier using conditional batch normalization. The bridge should encourage a form of lightweight semantic alignment between language and vision which could be useful for the classifier. However, after evaluating the proposed approach on two popular few-shot classification benchmarks we find that a) the improvements do not reproduce across benchmarks, and b) when they do, the improvements are due to the additional compute and parameters introduced by the bridge network. We contribute insights and recommendations for future work in multi-modal meta-learning, especially when using language representations.

Current large language model (LLM) evaluations rely on benchmarks to assess model capabilities and their encoded knowledge. However, these evaluations cannot reveal where a model encodes its knowledge, and thus little is known about which weights contain specific information. We propose a method to statically (without forward or backward passes) locate topical knowledge in the weight space of an LLM, building on a prior insight that parameters can be decoded into interpretable tokens. If parameters can be mapped into the embedding space, it should be possible to directly search for knowledge via embedding similarity. We study the validity of this assumption across several LLMs for a variety of concepts in the financial domain and a toxicity detection setup. Our analysis yields an improved understanding of the promises and limitations of static knowledge location in real-world scenarios.

This work presents the first large-scale biomedical Spanish language models trained from scratch, using large biomedical corpora consisting of a total of 1.1B tokens and an EHR corpus of 95M tokens. We compared them against general-domain and other domain-specific models for Spanish on three clinical NER tasks. As main results, our models are superior across the NER tasks, rendering them more convenient for clinical NLP applications. Furthermore, our findings indicate that when enough data is available, pre-training from scratch is better than continual pre-training when tested on clinical tasks, raising an exciting research question about which approach is optimal. Our models and fine-tuning scripts are publicly available at HuggingFace and GitHub.

In this work, we make the case of quality over quantity when training a MT system for a medium-to-low-resource language pair, namely Catalan-English. We compile our training corpus out of existing resources of varying quality and a new high-quality corpus. We also provide new evaluation translation datasets in three different domains. In the process of building Catalan-English parallel resources, we evaluate the impact of drastically filtering alignments in the resulting MT engines. Our results show that even when resources are limited, as in this case, it is worth filtering for quality. We further explore the cross-lingual transfer learning capabilities of the proposed model for parallel corpus filtering by applying it to other languages. All resources generated in this work are released under open license to encourage the development of language technology in Catalan.

In this work, we present the work that has been carried on in the MT4All CEF project and the resources that it has generated by leveraging recent research carried out in the field of unsupervised learning. In the course of the project 18 monolingual corpora for specific domains and languages have been collected, and 12 bilingual dictionaries and translation models have been generated. As part of the research, the unsupervised MT methodology based only on monolingual corpora (Artetxe et al., 2017) has been tested on a variety of languages and domains. Results show that in specialised domains, when there is enough monolingual in-domain data, unsupervised results are comparable to those of general domain supervised translation, and that, at any rate, unsupervised techniques can be used to boost results whenever very little data is available.

Generative Pre-trained Transformers (GPTs) have recently been scaled to unprecedented sizes in the history of machine learning. These models, solely trained on the language modeling objective, have been shown to exhibit outstanding zero, one, and few-shot learning capabilities in a number of different tasks. Nevertheless, aside from anecdotal experiences, little is known regarding their multilingual capabilities, given the fact that the pre-training corpus is almost entirely composed of English text. In this work, we investigate its potential and limits in three tasks: extractive question-answering, text summarization and natural language generation for five different languages, as well as the effect of scale in terms of model size. Our results show that GPT-3 can be almost as useful for many languages as it is for English, with room for improvement if optimization of the tokenization is addressed.

Introducing factors, that is to say, word features such as linguistic information referring to the source tokens, is known to improve the results of neural machine translation systems in certain settings, typically in recurrent architectures. This study proposes enhancing the current state-of-the-art neural machine translation architecture, the Transformer, so that it allows to introduce external knowledge. In particular, our proposed modification, the Factored Transformer, uses linguistic factors that insert additional knowledge into the machine translation system. Apart from using different kinds of features, we study the effect of different architectural configurations. Specifically, we analyze the performance of combining words and features at the embedding level or at the encoder level, and we experiment with two different combination strategies. With the best-found configuration, we show improvements of 0.8 BLEU over the baseline Transformer in the IWSLT German-to-English task. Moreover, we experiment with the more challenging FLoRes English-to-Nepali benchmark, which includes both extremely low-resourced and very distant languages, and obtain an improvement of 1.2 BLEU

This paper describes the participation of the BSC team in the WMT2021’s Multilingual Low-Resource Translation for Indo-European Languages Shared Task. The system aims to solve the Subtask 2: Wikipedia cultural heritage articles, which involves translation in four Romance languages: Catalan, Italian, Occitan and Romanian. The submitted system is a multilingual semi-supervised machine translation model. It is based on a pre-trained language model, namely XLM-RoBERTa, that is later fine-tuned with parallel data obtained mostly from OPUS. Unlike other works, we only use XLM to initialize the encoder and randomly initialize a shallow decoder. The reported results are robust and perform well for all tested languages.

Word embeddings are representations of words in a dense vector space. Although they are not recent phenomena in Natural Language Processing (NLP), they have gained momentum after the recent developments of neural methods and Word2Vec. Regarding their applications in medical and clinical NLP, they are invaluable resources when training in-domain named entity recognition systems, classifiers or taggers, for instance. Thus, the development of tailored word embeddings for medical NLP is of great interest. However, we identified a gap in the literature which we aim to fill in this paper: the availability of embeddings for medical NLP in Spanish, as well as a standardized form of intrinsic evaluation. Since most work has been done for English, some established datasets for intrinsic evaluation are already available. In this paper, we show the steps we employed to adapt such datasets for the first time to Spanish, of particular relevance due to the considerable volume of EHRs in this language, as well as the creation of in-domain medical word embeddings for the Spanish using the state-of-the-art FastText model. We performed intrinsic evaluation with our adapted datasets, as well as extrinsic evaluation with a named entity recognition systems using a baseline embedding of general-domain. Both experiments proved that our embeddings are suitable for use in medical NLP in the Spanish language, and are more accurate than general-domain ones.