Abstract:Pixel-based language models have emerged as a compelling alternative to subword-based language modelling, particularly because they can represent virtually any script. PIXEL, a canonical example of such a model, is a vision transformer that has been pre-trained on rendered text. While PIXEL has shown promising cross-script transfer abilities and robustness to orthographic perturbations, it falls short of outperforming monolingual subword counterparts like BERT in most other contexts. This discrepancy raises questions about the amount of linguistic knowledge learnt by these models and whether their performance in language tasks stems more from their visual capabilities than their linguistic ones. To explore this, we probe PIXEL using a variety of linguistic and visual tasks to assess its position on the vision-to-language spectrum. Our findings reveal a substantial gap between the model's visual and linguistic understanding. The lower layers of PIXEL predominantly capture superficial visual features, whereas the higher layers gradually learn more syntactic and semantic abstractions. Additionally, we examine variants of PIXEL trained with different text rendering strategies, discovering that introducing certain orthographic constraints at the input level can facilitate earlier learning of surface-level features. With this study, we hope to provide insights that aid the further development of pixel-based language models.
Abstract:Parameter-efficient fine-tuning (PEFT) methods are increasingly used with pre-trained language models (PLMs) for continual learning (CL). These methods involve training a PEFT module for each new task and using similarity-based selection to route modules during inference. However, they face two major limitations: 1) interference with already learned modules and 2) suboptimal routing when composing modules. In this paper, we introduce a method that isolates the training of PEFT modules for task specialization. Then, before evaluation, it learns to compose the previously learned modules by training a router that leverages samples from a small memory. We evaluate our method in two CL setups using several benchmarks. Our results show that our method provides a better composition of PEFT modules, leading to better generalization and performance compared to previous methods.
Abstract:Advancing representation learning in specialized fields like medicine remains challenging due to the scarcity of expert annotations for text and images. To tackle this issue, we present a novel two-stage framework designed to extract high-quality factual statements from free-text radiology reports in order to improve the representations of text encoders and, consequently, their performance on various downstream tasks. In the first stage, we propose a \textit{Fact Extractor} that leverages large language models (LLMs) to identify factual statements from well-curated domain-specific datasets. In the second stage, we introduce a \textit{Fact Encoder} (CXRFE) based on a BERT model fine-tuned with objective functions designed to improve its representations using the extracted factual data. Our framework also includes a new embedding-based metric (CXRFEScore) for evaluating chest X-ray text generation systems, leveraging both stages of our approach. Extensive evaluations show that our fact extractor and encoder outperform current state-of-the-art methods in tasks such as sentence ranking, natural language inference, and label extraction from radiology reports. Additionally, our metric proves to be more robust and effective than existing metrics commonly used in the radiology report generation literature. The code of this project is available at \url{https://github.com/PabloMessina/CXR-Fact-Encoder}.
Abstract:Visual Question Answering (VQA) is an important task in multimodal AI, and it is often used to test the ability of vision-language models to understand and reason on knowledge present in both visual and textual data. However, most of the current VQA models use datasets that are primarily focused on English and a few major world languages, with images that are typically Western-centric. While recent efforts have tried to increase the number of languages covered on VQA datasets, they still lack diversity in low-resource languages. More importantly, although these datasets often extend their linguistic range via translation or some other approaches, they usually keep images the same, resulting in narrow cultural representation. To address these limitations, we construct CVQA, a new Culturally-diverse multilingual Visual Question Answering benchmark, designed to cover a rich set of languages and cultures, where we engage native speakers and cultural experts in the data collection process. As a result, CVQA includes culturally-driven images and questions from across 28 countries on four continents, covering 26 languages with 11 scripts, providing a total of 9k questions. We then benchmark several Multimodal Large Language Models (MLLMs) on CVQA, and show that the dataset is challenging for the current state-of-the-art models. This benchmark can serve as a probing evaluation suite for assessing the cultural capability and bias of multimodal models and hopefully encourage more research efforts toward increasing cultural awareness and linguistic diversity in this field.
Abstract:We present the first shared task on Semantic Textual Relatedness (STR). While earlier shared tasks primarily focused on semantic similarity, we instead investigate the broader phenomenon of semantic relatedness across 14 languages: Afrikaans, Algerian Arabic, Amharic, English, Hausa, Hindi, Indonesian, Kinyarwanda, Marathi, Moroccan Arabic, Modern Standard Arabic, Punjabi, Spanish, and Telugu. These languages originate from five distinct language families and are predominantly spoken in Africa and Asia -- regions characterised by the relatively limited availability of NLP resources. Each instance in the datasets is a sentence pair associated with a score that represents the degree of semantic textual relatedness between the two sentences. Participating systems were asked to rank sentence pairs by their closeness in meaning (i.e., their degree of semantic relatedness) in the 14 languages in three main tracks: (a) supervised, (b) unsupervised, and (c) crosslingual. The task attracted 163 participants. We received 70 submissions in total (across all tasks) from 51 different teams, and 38 system description papers. We report on the best-performing systems as well as the most common and the most effective approaches for the three different tracks.
Abstract:Exploring and quantifying semantic relatedness is central to representing language. It holds significant implications across various NLP tasks, including offering insights into the capabilities and performance of Large Language Models (LLMs). While earlier NLP research primarily focused on semantic similarity, often within the English language context, we instead investigate the broader phenomenon of semantic relatedness. In this paper, we present SemRel, a new semantic relatedness dataset collection annotated by native speakers across 14 languages:Afrikaans, Algerian Arabic, Amharic, English, Hausa, Hindi, Indonesian, Kinyarwanda, Marathi, Moroccan Arabic, Modern Standard Arabic, Punjabi, Spanish, and Telugu. These languages originate from five distinct language families and are predominantly spoken in Africa and Asia -- regions characterised by a relatively limited availability of NLP resources. Each instance in the SemRel datasets is a sentence pair associated with a score that represents the degree of semantic textual relatedness between the two sentences. The scores are obtained using a comparative annotation framework. We describe the data collection and annotation processes, related challenges when building the datasets, and their impact and utility in NLP. We further report experiments for each language and across the different languages.
Abstract:In recent years, substantial advancements in pre-trained language models have paved the way for the development of numerous non-English language versions, with a particular focus on encoder-only and decoder-only architectures. While Spanish language models encompassing BERT, RoBERTa, and GPT have exhibited prowess in natural language understanding and generation, there remains a scarcity of encoder-decoder models designed for sequence-to-sequence tasks involving input-output pairs. This paper breaks new ground by introducing the implementation and evaluation of renowned encoder-decoder architectures, exclusively pre-trained on Spanish corpora. Specifically, we present Spanish versions of BART, T5, and BERT2BERT-style models and subject them to a comprehensive assessment across a diverse range of sequence-to-sequence tasks, spanning summarization, rephrasing, and generative question answering. Our findings underscore the competitive performance of all models, with BART and T5 emerging as top performers across all evaluated tasks. As an additional contribution, we have made all models publicly available to the research community, fostering future exploration and development in Spanish language processing.
Abstract:Existing question answering methods often assume that the input content (e.g., documents or videos) is always accessible to solve the task. Alternatively, memory networks were introduced to mimic the human process of incremental comprehension and compression of the information in a fixed-capacity memory. However, these models only learn how to maintain memory by backpropagating errors in the answers through the entire network. Instead, it has been suggested that humans have effective mechanisms to boost their memorization capacities, such as rehearsal and anticipation. Drawing inspiration from these, we propose a memory model that performs rehearsal and anticipation while processing inputs to memorize important information for solving question answering tasks from streaming data. The proposed mechanisms are applied self-supervised during training through masked modeling tasks focused on coreference information. We validate our model on a short-sequence (bAbI) dataset as well as large-sequence textual (NarrativeQA) and video (ActivityNet-QA) question answering datasets, where it achieves substantial improvements over previous memory network approaches. Furthermore, our ablation study confirms the proposed mechanisms' importance for memory models.
Abstract:Lifelong language learning seeks to have models continuously learn multiple tasks in a sequential order without suffering from catastrophic forgetting. State-of-the-art approaches rely on sparse experience replay as the primary approach to prevent forgetting. Experience replay usually adopts sampling methods for the memory population; however, the effect of the chosen sampling strategy on model performance has not yet been studied. In this paper, we investigate how relevant the selective memory population is in the lifelong learning process of text classification and question-answering tasks. We found that methods that randomly store a uniform number of samples from the entire data stream lead to high performances, especially for low memory size, which is consistent with computer vision studies.
Abstract:Continual Learning methods strive to mitigate Catastrophic Forgetting (CF), where knowledge from previously learned tasks is lost when learning a new one. Among those algorithms, some maintain a subset of samples from previous tasks when training. These samples are referred to as a memory. These methods have shown outstanding performance while being conceptually simple and easy to implement. Yet, despite their popularity, little has been done to understand which elements to be included into the memory. Currently, this memory is often filled via random sampling with no guiding principles that may aid in retaining previous knowledge. In this work, we propose a criterion based on the learning consistency of a sample called Consistency AWare Sampling (CAWS). This criterion prioritizes samples that are easier to learn by deep networks. We perform studies on three different memory-based methods: AGEM, GDumb, and Experience Replay, on MNIST, CIFAR-10 and CIFAR-100 datasets. We show that using the most consistent elements yields performance gains when constrained by a compute budget; when under no such constrain, random sampling is a strong baseline. However, using CAWS on Experience Replay yields improved performance over the random baseline. Finally, we show that CAWS achieves similar results to a popular memory selection method while requiring significantly less computational resources.