Albert Gatt

Mika Sie, Ruby Beek, Michiel Bots et al.

Due to their length and complexity, long regulatory texts are challenging to summarize. To address this, a multi-step extractive-abstractive architecture is proposed to handle lengthy regulatory documents more effectively. In this paper, we show that the effectiveness of a two-step architecture for summarizing long regulatory texts varies significantly depending on the model used. Specifically, the two-step architecture improves the performance of decoder-only models. For abstractive encoder-decoder models with short context lengths, the effectiveness of an extractive step varies, whereas for long-context encoder-decoder models, the extractive step worsens their performance. This research also highlights the challenges of evaluating generated texts, as evidenced by the differing results from human and automated evaluations. Most notably, human evaluations favoured language models pretrained on legal text, while automated metrics rank general-purpose language models higher. The results underscore the importance of selecting the appropriate summarization strategy based on model architecture and context length.

Ilker Kesen, Andrea Pedrotti, Mustafa Dogan et al.

With the ever-increasing popularity of pretrained Video-Language Models (VidLMs), there is a pressing need to develop robust evaluation methodologies that delve deeper into their visio-linguistic capabilities. To address this challenge, we present ViLMA (Video Language Model Assessment), a task-agnostic benchmark that places the assessment of fine-grained capabilities of these models on a firm footing. Task-based evaluations, while valuable, fail to capture the complexities and specific temporal aspects of moving images that VidLMs need to process. Through carefully curated counterfactuals, ViLMA offers a controlled evaluation suite that sheds light on the true potential of these models, as well as their performance gaps compared to human-level understanding. ViLMA also includes proficiency tests, which assess basic capabilities deemed essential to solving the main counterfactual tests. We show that current VidLMs' grounding abilities are no better than those of vision-language models which use static images. This is especially striking once the performance on proficiency tests is factored in. Our benchmark serves as a catalyst for future research on VidLMs, helping to highlight areas that still need to be explored.

Kurt Micallef, Albert Gatt, Marc Tanti et al.

Multilingual language models such as mBERT have seen impressive cross-lingual transfer to a variety of languages, but many languages remain excluded from these models. In this paper, we analyse the effect of pre-training with monolingual data for a low-resource language that is not included in mBERT -- Maltese -- with a range of pre-training set ups. We conduct evaluations with the newly pre-trained models on three morphosyntactic tasks -- dependency parsing, part-of-speech tagging, and named-entity recognition -- and one semantic classification task -- sentiment analysis. We also present a newly created corpus for Maltese, and determine the effect that the pre-training data size and domain have on the downstream performance. Our results show that using a mixture of pre-training domains is often superior to using Wikipedia text only. We also find that a fraction of this corpus is enough to make significant leaps in performance over Wikipedia-trained models. We pre-train and compare two models on the new corpus: a monolingual BERT model trained from scratch (BERTu), and a further pre-trained multilingual BERT (mBERTu). The models achieve state-of-the-art performance on these tasks, despite the new corpus being considerably smaller than typically used corpora for high-resourced languages. On average, BERTu outperforms or performs competitively with mBERTu, and the largest gains are observed for higher-level tasks.

Patrícia Schmidtová, Saad Mahamood, Simone Balloccu et al.

Automatic metrics are extensively used to evaluate natural language processing systems. However, there has been increasing focus on how they are used and reported by practitioners within the field. In this paper, we have conducted a survey on the use of automatic metrics, focusing particularly on natural language generation (NLG) tasks. We inspect which metrics are used as well as why they are chosen and how their use is reported. Our findings from this survey reveal significant shortcomings, including inappropriate metric usage, lack of implementation details and missing correlations with human judgements. We conclude with recommendations that we believe authors should follow to enable more rigour within the field.

Michele Cafagna, Lina M. Rojas-Barahona, Kees van Deemter et al.

When applied to Image-to-text models, interpretability methods often provide token-by-token explanations namely, they compute a visual explanation for each token of the generated sequence. Those explanations are expensive to compute and unable to comprehensively explain the model's output. Therefore, these models often require some sort of approximation that eventually leads to misleading explanations. We develop a framework based on SHAP, that allows for generating comprehensive, meaningful explanations leveraging the meaning representation of the output sequence as a whole. Moreover, by exploiting semantic priors in the visual backbone, we extract an arbitrary number of features that allows the efficient computation of Shapley values on large-scale models, generating at the same time highly meaningful visual explanations. We demonstrate that our method generates semantically more expressive explanations than traditional methods at a lower compute cost and that it can be generalized over other explainability methods.

Claudia Tagliaferri, Sofia Axioti, Albert Gatt et al.

Derivationally related words, such as "runner" and "running", exhibit semantic differences which also elicit different visual scenarios. In this paper, we ask whether Vision and Language (V\&L) models capture such distinctions at the morphological level, using a a new methodology and dataset. We compare the results from V\&L models to human judgements and find that models' predictions differ from those of human participants, in particular displaying a grammatical bias. We further investigate whether the human-model misalignment is related to model architecture. Our methodology, developed on one specific morphological contrast, can be further extended for testing models on capturing other nuanced language features.

Marc Tanti, Shaun Abdilla, Adrian Muscat et al.

Current image description generation models do not transfer well to the task of describing human faces. To encourage the development of more human-focused descriptions, we developed a new data set of facial descriptions based on the CelebA image data set. We describe the properties of this data set, and present results from a face description generator trained on it, which explores the feasibility of using transfer learning from VGGFace/ResNet CNNs. Comparisons are drawn through both automated metrics and human evaluation by 76 English-speaking participants. The descriptions generated by the VGGFace-LSTM + Attention model are closest to the ground truth according to human evaluation whilst the ResNet-LSTM + Attention model obtained the highest CIDEr and CIDEr-D results (1.252 and 0.686 respectively). Together, the new data set and these experimental results provide data and baselines for future work in this area.

Michele Cafagna, Kees van Deemter, Albert Gatt

Current captioning datasets focus on object-centric captions, describing the visible objects in the image, e.g. "people eating food in a park". Although these datasets are useful to evaluate the ability of Vision & Language models to recognize and describe visual content, they do not support controlled experiments involving model testing or fine-tuning, with more high-level captions, which humans find easy and natural to produce. For example, people often describe images based on the type of scene they depict ('people at a holiday resort') and the actions they perform ('people having a picnic'). Such descriptions draw on personal experience and commonsense assumptions. We present the High-Level Dataset a dataset extending 14997 images from the COCO dataset, aligned with a new set of 134,973 human-annotated (high-level) captions collected along three axes: scenes, actions, and rationales. We further extend this dataset with confidence scores collected from an independent set of readers, as well as a set of narrative captions generated synthetically, by combining each of the three axes. We describe this dataset and analyse it extensively. We also present baseline results for the High-Level Captioning task.

Michele Cafagna, Kees van Deemter, Albert Gatt

Image captioning models tend to describe images in an object-centric way, emphasising visible objects. But image descriptions can also abstract away from objects and describe the type of scene depicted. In this paper, we explore the potential of a state-of-the-art Vision and Language model, VinVL, to caption images at the scene level using (1) a novel dataset which pairs images with both object-centric and scene descriptions. Through (2) an in-depth analysis of the effect of the fine-tuning, we show (3) that a small amount of curated data suffices to generate scene descriptions without losing the capability to identify object-level concepts in the scene; the model acquires a more holistic view of the image compared to when object-centric descriptions are generated. We discuss the parallels between these results and insights from computational and cognitive science research on scene perception.

Burak Kilic, Florix Bex, Albert Gatt

In this study, we analyze data-scarce classification scenarios, where available labeled legal data is small and imbalanced, potentially hurting the quality of the results. We focused on two finetuning objectives; SetFit (Sentence Transformer Finetuning), a contrastive learning setup, and a vanilla finetuning setup on a legal provision classification task. Additionally, we compare the features that are extracted with LIME (Local Interpretable Model-agnostic Explanations) to see which particular features contributed to the model's classification decisions. The results show that a contrastive setup with SetFit performed better than vanilla finetuning while using a fraction of the training samples. LIME results show that the contrastive learning approach helps boost both positive and negative features which are legally informative and contribute to the classification results. Thus a model finetuned with a contrastive objective seems to base its decisions more confidently on legally informative features.

Ivana Beňová, Michal Gregor, Albert Gatt

How do vision-language (VL) transformer models ground verb phrases and do they integrate contextual and world knowledge in this process? We introduce the CV-Probes dataset, containing image-caption pairs involving verb phrases that require both social knowledge and visual context to interpret (e.g., "beg"), as well as pairs involving verb phrases that can be grounded based on information directly available in the image (e.g., "sit"). We show that VL models struggle to ground VPs that are strongly context-dependent. Further analysis using explainable AI techniques shows that such models may not pay sufficient attention to the verb token in the captions. Our results suggest a need for improved methodologies in VL model training and evaluation. The code and dataset will be available https://github.com/ivana-13/CV-Probes.

Vincent Quantmeyer, Pablo Mosteiro, Albert Gatt

Various benchmarks have been proposed to test linguistic understanding in pre-trained vision \& language (VL) models. Here we build on the existence task from the VALSE benchmark (Parcalabescu et al, 2022) which we use to test models' understanding of negation, a particularly interesting issue for multimodal models. However, while such VL benchmarks are useful for measuring model performance, they do not reveal anything about the internal processes through which these models arrive at their outputs in such visio-linguistic tasks. We take inspiration from the growing literature on model interpretability to explain the behaviour of VL models on the understanding of negation. Specifically, we approach these questions through an in-depth analysis of the text encoder in CLIP (Radford et al, 2021), a highly influential VL model. We localise parts of the encoder that process negation and analyse the role of attention heads in this task. Our contributions are threefold. We demonstrate how methods from the language model interpretability literature (such as causal tracing) can be translated to multimodal models and tasks; we provide concrete insights into how CLIP processes negation on the VALSE existence task; and we highlight inherent limitations in the VALSE dataset as a benchmark for linguistic understanding.

Juliette Faille, Albert Gatt, Claire Gardent

In Natural Language Generation (NLG), important information is sometimes omitted in the output text. To better understand and analyse how this type of mistake arises, we focus on RDF-to-Text generation and explore two methods of probing omissions in the encoder output of BART (Lewis et al, 2020) and of T5 (Raffel et al, 2019): (i) a novel parameter-free probing method based on the computation of cosine similarity between embeddings of RDF graphs and of RDF graphs in which we removed some entities and (ii) a parametric probe which performs binary classification on the encoder embeddings to detect omitted entities. We also extend our analysis to distorted entities, i.e. entities that are not fully correctly mentioned in the generated text (e.g. misspelling of entity, wrong units of measurement). We found that both omitted and distorted entities can be probed in the encoder's output embeddings. This suggests that the encoder emits a weaker signal for these entities and therefore is responsible for some loss of information. This also shows that probing methods can be used to detect mistakes in the output of NLG models.

Yupei Du, Albert Gatt, Dong Nguyen

Despite the massive success of fine-tuning Pre-trained Language Models (PLMs), they remain susceptible to out-of-distribution input. Dataset cartography is a simple yet effective dual-model approach that improves the robustness of fine-tuned PLMs. It involves fine-tuning a model on the original training set (i.e. reference model), selecting a subset of important training instances based on the training dynamics, and fine-tuning again only on these selected examples (i.e. main model). However, this approach requires fine-tuning the same model twice, which is computationally expensive for large PLMs. In this paper, we show that (1) training dynamics are highly transferable across model sizes and pre-training methods, and that (2) fine-tuning main models using these selected training instances achieves higher training efficiency than empirical risk minimization (ERM). Building on these observations, we propose a novel fine-tuning approach: Fine-Tuning by transFerring Training dynamics (FTFT). Compared with dataset cartography, FTFT uses more efficient reference models and aggressive early stopping. FTFT achieves robustness improvements over ERM while lowering the training cost by up to $\sim 50\%$.

Ivan Martínez-Murillo, Elena Lloret, Paloma Moreda et al.

This paper explores the multilingual commonsense generation abilities of Large Language Models (LLMs). To facilitate this investigation, we introduce MULTICOM, a novel benchmark that extends the COCOTEROS dataset to four languages: English, Spanish, Dutch, and Valencian. The task involves generating a commonsensical sentence that includes a given triplet of words. We evaluate a range of open-source LLMs, including LLaMA, Qwen, Gemma, EuroLLM, and Salamandra, on this benchmark. Our evaluation combines automatic metrics, LLM-as-a-judge approaches (using Prometheus and JudgeLM), and human annotations. Results consistently show superior performance in English, with significantly lower performance in less-resourced languages. While contextual support yields mixed results, it tends to benefit underrepresented languages. These findings underscore the current limitations of LLMs in multilingual commonsense generation. The dataset is publicly available at https://huggingface.co/datasets/gplsi/MULTICOM.

Filippo Merlo, Ece Takmaz, Wenkai Chen et al.

Natural scenes provide us with rich contexts for object recognition and reference. In particular, knowing what type of scene one is looking at generates expectations about which objects will occur, and what their spatial configuration should be. Do Vision-Language Models (VLMs) learn to rely on scene contexts in a similar way, when generating references to objects? To address this question, we introduce the \textit{Common Objects Out-of-Context (COOCO)} dataset and test to what extent VLMs rely on scene context to refer to objects under different degrees of scene-object congruency, and different perturbations. Our findings show that models leverage scene context adaptively, depending on both the semantic relatedness between object and scene and the level of noise. In particular, models rely more on context under high target-scene congruence or when objects are degraded. Attention analysis reveals that successful object categorisation involves increased focus on the target in mid-level layers, especially under moderate noise, suggesting that VLMs dynamically balance local and contextual information for reference generation. We make our dataset, code and models available at \href{https://github.com/cs-nlp-uu/scenereg}{https://github.com/cs-nlp-uu/scenereg}.

Yingjin Song, Yupei Du, Denis Paperno et al.

This paper introduces the TempVS benchmark, which focuses on temporal grounding and reasoning capabilities of Multimodal Large Language Models (MLLMs) in image sequences. TempVS consists of three main tests (i.e., event relation inference, sentence ordering and image ordering), each accompanied with a basic grounding test. TempVS requires MLLMs to rely on both visual and linguistic modalities to understand the temporal order of events. We evaluate 38 state-of-the-art MLLMs, demonstrating that models struggle to solve TempVS, with a substantial performance gap compared to human capabilities. We also provide fine-grained insights that suggest promising directions for future research. Our TempVS benchmark data and code are available at https://github.com/yjsong22/TempVS.

Yingjin Song, Denis Paperno, Albert Gatt

Visual storytelling systems generate multi-sentence stories from image sequences. In this task, capturing contextual information and bridging visual variation bring additional challenges. We propose a simple yet effective framework that leverages the generalization capabilities of pretrained foundation models, only training a lightweight vision-language mapping network to connect modalities, while incorporating context to enhance coherence. We introduce a multimodal contrastive objective that also improves visual relevance and story informativeness. Extensive experimental results, across both automatic metrics and human evaluations, demonstrate that the stories generated by our framework are diverse, coherent, informative, and interesting.

Nan Li, Albert Gatt, Massimo Poesio

Collaborative dialogue relies on participants incrementally establishing common ground, yet in asymmetric settings they may believe they agree while referring to different entities. We introduce a perspectivist annotation scheme for the HCRC MapTask corpus (Anderson et al., 1991) that separately captures speaker and addressee grounded interpretations for each reference expression, enabling us to trace how understanding emerges, diverges, and repairs over time. Using a scheme-constrained LLM annotation pipeline, we obtain 13k annotated reference expressions with reliability estimates and analyze the resulting understanding states. The results show that full misunderstandings are rare once lexical variants are unified, but multiplicity discrepancies systematically induce divergences, revealing how apparent grounding can mask referential misalignment. Our framework provides both a resource and an analytic lens for studying grounded misunderstanding and for evaluating (V)LLMs' capacity to model perspective-dependent grounding in collaborative dialogue.

Hugh Mee Wong, Rick Nouwen, Albert Gatt

Multiple-choice question answering (MCQA) is easy to evaluate but adds a meta-task: models must both solve the problem and output the symbol that *represents* the answer, conflating reasoning errors with symbol-binding failures. We study how language models implement MCQA internally using representational analyses (PCA, linear probes) as well as causal interventions. We find that option-boundary (newline) residual states often contain strong linearly decodable signals related to per-option correctness. Winner-identity probing reveals a two-stage progression: the winning *content position* becomes decodable immediately after the final option is processed, while the *output symbol* is represented closer to the answer emission position. Tests under symbol and content permutations support a two-stage mechanism in which models first select a winner in content space and then bind or route that winner to the appropriate symbol to emit.

Max Schaffelder, Albert Gatt

As synthetic data becomes widely used in language model development, understanding its impact on model behavior is crucial. This paper investigates the impact of the diversity of sources of synthetic data on fine-tuned large language models. We focus on three key dimensions: distribution collapse, adversarial robustness, and self-preference bias. Our findings reveal that fine-tuning models on synthetic data from diverse sources can mitigate distribution collapse, preserving the breadth of the output distribution and the diversity of the output text. Furthermore, while both human and synthetic fine-tuning data can remove safeguards, the latter preserves higher output quality, thus making outputs potentially more usable and dangerous. Finally, fine-tuning reduces self-preference bias, with human data being the most effective, followed by multi-source synthetic data.

Daniil Ignatev, Ayman Santeer, Albert Gatt et al.

We propose a zero-shot method for Natural Language Inference (NLI) that leverages multimodal representations by grounding language in visual contexts. Our approach generates visual representations of premises using text-to-image models and performs inference by comparing these representations with textual hypotheses. We evaluate two inference techniques: cosine similarity and visual question answering. Our method achieves high accuracy without task-specific fine-tuning, demonstrating robustness against textual biases and surface heuristics. Additionally, we design a controlled adversarial dataset to validate the robustness of our approach. Our findings suggest that leveraging visual modality as a meaning representation provides a promising direction for robust natural language understanding.

Dongxu Lu, Johan Jeuring, Albert Gatt

Evaluating large language models (LLMs) in long-form, knowledge-grounded role-play dialogues remains challenging. This study compares LLM-generated and human-authored responses in multi-turn professional training simulations through human evaluation ($N=38$) and automated LLM-as-a-judge assessment. Human evaluation revealed significant degradation in LLM-generated response quality across turns, particularly in naturalness, context maintenance and overall quality, while human-authored responses progressively improved. In line with this finding, participants also indicated a consistent preference for human-authored dialogue. These human judgements were validated by our automated LLM-as-a-judge evaluation, where Gemini 2.0 Flash achieved strong alignment with human evaluators on both zero-shot pairwise preference and stochastic 6-shot construct ratings, confirming the widening quality gap between LLM and human responses over time. Our work contributes a multi-turn benchmark exposing LLM degradation in knowledge-grounded role-play dialogues and provides a validated hybrid evaluation framework to guide the reliable integration of LLMs in training simulations.

Ece Takmaz, Albert Gatt, Jakub Dotlacil

Images vary in how memorable they are to humans. Inspired by findings from cognitive science and computer vision, this paper explores the correlates of image memorability in pretrained vision encoders, focusing on latent activations, attention distributions, and the uniformity of image patches. We find that these features correlate with memorability to some extent. Additionally, we explore sparse autoencoder loss over the representations of vision transformers as a proxy for memorability, which yields results outperforming past methods using convolutional neural network representations. Our results shed light on the relationship between model-internal features and memorability. They show that some features are informative predictors of what makes images memorable to humans.

Admitos Passadakis, Yingjin Song, Albert Gatt

Visual Storytelling is a challenging multimodal task between Vision & Language, where the purpose is to generate a story for a stream of images. Its difficulty lies on the fact that the story should be both grounded to the image sequence but also narrative and coherent. The aim of this work is to balance between these aspects, by treating Visual Storytelling as a superset of Image Captioning, an approach quite different compared to most of prior relevant studies. This means that we firstly employ a vision-to-language model for obtaining captions of the input images, and then, these captions are transformed into coherent narratives using language-to-language methods. Our multifarious evaluation shows that integrating captioning and storytelling under a unified framework, has a positive impact on the quality of the produced stories. In addition, compared to numerous previous studies, this approach accelerates training time and makes our framework readily reusable and reproducible by anyone interested. Lastly, we propose a new metric/tool, named ideality, that can be used to simulate how far some results are from an oracle model, and we apply it to emulate human-likeness in visual storytelling.

Yupei Du, Yingjin Song, Hugh Mee Wong et al.

Data pruning, selecting small but impactful subsets, offers a promising way to efficiently scale NLP model training. However, existing methods often involve many different design choices, which have not been systematically studied. This limits future developments. In this work, we decompose data pruning into two key components: the data representation and the selection algorithm, and we systematically analyze their influence on the selection of instances. Our theoretical and empirical results highlight the crucial role of representations: better representations, e.g., training gradients, generally lead to a better selection of instances, regardless of the chosen selection algorithm. Furthermore, different selection algorithms excel in different settings, and none consistently outperforms the others. Moreover, the selection algorithms do not always align with their intended objectives: for example, algorithms designed for the same objective can select drastically different instances, highlighting the need for careful evaluation.

Silvia Casola, Yang Janet Liu, Siyao Peng et al.

Human language production exhibits remarkable richness and variation, reflecting diverse communication styles and intents. However, this variation is often overlooked in summarization evaluation. While having multiple reference summaries is known to improve correlation with human judgments, the impact of the reference set on reference-based metrics has not been systematically investigated. This work examines the sensitivity of widely used reference-based metrics in relation to the choice of reference sets, analyzing three diverse multi-reference summarization datasets: SummEval, GUMSum, and DUC2004. We demonstrate that many popular metrics exhibit significant instability. This instability is particularly concerning for n-gram-based metrics like ROUGE, where model rankings vary depending on the reference sets, undermining the reliability of model comparisons. We also collect human judgments on LLM outputs for genre-diverse data and examine their correlation with metrics to supplement existing findings beyond newswire summaries, finding weak-to-no correlation. Taken together, we recommend incorporating reference set variation into summarization evaluation to enhance consistency alongside correlation with human judgments, especially when evaluating LLMs.

Hugh Mee Wong, Rick Nouwen, Albert Gatt

Vague quantifiers such as "a few" and "many" are influenced by various contextual factors, including the number of objects present in a given context. In this work, we evaluate the extent to which vision-and-language models (VLMs) are compatible with humans when producing or judging the appropriateness of vague quantifiers in visual contexts. We release a novel dataset, VAQUUM, containing 20,300 human ratings on quantified statements across a total of 1089 images. Using this dataset, we compare human judgments and VLM predictions using three different evaluation methods. Our findings show that VLMs, like humans, are influenced by object counts in vague quantifier use. However, we find significant inconsistencies across models in different evaluation settings, suggesting that judging and producing vague quantifiers rely on two different processes.

Anna Bavaresco, Raffaella Bernardi, Leonardo Bertolazzi et al.

There is an increasing trend towards evaluating NLP models with LLMs instead of human judgments, raising questions about the validity of these evaluations, as well as their reproducibility in the case of proprietary models. We provide JUDGE-BENCH, an extensible collection of 20 NLP datasets with human annotations covering a broad range of evaluated properties and types of data, and comprehensively evaluate 11 current LLMs, covering both open-weight and proprietary models, for their ability to replicate the annotations. Our evaluations show substantial variance across models and datasets. Models are reliable evaluators on some tasks, but overall display substantial variability depending on the property being evaluated, the expertise level of the human judges, and whether the language is human or model-generated. We conclude that LLMs should be carefully validated against human judgments before being used as evaluators.

Leonardo Bertolazzi, Albert Gatt, Raffaella Bernardi

The reasoning abilities of Large Language Models (LLMs) are becoming a central focus of study in NLP. In this paper, we consider the case of syllogistic reasoning, an area of deductive reasoning studied extensively in logic and cognitive psychology. Previous research has shown that pre-trained LLMs exhibit reasoning biases, such as $\textit{content effects}$, avoid answering that $\textit{no conclusion follows}$, display human-like difficulties, and struggle with multi-step reasoning. We contribute to this research line by systematically investigating the effects of chain-of-thought reasoning, in-context learning (ICL), and supervised fine-tuning (SFT) on syllogistic reasoning, considering syllogisms with conclusions that support or violate world knowledge, as well as ones with multiple premises. Crucially, we go beyond the standard focus on accuracy, with an in-depth analysis of the conclusions generated by the models. Our results suggest that the behavior of pre-trained LLMs can be explained by heuristics studied in cognitive science and that both ICL and SFT improve model performance on valid inferences, although only the latter mitigates most reasoning biases without harming model consistency.

Anya Belz, Craig Thomson, Ehud Reiter et al.

We report our efforts in identifying a set of previous human evaluations in NLP that would be suitable for a coordinated study examining what makes human evaluations in NLP more/less reproducible. We present our results and findings, which include that just 13\% of papers had (i) sufficiently low barriers to reproduction, and (ii) enough obtainable information, to be considered for reproduction, and that all but one of the experiments we selected for reproduction was discovered to have flaws that made the meaningfulness of conducting a reproduction questionable. As a result, we had to change our coordinated study design from a reproduce approach to a standardise-then-reproduce-twice approach. Our overall (negative) finding that the great majority of human evaluations in NLP is not repeatable and/or not reproducible and/or too flawed to justify reproduction, paints a dire picture, but presents an opportunity for a rethink about how to design and report human evaluations in NLP.

Letitia Parcalabescu, Michele Cafagna, Lilitta Muradjan et al.

We propose VALSE (Vision And Language Structured Evaluation), a novel benchmark designed for testing general-purpose pretrained vision and language (V&L) models for their visio-linguistic grounding capabilities on specific linguistic phenomena. VALSE offers a suite of six tests covering various linguistic constructs. Solving these requires models to ground linguistic phenomena in the visual modality, allowing more fine-grained evaluations than hitherto possible. We build VALSE using methods that support the construction of valid foils, and report results from evaluating five widely-used V&L models. Our experiments suggest that current models have considerable difficulty addressing most phenomena. Hence, we expect VALSE to serve as an important benchmark to measure future progress of pretrained V&L models from a linguistic perspective, complementing the canonical task-centred V&L evaluations.

Andrea DeMarco, Carlos Mena, Albert Gatt et al.

Recent years have seen an increased interest in the computational speech processing of Maltese, but resources remain sparse. In this paper, we consider data augmentation techniques for improving speech recognition for low-resource languages, focusing on Maltese as a test case. We consider three different types of data augmentation: unsupervised training, multilingual training and the use of synthesized speech as training data. The goal is to determine which of these techniques, or combination of them, is the most effective to improve speech recognition for languages where the starting point is a small corpus of approximately 7 hours of transcribed speech. Our results show that combining the data augmentation techniques studied here lead us to an absolute WER improvement of 15% without the use of a language model.

Michele Cafagna, Kees van Deemter, Albert Gatt

Images can be described in terms of the objects they contain, or in terms of the types of scene or place that they instantiate. In this paper we address to what extent pretrained Vision and Language models can learn to align descriptions of both types with images. We compare 3 state-of-the-art models, VisualBERT, LXMERT and CLIP. We find that (i) V&L models are susceptible to stylistic biases acquired during pretraining; (ii) only CLIP performs consistently well on both object- and scene-level descriptions. A follow-up ablation study shows that CLIP uses object-level information in the visual modality to align with scene-level textual descriptions.

Marc Tanti, Lonneke van der Plas, Claudia Borg et al.

Recent work has shown evidence that the knowledge acquired by multilingual BERT (mBERT) has two components: a language-specific and a language-neutral one. This paper analyses the relationship between them, in the context of fine-tuning on two tasks -- POS tagging and natural language inference -- which require the model to bring to bear different degrees of language-specific knowledge. Visualisations reveal that mBERT loses the ability to cluster representations by language after fine-tuning, a result that is supported by evidence from language identification experiments. However, further experiments on 'unlearning' language-specific representations using gradient reversal and iterative adversarial learning are shown not to add further improvement to the language-independent component over and above the effect of fine-tuning. The results presented here suggest that the process of fine-tuning causes a reorganisation of the model's limited representational capacity, enhancing language-independent representations at the expense of language-specific ones.

Lorenzo De Mattei, Michele Cafagna, Huiyuan Lai et al.

An ongoing debate in the NLG community concerns the best way to evaluate systems, with human evaluation often being considered the most reliable method, compared to corpus-based metrics. However, tasks involving subtle textual differences, such as style transfer, tend to be hard for humans to perform. In this paper, we propose an evaluation method for this task based on purposely-trained classifiers, showing that it better reflects system differences than traditional metrics such as BLEU and ROUGE.

Letitia Parcalabescu, Albert Gatt, Anette Frank et al.

We investigate the reasoning ability of pretrained vision and language (V&L) models in two tasks that require multimodal integration: (1) discriminating a correct image-sentence pair from an incorrect one, and (2) counting entities in an image. We evaluate three pretrained V&L models on these tasks: ViLBERT, ViLBERT 12-in-1 and LXMERT, in zero-shot and finetuned settings. Our results show that models solve task (1) very well, as expected, since all models are pretrained on task (1). However, none of the pretrained V&L models is able to adequately solve task (2), our counting probe, and they cannot generalise to out-of-distribution quantities. We propose a number of explanations for these findings: LXMERT (and to some extent ViLBERT 12-in-1) show some evidence of catastrophic forgetting on task (1). Concerning our results on the counting probe, we find evidence that all models are impacted by dataset bias, and also fail to individuate entities in the visual input. While a selling point of pretrained V&L models is their ability to solve complex tasks, our findings suggest that understanding their reasoning and grounding capabilities requires more targeted investigations on specific phenomena.

Gaetana Ruggiero, Albert Gatt, Malvina Nissim

Existing research on Authorship Attribution (AA) focuses on texts for which a lot of data is available (e.g novels), mainly in English. We approach AA via Authorship Verification on short Italian texts in two novel datasets, and analyze the interaction between genre, topic, gender and length. Results show that AV is feasible even with little data, but more evidence helps. Gender and topic can be indicative clues, and if not controlled for, they might overtake more specific aspects of personal style.

Marion Bartl, Malvina Nissim, Albert Gatt

Contextualized word embeddings have been replacing standard embeddings as the representational knowledge source of choice in NLP systems. Since a variety of biases have previously been found in standard word embeddings, it is crucial to assess biases encoded in their replacements as well. Focusing on BERT (Devlin et al., 2018), we measure gender bias by studying associations between gender-denoting target words and names of professions in English and German, comparing the findings with real-world workforce statistics. We mitigate bias by fine-tuning BERT on the GAP corpus (Webster et al., 2018), after applying Counterfactual Data Substitution (CDS) (Maudslay et al., 2019). We show that our method of measuring bias is appropriate for languages such as English, but not for languages with a rich morphology and gender-marking, such as German. Our results highlight the importance of investigating bias and mitigation techniques cross-linguistically, especially in view of the current emphasis on large-scale, multilingual language models.

Stavros Assimakopoulos, Rebecca Vella Muskat, Lonneke van der Plas et al.

This paper presents a novel scheme for the annotation of hate speech in corpora of Web 2.0 commentary. The proposed scheme is motivated by the critical analysis of posts made in reaction to news reports on the Mediterranean migration crisis and LGBTIQ+ matters in Malta, which was conducted under the auspices of the EU-funded C.O.N.T.A.C.T. project. Based on the realization that hate speech is not a clear-cut category to begin with, appears to belong to a continuum of discriminatory discourse and is often realized through the use of indirect linguistic means, it is argued that annotation schemes for its detection should refrain from directly including the label 'hate speech,' as different annotators might have different thresholds as to what constitutes hate speech and what not. In view of this, we suggest a multi-layer annotation scheme, which is pilot-tested against a binary +/- hate speech classification and appears to yield higher inter-annotator agreement. Motivating the postulation of our scheme, we then present the MaNeCo corpus on which it will eventually be used; a substantial corpus of on-line newspaper comments spanning 10 years.

Carlos Mena, Albert Gatt, Andrea DeMarco et al.

Maltese, the national language of Malta, is spoken by approximately 500,000 people. Speech processing for Maltese is still in its early stages of development. In this paper, we present the first spoken Maltese corpus designed purposely for Automatic Speech Recognition (ASR). The MASRI-HEADSET corpus was developed by the MASRI project at the University of Malta. It consists of 8 hours of speech paired with text, recorded by using short text snippets in a laboratory environment. The speakers were recruited from different geographical locations all over the Maltese islands, and were roughly evenly distributed by gender. This paper also presents some initial results achieved in baseline experiments for Maltese ASR using Sphinx and Kaldi. The MASRI-HEADSET Corpus is publicly available for research/academic purposes.

Marc Tanti, Albert Gatt, Kenneth P. Camilleri

A neural language model can be conditioned into generating descriptions for images by providing visual information apart from the sentence prefix. This visual information can be included into the language model through different points of entry resulting in different neural architectures. We identify four main architectures which we call init-inject, pre-inject, par-inject, and merge. We analyse these four architectures and conclude that the best performing one is init-inject, which is when the visual information is injected into the initial state of the recurrent neural network. We confirm this using both automatic evaluation measures and human annotation. We then analyse how much influence the images have on each architecture. This is done by measuring how different the output probabilities of a model are when a partial sentence is combined with a completely different image from the one it is meant to be combined with. We find that init-inject tends to quickly become less influenced by the image as more words are generated. A different architecture called merge, which is when the visual information is merged with the recurrent neural network's hidden state vector prior to output, loses visual influence much more slowly, suggesting that it would work better for generating longer sentences. We also observe that the merge architecture can have its recurrent neural network pre-trained in a text-only language model (transfer learning) rather than be initialised randomly as usual. This results in even better performance than the other architectures, provided that the source language model is not too good at language modelling or it will overspecialise and be less effective at image description generation. Our work opens up new avenues of research in neural architectures, explainable AI, and transfer learning.

Somaye Jafaritazehjani, Albert Gatt, Marc Tanti

Natural Language Inference (NLI) is the task of determining the semantic relationship between a premise and a hypothesis. In this paper, we focus on the {\em generation} of hypotheses from premises in a multimodal setting, to generate a sentence (hypothesis) given an image and/or its description (premise) as the input. The main goals of this paper are (a) to investigate whether it is reasonable to frame NLI as a generation task; and (b) to consider the degree to which grounding textual premises in visual information is beneficial to generation. We compare different neural architectures, showing through automatic and human evaluation that entailments can indeed be generated successfully. We also show that multimodal models outperform unimodal models in this task, albeit marginally.

Angelo Basile, Albert Gatt, Malvina Nissim

Inspired by Labov's seminal work on stylistic variation as a function of social stratification, we develop and compare neural models that predict a person's presumed socio-economic status, obtained through distant supervision,from their writing style on social media. The focus of our work is on identifying the most important stylistic parameters to predict socio-economic group. In particular, we show the effectiveness of morpho-syntactic features as stylistic predictors of socio-economic group,in contrast to lexical features, which are good predictors of topic.

Marc Tanti, Albert Gatt, Kenneth P. Camilleri

When designing a neural caption generator, a convolutional neural network can be used to extract image features. Is it possible to also use a neural language model to extract sentence prefix features? We answer this question by trying different ways to transfer the recurrent neural network and embedding layer from a neural language model to an image caption generator. We find that image caption generators with transferred parameters perform better than those trained from scratch, even when simply pre-training them on the text of the same captions dataset it will later be trained on. We also find that the best language models (in terms of perplexity) do not result in the best caption generators after transfer learning.

Marc Tanti, Albert Gatt, Kenneth P. Camilleri

This paper addresses the sensitivity of neural image caption generators to their visual input. A sensitivity analysis and omission analysis based on image foils is reported, showing that the extent to which image captioning architectures retain and are sensitive to visual information varies depending on the type of word being generated and the position in the caption as a whole. We motivate this work in the context of broader goals in the field to achieve more explainability in AI.

Marc Tanti, Albert Gatt, Adrian Muscat

Image caption generation systems are typically evaluated against reference outputs. We show that it is possible to predict output quality without generating the captions, based on the probability assigned by the neural model to the reference captions. Such pre-gen metrics are strongly correlated to standard evaluation metrics.

Albert Gatt, Nicolás Marín, Gustavo Rivas-Gervilla et al.

In this paper we study empirically the validity of measures of referential success for referring expressions involving gradual properties. More specifically, we study the ability of several measures of referential success to predict the success of a user in choosing the right object, given a referring expression. Experimental results indicate that certain fuzzy measures of success are able to predict human accuracy in reference resolution. Such measures are therefore suitable for the estimation of the success or otherwise of a referring expression produced by a generation algorithm, especially in case the properties in a domain cannot be assumed to have crisp denotations.

Alejandro Ramos-Soto, Ehud Reiter, Kees van Deemter et al.

We present a data resource which can be useful for research purposes on language grounding tasks in the context of geographical referring expression generation. The resource is composed of two data sets that encompass 25 different geographical descriptors and a set of associated graphical representations, drawn as polygons on a map by two groups of human subjects: teenage students and expert meteorologists.

Steffen Pauws, Albert Gatt, Emiel Krahmer et al.

Healthcare organizations are in a continuous effort to improve health outcomes, reduce costs and enhance patient experience of care. Data is essential to measure and help achieving these improvements in healthcare delivery. Consequently, a data influx from various clinical, financial and operational sources is now overtaking healthcare organizations and their patients. The effective use of this data, however, is a major challenge. Clearly, text is an important medium to make data accessible. Financial reports are produced to assess healthcare organizations on some key performance indicators to steer their healthcare delivery. Similarly, at a clinical level, data on patient status is conveyed by means of textual descriptions to facilitate patient review, shift handover and care transitions. Likewise, patients are informed about data on their health status and treatments via text, in the form of reports or via ehealth platforms by their doctors. Unfortunately, such text is the outcome of a highly labour-intensive process if it is done by healthcare professionals. It is also prone to incompleteness, subjectivity and hard to scale up to different domains, wider audiences and varying communication purposes. Data-to-text is a recent breakthrough technology in artificial intelligence which automatically generates natural language in the form of text or speech from data. This chapter provides a survey of data-to-text technology, with a focus on how it can be deployed in a healthcare setting. It will (1) give an up-to-date synthesis of data-to-text approaches, (2) give a categorized overview of use cases in healthcare, (3) seek to make a strong case for evaluating and implementing data-to-text in a healthcare setting, and (4) highlight recent research challenges.