Iacer Calixto

Ningyuan Huang, Yash R. Deshpande, Yibo Liu et al. · amazon-science

We propose a method to make natural language understanding models more parameter efficient by storing knowledge in an external knowledge graph (KG) and retrieving from this KG using a dense index. Given (possibly multilingual) downstream task data, e.g., sentences in German, we retrieve entities from the KG and use their multimodal representations to improve downstream task performance. We use the recently released VisualSem KG as our external knowledge repository, which covers a subset of Wikipedia and WordNet entities, and compare a mix of tuple-based and graph-based algorithms to learn entity and relation representations that are grounded on the KG multimodal information. We demonstrate the usefulness of the learned entity representations on two downstream tasks, and show improved performance on the multilingual named entity recognition task by $0.3\%$--$0.7\%$ F1, while we achieve up to $2.5\%$ improvement in accuracy on the visual sense disambiguation task. All our code and data are available in: \url{https://github.com/iacercalixto/visualsem-kg}.

Auke Elfrink, Iacopo Vagliano, Ameen Abu-Hanna et al.

We investigate different natural language processing (NLP) approaches based on contextualised word representations for the problem of early prediction of lung cancer using free-text patient medical notes of Dutch primary care physicians. Because lung cancer has a low prevalence in primary care, we also address the problem of classification under highly imbalanced classes. Specifically, we use large Transformer-based pretrained language models (PLMs) and investigate: 1) how \textit{soft prompt-tuning} -- an NLP technique used to adapt PLMs using small amounts of training data -- compares to standard model fine-tuning; 2) whether simpler static word embedding models (WEMs) can be more robust compared to PLMs in highly imbalanced settings; and 3) how models fare when trained on notes from a small number of patients. We find that 1) soft-prompt tuning is an efficient alternative to standard model fine-tuning; 2) PLMs show better discrimination but worse calibration compared to simpler static word embedding models as the classification problem becomes more imbalanced; and 3) results when training models on small number of patients are mixed and show no clear differences between PLMs and WEMs. All our code is available open source in \url{https://bitbucket.org/aumc-kik/prompt_tuning_cancer_prediction/}.

Mustafa Dogan, Ilker Kesen, Iacer Calixto et al.

As multimodal large language models (MLLMs) advance in handling interleaved image-text data, assessing their few-shot learning capabilities remains an open challenge. In this paper, we introduce FewMMBench, a comprehensive benchmark designed to evaluate MLLMs under few-shot conditions, with a focus on In-Context Learning (ICL) and Chain-of-Thought (CoT) prompting. Covering a diverse suite of multimodal understanding tasks, from attribute recognition to temporal reasoning, FewMMBench enables systematic analysis across task types, model families, and prompting strategies. We evaluate 26 open-weight MLLMs from six model families across zero-shot, few-shot, and CoT-augmented few-shot settings. Our findings reveal that instruction-tuned models exhibit strong zero-shot performance but benefit minimally, or even regress, with additional demonstrations or CoT reasoning. Retrieval-based demonstrations and increased context size also yield limited gains. These results highlight FewMMBench as a rigorous testbed for diagnosing and advancing few-shot capabilities in multimodal LLMs. The data is available at: https://huggingface.co/datasets/mustafaa/FewMMBench

Nishant Mishra, Gaurav Sahu, Iacer Calixto et al.

Generating high-quality summaries for chat dialogs often requires large labeled datasets. We propose a method to efficiently use unlabeled data for extractive summarization of customer-agent dialogs. In our method, we frame summarization as a question-answering problem and use state-of-the-art large language models (LLMs) to generate pseudo-labels for a dialog. We then use these pseudo-labels to fine-tune a chat summarization model, effectively transferring knowledge from the large LLM into a smaller specialized model. We demonstrate our method on the \tweetsumm dataset, and show that using 10% of the original labelled data set we can achieve 65.9/57.0/61.0 ROUGE-1/-2/-L, whereas the current state-of-the-art trained on the entire training data set obtains 65.16/55.81/64.37 ROUGE-1/-2/-L. In other words, in the worst case (i.e., ROUGE-L) we still effectively retain 94.7% of the performance while using only 10% of the data.

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.

İlker Kesen, Aykut Erdem, Erkut Erdem et al.

This paper describes our two-stage system for the Euphemism Detection shared task hosted by the 3rd Workshop on Figurative Language Processing in conjunction with EMNLP 2022. Euphemisms tone down expressions about sensitive or unpleasant issues like addiction and death. The ambiguous nature of euphemistic words or expressions makes it challenging to detect their actual meaning within a context. In the first stage, we seek to mitigate this ambiguity by incorporating literal descriptions into input text prompts to our baseline model. It turns out that this kind of direct supervision yields remarkable performance improvement. In the second stage, we integrate visual supervision into our system using visual imageries, two sets of images generated by a text-to-image model by taking terms and descriptions as input. Our experiments demonstrate that visual supervision also gives a statistically significant performance boost. Our system achieved the second place with an F1 score of 87.2%, only about 0.9% worse than the best submission.

Zeyu Zhang, Paul Groth, Iacer Calixto et al.

Entity matching (EM) is the problem of determining whether two records refer to same real-world entity, which is crucial in data integration, e.g., for product catalogs or address databases. A major drawback of many EM approaches is their dependence on labelled examples. We thus focus on the challenging setting of zero-shot entity matching where no labelled examples are available for an unseen target dataset. Recently, large language models (LLMs) have shown promising results for zero-shot EM, but their low throughput and high deployment cost limit their applicability and scalability. We revisit the zero-shot EM problem with AnyMatch, a small language model fine-tuned in a transfer learning setup. We propose several novel data selection techniques to generate fine-tuning data for our model, e.g., by selecting difficult pairs to match via an AutoML filter, by generating additional attribute-level examples, and by controlling label imbalance in the data. We conduct an extensive evaluation of the prediction quality and deployment cost of our model, in a comparison to thirteen baselines on nine benchmark datasets. We find that AnyMatch provides competitive prediction quality despite its small parameter size: it achieves the second-highest F1 score overall, and outperforms several other approaches that employ models with hundreds of billions of parameters. Furthermore, our approach exhibits major cost benefits: the average prediction quality of AnyMatch is within 4.4% of the state-of-the-art method MatchGPT with the proprietary trillion-parameter model GPT-4, yet AnyMatch requires four orders of magnitude less parameters and incurs a 3,899 times lower inference cost (in dollars per 1,000 tokens).

Mustafa Dogan, Ilker Kesen, Iacer Calixto et al.

The linguistic capabilities of Multimodal Large Language Models (MLLMs) are critical for their effective application across diverse tasks. This study aims to evaluate the performance of MLLMs on the VALSE benchmark, focusing on the efficacy of few-shot In-Context Learning (ICL), and Chain-of-Thought (CoT) prompting. We conducted a comprehensive assessment of state-of-the-art MLLMs, varying in model size and pretraining datasets. The experimental results reveal that ICL and CoT prompting significantly boost model performance, particularly in tasks requiring complex reasoning and contextual understanding. Models pretrained on captioning datasets show superior zero-shot performance, while those trained on interleaved image-text data benefit from few-shot learning. Our findings provide valuable insights into optimizing MLLMs for better grounding of language in visual contexts, highlighting the importance of the composition of pretraining data and the potential of few-shot learning strategies to improve the reasoning abilities of MLLMs.

Michele Miranda, Xinlan Yan, Nishant Mishra et al.

Protecting patient privacy in clinical narratives is essential for enabling secondary use of healthcare data under regulations such as GDPR and HIPAA. While manual de-identification remains the gold standard, it is costly and slow, motivating the need for automated methods that combine privacy guarantees with high utility. Most automated text de-identification pipelines employed named entity recognition (NER) to identify protected entities for redaction. Although methods based on differential privacy (DP) provide formal privacy guarantees, more recently also large language models (LLMs) are increasingly used for text de-identification in the clinical domain. In this work, we present the first comparative study of DP, NER, and LLMs for Dutch clinical text de-identification. We investigate these methods separately as well as hybrid strategies that apply NER or LLM preprocessing prior to DP, and assess performance in terms of privacy leakage and extrinsic evaluation (entity and relation classification). We show that DP mechanisms alone degrade utility substantially, but combining them with linguistic preprocessing, especially LLM-based redaction, significantly improves the privacy-utility trade-off.

Alberto Testoni, Iacer Calixto

Safe clinical deployment of Large Language Models (LLMs) requires not only high accuracy but also robust uncertainty calibration to ensure models defer to clinicians when appropriate. Our paper investigates how social descriptors of a patient (specifically sexual orientation and religious affiliation) distort these uncertainty signals and model accuracy. Evaluating nine general-purpose and biomedical LLMs on 2,364 medical questions and their counterfactual variants, we demonstrate that identity markers cause a "calibration crisis". "Homosexual" markers consistently trigger performance drops, and intersectional identities produce idiosyncratic, non-additive harms to calibration. Moreover, a clinician-validated case study in an open-ended generation setting confirms that these failures are not an artifact of the multiple-choice format. Our results demonstrate that the presence of social identity cues does not merely shift predictions; it affects the reliability of confidence signals, posing a significant risk to equitable care and safe deployment in confidence-based clinical workflows.

Giovanni Cinà, Daniel Fernandez-Llaneza, Ludovico Deponte et al.

Feature attribution methods have become a staple method to disentangle the complex behavior of black box models. Despite their success, some scholars have argued that such methods suffer from a serious flaw: they do not allow a reliable interpretation in terms of human concepts. Simply put, visualizing an array of feature contributions is not enough for humans to conclude something about a model's internal representations, and confirmation bias can trick users into false beliefs about model behavior. We argue that a structured approach is required to test whether our hypotheses on the model are confirmed by the feature attributions. This is what we call the "semantic match" between human concepts and (sub-symbolic) explanations. Building on the conceptual framework put forward in Cinà et al. [2023], we propose a structured approach to evaluate semantic match in practice. We showcase the procedure in a suite of experiments spanning tabular and image data, and show how the assessment of semantic match can give insight into both desirable (e.g., focusing on an object relevant for prediction) and undesirable model behaviors (e.g., focusing on a spurious correlation). We couple our experimental results with an analysis on the metrics to measure semantic match, and argue that this approach constitutes the first step towards resolving the issue of confirmation bias in XAI.

Nishant Mishra, Wilker Aziz, Iacer Calixto

Progress in biomedical Named Entity Recognition (NER) and Entity Linking (EL) is currently hindered by a fragmented data landscape, a lack of resources for building explainable models, and the limitations of semantically-blind evaluation metrics. To address these challenges, we present MedPath, a large-scale and multi-domain biomedical EL dataset that builds upon nine existing expert-annotated EL datasets. In MedPath, all entities are 1) normalized using the latest version of the Unified Medical Language System (UMLS), 2) augmented with mappings to 62 other biomedical vocabularies and, crucially, 3) enriched with full ontological paths -- i.e., from general to specific -- in up to 11 biomedical vocabularies. MedPath directly enables new research frontiers in biomedical NLP, facilitating training and evaluation of semantic-rich and interpretable EL systems, and the development of the next generation of interoperable and explainable clinical NLP models.

Alberto Testoni, Iacer Calixto

Reliable uncertainty quantification (UQ) is essential when employing large language models (LLMs) in high-risk domains such as clinical question answering (QA). In this work, we evaluate uncertainty estimation methods for clinical QA focusing, for the first time, on eleven clinical specialties and six question types, and across ten open-source LLMs (general-purpose, biomedical, and reasoning models). We analyze score-based UQ methods, present a case study introducing a novel lightweight method based on behavioral features derived from reasoning-oriented models, and examine conformal prediction as a complementary set-based approach. Our findings reveal that uncertainty reliability is not a monolithic property, but one that depends on clinical specialty and question type due to shifts in calibration and discrimination. Our results highlight the need to select or ensemble models based on their distinct, complementary strengths and clinical use.

Victor Milewski, Marie-Francine Moens, Iacer Calixto

Many top-performing image captioning models rely solely on object features computed with an object detection model to generate image descriptions. However, recent studies propose to directly use scene graphs to introduce information about object relations into captioning, hoping to better describe interactions between objects. In this work, we thoroughly investigate the use of scene graphs in image captioning. We empirically study whether using additional scene graph encoders can lead to better image descriptions and propose a conditional graph attention network (C-GAT), where the image captioning decoder state is used to condition the graph updates. Finally, we determine to what extent noise in the predicted scene graphs influence caption quality. Overall, we find no significant difference between models that use scene graph features and models that only use object detection features across different captioning metrics, which suggests that existing scene graph generation models are still too noisy to be useful in image captioning. Moreover, although the quality of predicted scene graphs is very low in general, when using high quality scene graphs we obtain gains of up to 3.3 CIDEr compared to a strong Bottom-Up Top-Down baseline. We open source code to reproduce all our experiments in https://github.com/iacercalixto/butd-image-captioning.

Houda Alberts, Iacer Calixto

Datasets (semi-)automatically collected from the web can easily scale to millions of entries, but a dataset's usefulness is directly related to how clean and high-quality its examples are. In this paper, we describe and publicly release an image dataset along with pretrained models designed to (semi-)automatically filter out undesirable images from very large image collections, possibly obtained from the web. Our dataset focusses on photographic and/or natural images, a very common use-case in computer vision research. We provide annotations for coarse prediction, i.e. photographic vs. non-photographic, and smaller fine-grained prediction tasks where we further break down the non-photographic class into five classes: maps, drawings, graphs, icons, and sketches. Results on held out validation data show that a model architecture with reduced memory footprint achieves over 96% accuracy on coarse-prediction. Our best model achieves 88% accuracy on the hardest fine-grained classification task available. Dataset and pretrained models are available at: https://github.com/houda96/imagi-filter.

Houda Alberts, Teresa Huang, Yash Deshpande et al.

An exciting frontier in natural language understanding (NLU) and generation (NLG) calls for (vision-and-) language models that can efficiently access external structured knowledge repositories. However, many existing knowledge bases only cover limited domains, or suffer from noisy data, and most of all are typically hard to integrate into neural language pipelines. To fill this gap, we release VisualSem: a high-quality knowledge graph (KG) which includes nodes with multilingual glosses, multiple illustrative images, and visually relevant relations. We also release a neural multi-modal retrieval model that can use images or sentences as inputs and retrieves entities in the KG. This multi-modal retrieval model can be integrated into any (neural network) model pipeline. We encourage the research community to use VisualSem for data augmentation and/or as a source of grounding, among other possible uses. VisualSem as well as the multi-modal retrieval models are publicly available and can be downloaded in this URL: https://github.com/iacercalixto/visualsem

Camila Zurdo Tagliabue, Heloisa Oss Boll, Aykut Erdem et al.

Large language models (LLMs) are increasingly used in clinical decision support, yet current evaluation methods often fail to distinguish genuine medical reasoning from superficial patterns. We introduce DeVisE (Demographics and Vital signs Evaluation), a behavioral testing framework for probing fine-grained clinical understanding. We construct a dataset of ICU discharge notes from MIMIC-IV, generating both raw (real-world) and template-based (synthetic) versions with controlled single-variable counterfactuals targeting demographic (age, gender, ethnicity) and vital sign attributes. We evaluate five LLMs spanning general-purpose and medically fine-tuned variants, under both zero-shot and fine-tuned settings. We assess model behavior via (1) input-level sensitivity - how counterfactuals alter the likelihood of a note; and (2) downstream reasoning - how they affect predicted hospital length-of-stay. Our results show that zero-shot models exhibit more coherent counterfactual reasoning patterns, while fine-tuned models tend to be more stable yet less responsive to clinically meaningful changes. Notably, demographic factors subtly but consistently influence outputs, emphasizing the importance of fairness-aware evaluation. This work highlights the utility of behavioral testing in exposing the reasoning strategies of clinical LLMs and informing the design of safer, more transparent medical AI systems.

Heloisa Oss Boll, Antonio Oss Boll, Leticia Puttlitz Boll et al.

Large language models (LLMs) offer unprecedented opportunities to generate concise summaries of patient information and alleviate the burden of clinical documentation that overwhelms healthcare providers. We present Distillnote, a framework for LLM-based clinical note summarization, and generate over 64,000 admission note summaries through three techniques: (1) One-step, direct summarization, and a divide-and-conquer approach involving (2) Structured summarization focused on independent clinical insights, and (3) Distilled summarization that further condenses the Structured summaries. We test how useful are the summaries by using them to predict heart failure compared to a model trained on the original notes. Distilled summaries achieve 79% text compression and up to 18.2% improvement in AUPRC compared to an LLM trained on the full notes. We also evaluate the quality of the generated summaries in an LLM-as-judge evaluation as well as through blinded pairwise comparisons with clinicians. Evaluations indicate that one-step summaries are favoured by clinicians according to relevance and clinical actionability, while distilled summaries offer optimal efficiency (avg. 6.9x compression-to-performance ratio) and significantly reduce hallucinations. We release our summaries on PhysioNet to encourage future research.

Xinlan Yan, Di Wu, Yibin Lei et al.

In this paper, we introduce S-MedQA, an English medical question-answering (QA) dataset for benchmarking large language models (LLMs) in fine-grained clinical specialties. S-MedQA has over 20k examples, covers 15 medical specialties, and QA pairs can have multiple specialty annotations (e.g., when a question is cross-disciplinary), constructed with both machine and expert verification to maximize data availability. We use S-MedQA to investigate the role of clinical specialty data in the knowledge-intensive scenario of medical QA. Our results show that 1) training on data from a clinical specialty does not necessarily lead to best performance on that specialty, and 2) regardless of the specialty the LLM was fine-tuned on, token probabilities of clinically relevant terms increase consistently across all specialties. Thus, we hypothesize improvement gains are derived mostly from domain shifting (e.g., general to medical) rather than specialty-specific knowledge injection, and suggest rethinking the role of fine-tuning data in the medical domain.

Rachel M. Murphy, Nishant Mishra, Nicolette F. de Keizer et al.

In this study, we establish a benchmark for adverse drug event (ADE) detection in Dutch clinical free-text documents using several transformer models, clinical scenarios, and fit-for-purpose performance measures. We trained a Bidirectional Long Short-Term Memory (Bi-LSTM) model and four transformer-based Dutch and/or multilingual encoder models (BERTje, RobBERT, MedRoBERTa(.)nl, and NuNER) for the tasks of named entity recognition (NER) and relation classification (RC) using 102 richly annotated Dutch ICU clinical progress notes. Anonymized free-text clinical progress notes of patients admitted to the intensive care unit (ICU) of one academic hospital and discharge letters of patients admitted to Internal Medicine wards of two non-academic hospitals were reused. We evaluated our ADE RC models internally using the gold standard (two-step task) and predicted entities (end-to-end task). In addition, all models were externally validated for detecting ADEs at the document level. We report both micro- and macro-averaged F1 scores, given the dataset imbalance in ADEs. Although differences for the ADE RC task between the models were small, MedRoBERTa(.)nl was the best performing model with a macro-averaged F1 score of 0.63 using the gold standard and 0.62 using predicted entities. The MedRoBERTa(.)nl models also performed the best in our external validation and achieved a recall of between 0.67 to 0.74 using predicted entities, meaning between 67 to 74% of discharge letters with ADEs were detected. Our benchmark study presents a robust and clinically meaningful approach for evaluating language models for ADE detection in clinical free-text documents. Our study highlights the need to use appropriate performance measures fit for the task of ADE detection in clinical free-text documents and envisioned future clinical use.

Vittorio Torri, Machteld J. Boonstra, Marielle C. van de Veerdonk et al.

Objective: Heart failure (HF) patients present with diverse phenotypes affecting treatment and prognosis. This study evaluates models for phenotyping HF patients based on left ventricular ejection fraction (LVEF) classes, using structured and unstructured data, assessing performance and interpretability. Materials and Methods: The study analyzes all HF hospitalizations at both Amsterdam UMC hospitals (AMC and VUmc) from 2015 to 2023 (33,105 hospitalizations, 16,334 patients). Data from AMC were used for model training, and from VUmc for external validation. The dataset was unlabelled and included tabular clinical measurements and discharge letters. Silver labels for LVEF classes were generated by combining diagnosis codes, echocardiography results, and textual mentions. Gold labels were manually annotated for 300 patients for testing. Multiple Transformer-based (black-box) and Aug-Linear (white-box) models were trained and compared with baselines on structured and unstructured data. To evaluate interpretability, two clinicians annotated 20 discharge letters by highlighting information they considered relevant for LVEF classification. These were compared to SHAP and LIME explanations from black-box models and the inherent explanations of Aug-Linear models. Results: BERT-based and Aug-Linear models, using discharge letters alone, achieved the highest classification results (AUC=0.84 for BERT, 0.81 for Aug-Linear on external validation), outperforming baselines. Aug-Linear explanations aligned more closely with clinicians' explanations than post-hoc explanations on black-box models. Conclusions: Discharge letters emerged as the most informative source for phenotyping HF patients. Aug-Linear models matched black-box performance while providing clinician-aligned interpretability, supporting their use in transparent clinical decision-making.

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.

Ruan Chaves Rodrigues, Marcelo Akira Inuzuka, Juliana Resplande Sant'Anna Gomes et al.

Hashtag segmentation, also known as hashtag decomposition, is a common step in preprocessing pipelines for social media datasets. It usually precedes tasks such as sentiment analysis and hate speech detection. For sentiment analysis in medium to low-resourced languages, previous research has demonstrated that a multilingual approach that resorts to machine translation can be competitive or superior to previous approaches to the task. We develop a zero-shot hashtag segmentation framework and demonstrate how it can be used to improve the accuracy of multilingual sentiment analysis pipelines. Our zero-shot framework establishes a new state-of-the-art for hashtag segmentation datasets, surpassing even previous approaches that relied on feature engineering and language models trained on in-domain data.

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.

Elham J. Barezi, Iacer Calixto, Kyunghyun Cho et al.

Extreme classification tasks are multi-label tasks with an extremely large number of labels (tags). These tasks are hard because the label space is usually (i) very large, e.g. thousands or millions of labels, (ii) very sparse, i.e. very few labels apply to each input document, and (iii) highly correlated, meaning that the existence of one label changes the likelihood of predicting all other labels. In this work, we propose a self-attention based variational encoder-model to extract the label-label and label-feature dependencies jointly and to predict labels for a given input. In more detail, we propose a non-autoregressive latent variable model and compare it to a strong autoregressive baseline that predicts a label based on all previously generated labels. Our model can therefore be used to predict all labels in parallel while still including both label-label and label-feature dependencies through latent variables, and compares favourably to the autoregressive baseline. We apply our models to four standard extreme classification natural language data sets, and one news videos dataset for automated label detection from a lexicon of semantic concepts. Experimental results show that although the autoregressive models, where use a given order of the labels for chain-order label prediction, work great for the small scale labels or the prediction of the highly ranked label, but our non-autoregressive model surpasses them by around 2% to 6% when we need to predict more labels, or the dataset has a larger number of the labels.

Jason Phang, Iacer Calixto, Phu Mon Htut et al.

Intermediate-task training---fine-tuning a pretrained model on an intermediate task before fine-tuning again on the target task---often improves model performance substantially on language understanding tasks in monolingual English settings. We investigate whether English intermediate-task training is still helpful on non-English target tasks. Using nine intermediate language-understanding tasks, we evaluate intermediate-task transfer in a zero-shot cross-lingual setting on the XTREME benchmark. We see large improvements from intermediate training on the BUCC and Tatoeba sentence retrieval tasks and moderate improvements on question-answering target tasks. MNLI, SQuAD and HellaSwag achieve the best overall results as intermediate tasks, while multi-task intermediate offers small additional improvements. Using our best intermediate-task models for each target task, we obtain a 5.4 point improvement over XLM-R Large on the XTREME benchmark, setting the state of the art as of June 2020. We also investigate continuing multilingual MLM during intermediate-task training and using machine-translated intermediate-task data, but neither consistently outperforms simply performing English intermediate-task training.

Iacer Calixto, Miguel Rios, Wilker Aziz

In this work, we propose to model the interaction between visual and textual features for multi-modal neural machine translation (MMT) through a latent variable model. This latent variable can be seen as a multi-modal stochastic embedding of an image and its description in a foreign language. It is used in a target-language decoder and also to predict image features. Importantly, our model formulation utilises visual and textual inputs during training but does not require that images be available at test time. We show that our latent variable MMT formulation improves considerably over strong baselines, including a multi-task learning approach (Elliott and Kádár, 2017) and a conditional variational auto-encoder approach (Toyama et al., 2016). Finally, we show improvements due to (i) predicting image features in addition to only conditioning on them, (ii) imposing a constraint on the minimum amount of information encoded in the latent variable, and (iii) by training on additional target-language image descriptions (i.e. synthetic data).

Iacer Calixto, Qun Liu, Nick Campbell

We introduce a Multi-modal Neural Machine Translation model in which a doubly-attentive decoder naturally incorporates spatial visual features obtained using pre-trained convolutional neural networks, bridging the gap between image description and translation. Our decoder learns to attend to source-language words and parts of an image independently by means of two separate attention mechanisms as it generates words in the target language. We find that our model can efficiently exploit not just back-translated in-domain multi-modal data but also large general-domain text-only MT corpora. We also report state-of-the-art results on the Multi30k data set.

Iacer Calixto, Qun Liu, Nick Campbell

We propose a novel discriminative model that learns embeddings from multilingual and multi-modal data, meaning that our model can take advantage of images and descriptions in multiple languages to improve embedding quality. To that end, we introduce a modification of a pairwise contrastive estimation optimisation function as our training objective. We evaluate our embeddings on an image-sentence ranking (ISR), a semantic textual similarity (STS), and a neural machine translation (NMT) task. We find that the additional multilingual signals lead to improvements on both the ISR and STS tasks, and the discriminative cost can also be used in re-ranking $n$-best lists produced by NMT models, yielding strong improvements.

Iacer Calixto, Qun Liu, Nick Campbell

We introduce multi-modal, attention-based neural machine translation (NMT) models which incorporate visual features into different parts of both the encoder and the decoder. We utilise global image features extracted using a pre-trained convolutional neural network and incorporate them (i) as words in the source sentence, (ii) to initialise the encoder hidden state, and (iii) as additional data to initialise the decoder hidden state. In our experiments, we evaluate how these different strategies to incorporate global image features compare and which ones perform best. We also study the impact that adding synthetic multi-modal, multilingual data brings and find that the additional data have a positive impact on multi-modal models. We report new state-of-the-art results and our best models also significantly improve on a comparable phrase-based Statistical MT (PBSMT) model trained on the Multi30k data set according to all metrics evaluated. To the best of our knowledge, it is the first time a purely neural model significantly improves over a PBSMT model on all metrics evaluated on this data set.