Most Large Vision-Language Models (LVLMs) to date are trained predominantly on English data, which makes them struggle to understand non-English input and fail to generate output in the desired target language. Existing efforts mitigate these issues by adding multilingual training data, but do so in a largely ad-hoc manner, lacking insight into how different training mixes tip the scale for different groups of languages. In this work, we present a comprehensive investigation into the training strategies for massively multilingual LVLMs. First, we conduct a series of multi-stage experiments spanning 13 downstream vision-language tasks and 43 languages, systematically examining: (1) the number of training languages that can be included without degrading English performance and (2) optimal language distributions of pre-training as well as (3) instruction-tuning data. Further, we (4) investigate how to improve multilingual text-in-image understanding, and introduce a new benchmark for the task. Surprisingly, our analysis reveals that one can (i) include as many as 100 training languages simultaneously (ii) with as little as 25-50% of non-English data, to greatly improve multilingual performance while retaining strong English performance. We further find that (iii) including non-English OCR data in pre-training and instruction-tuning is paramount for improving multilingual text-in-image understanding. Finally, we put all our findings together and train , a 100-language LVLM, offering state-of-the-art performance in an evaluation covering 14 tasks and 56 languages.

Even with various regulations in place across countries and social media platforms (Government of India, 2021; European Parliament and Council of the European Union, 2022), digital abusive speech remains a significant issue. One potential approach to address this challenge is automatic text detoxification, a text style transfer (TST) approach that transforms toxic language into a more neutral or non-toxic form. To date, the availability of parallel corpora for the text detoxification task (Logacheva et al., 2022; Atwell et al., 2022; Dementieva et al., 2024a) has proven to be crucial for state-of-the-art approaches. With this work, we extend parallel text detoxification corpus to new languages—German, Chinese, Arabic, Hindi, and Amharic—testing in the extensive multilingual setup TST baselines. Next, we conduct the first of its kind an automated, explainable analysis of the descriptive features of both toxic and non-toxic sentences, diving deeply into the nuances, similarities, and differences of toxicity and detoxification across 9 languages. Finally, based on the obtained insights, we experiment with a novel text detoxification method inspired by the Chain-of-Thoughts reasoning approach, enhancing the prompting process through clustering on relevant descriptive attributes.

Large Vision-Language Models (LVLMs) have recently gained attention due to their distinctive performance and broad applicability. While it has been previously shown that their efficacy in usage scenarios involving non-Western contexts falls short, existing studies are limited in scope, covering just a narrow range of cultures, focusing exclusively on a small number of cultural aspects, or evaluating a limited selection of models on a single task only. Towards globally inclusive LVLM research, we introduce GIMMICK, an extensive multimodal benchmark designed to assess a broad spectrum of cultural knowledge across 144 countries representing six global macro-regions. GIMMICK comprises six tasks built upon three new datasets that span 728 unique cultural events or facets on which we evaluated 20 LVLMs and 11 LLMs, including five proprietary and 26 open-weight models of all sizes. We systematically examine (1) regional cultural biases, (2) the influence of model size, (3) input modalities, and (4) external cues. Our analyses reveal strong biases toward Western cultures across models and tasks and highlight strong correlations between model size and performance, as well as the effectiveness of multimodal input and external geographic cues. We further find that models have more knowledge of tangible than intangible aspects (e.g., food vs. rituals) and that they excel in recognizing broad cultural origins but struggle with a more nuanced understanding.

Existing multilingual vision-language (VL) benchmarks often only cover a handful of languages. Consequently, evaluations of large vision-language models (LVLMs) predominantly target high-resource languages, underscoring the need for evaluation data for low-resource languages. To address this limitation, we introduce MVL-SIB, a massively multilingual vision-language benchmark that evaluates both cross-modal and text-only topical matching across 205 languages – over 100 more than the most multilingual existing VL benchmarks encompass. We then benchmark a range of of open-weight LVLMs together with GPT-4o(-mini) on MVL-SIB. Our results reveal that LVLMs struggle in cross-modal topic matching in lower-resource languages, performing no better than chance on languages like N’Koo. Our analysis further reveals that VL support in LVLMs declines disproportionately relative to textual support for lower-resource languages, as evidenced by comparison of cross-modal and text-only topical matching performance. We further observe that open-weight LVLMs do not benefit from representing a topic with more than one image, suggesting that these models are not yet fully effective at handling multi-image tasks. By correlating performance on MVL-SIB with other multilingual VL benchmarks, we highlight that MVL-SIB serves as a comprehensive probe of multilingual VL understanding in LVLMs.

Large Language Models (LLMs) such as ChatGPT, GitHub Copilot, Llama, or Mistral assist programmers as copilots and knowledge sources to make the coding process faster and more efficient. This paper aims to improve the copilot performance by implementing different self-alignment processes and retrieval-augmented generation (RAG) pipelines, as well as their combination. To test the effectiveness of all approaches, we create a dataset and apply a model-based evaluation, using LLM as a judge. It is designed to check the model’s abilities to understand the source code semantics, the dependency between files, and the overall meta-information about the repository. We also compare our approach with other existing solutions, e.g. ChatGPT-3.5, and evaluate on the existing benchmarks. Code and dataset are available online (https://anonymous.4open.science/r/ma_llm-382D).

This paper presents WISMIR3, a multi-modal dataset comprising roughly 300K text-image pairs from Wikipedia. With a sophisticated automatic ETL pipeline, we scraped, filtered, and transformed the data so that WISMIR3 intrinsically differs from other popular text-image datasets like COCO and Flickr30k. We prove this difference by comparing various linguistic statistics between the three datasets computed using the pipeline. The primary purpose of WISMIR3 is to use it as a benchmark to challenge state-of-the-art text-image retrieval approaches, which already reach around 90% Recall@5 scores on the mentioned popular datasets. Therefore, we ran several text-image retrieval experiments on our dataset using current models, which show that the models, in fact, perform significantly worse compared to evaluation results on COCO and Flickr30k. In addition, for each text-image pair, we release features computed by Faster-R-CNN and CLIP models. With this, we want to ease and motivate the use of the dataset for other researchers.

Since the release of ChatGPT, the field of Natural Language Processing has experienced rapid advancements, particularly in Large Language Models (LLMs) and their multimodal counterparts, Large Multimodal Models (LMMs). Despite their impressive capabilities, LLMs often exhibit significant performance disparities across different languages and cultural contexts, as demonstrated by various text-only benchmarks. However, current research lacks such benchmarks for multimodal visio-linguistic settings. This work fills this gap by introducing M5, the first comprehensive benchmark designed to evaluate LMMs on diverse vision-language tasks within a multilingual and multicultural context. M5 includes eight datasets covering five tasks and 41 languages, with a focus on underrepresented languages and culturally diverse images. Furthermore, we introduce two novel datasets, M5-VGR and M5-VLOD, including a new Visio-Linguistic Outlier Detection task, in which all evaluated open-source models fail to significantly surpass the random baseline. Through extensive evaluation and analyses, we highlight substantial task-agnostic performance disparities between high- and low-resource languages. Moreover, we show that larger models do not necessarily outperform smaller ones in a multilingual setting.

We uncover a surprising multilingual bias occurring in a popular class of multimodal vision-language models (VLMs). Including an image in the query to a LLaVA-style VLM significantly increases the likelihood of the model returning an English response, regardless of the language of the query. This paper investigates the causes of this loss with a two-pronged approach that combines extensive ablation of the design space with a mechanistic analysis of the models’ internal representations of image and text inputs. Both approaches indicate that the issue stems in the language modeling component of the LLaVA model. Statistically, we find that switching the language backbone for a bilingual language model has the strongest effect on reducing this error. Mechanistically, we provide compelling evidence that visual inputs are not mapped to a similar space as text ones, and that intervening on intermediary attention layers can reduce this bias. Our findings provide important insights to researchers and engineers seeking to understand the crossover between multimodal and multilingual spaces, and contribute to the goal of developing capable and inclusive VLMs for non-English contexts.

In this system demonstration paper, we describe the Whiteboards extension for an existing web-based platform for digital qualitative discourse analysis. Whiteboards comprise interactive graph-based interfaces to organize and manipulate objects, which can be qualitative research data, such as documents, images, etc., and analyses of these research data, such as annotations, tags, and code structures. The proposed extension offers a customizable view of the material and a wide range of actions that enable new ways of interacting and working with such resources. We show that the visualizations facilitate various use cases of qualitative data analysis, including reflection of the research process through sampling maps, creation of actor networks, and refining code taxonomies.

In this system demonstration paper, we present the Concept Over Time Analysis extension for the Discourse Analysis Tool Suite.The proposed tool empowers users to define, refine, and visualize their concepts of interest within an interactive interface. Adhering to the Human-in-the-loop paradigm, users can give feedback through sentence annotations. Utilizing few-shot sentence classification, the system employs Sentence Transformers to compute representations of sentences and concepts. Through an iterative process involving semantic similarity searches, sentence annotation, and fine-tuning with contrastive data, the model continuously refines, providing users with enhanced analysis outcomes. The final output is a timeline visualization of sentences classified to concepts. Especially suited for the Digital Humanities, Concept Over Time Analysis serves as a valuable tool for qualitative data analysis within extensive datasets. The chronological overview of concepts enables researchers to uncover patterns, trends, and shifts in discourse over time.

Images increasingly constitute a larger portion of internet content, encoding even more complex meanings. Recent studies have highlight the pivotal role of visual communication in the spread of extremist content, particularly that associated with right-wing political ideologies. However, the capability of machine learning systems to recognize such meanings, sometimes implicit, remains limited. To enable future research in this area, we introduce and release VIDA, the Visual Incel Data Archive, a multimodal dataset comprising visual material and internet memes collected from two main Incel communities (Italian and Anglophone) known for their extremist misogynistic content. Following the analytical framework of Shifman (2014), we propose a new taxonomy for annotation across three main levels of analysis: content, form, and stance (hate). This allows for the association of images with fine-grained contextual information that help to identify the presence of offensiveness and a broader set of cultural references, enhancing the understanding of more nuanced aspects in visual communication. In this work we present a statistical analysis of the annotated dataset as well as discuss annotation examples and future line of research.

This work introduces the D-WISE Tool Suite (DWTS), a novel working environment for digital qualitative discourse analysis in the Digital Humanities (DH). The DWTS addresses limitations of current DH tools induced by the ever-increasing amount of heterogeneous, unstructured, and multi-modal data in which the discourses of contemporary societies are encoded. To provide meaningful insights from such data, our system leverages and combines state-of-the-art machine learning technologies from Natural Language Processing and Com-puter Vision. Further, the DWTS is conceived and developed by an interdisciplinary team ofcultural anthropologists and computer scientists to ensure the tool’s usability for modernDH research. Central features of the DWTS are: a) import of multi-modal data like text, image, audio, and video b) preprocessing pipelines for automatic annotations c) lexical and semantic search of documents d) manual span, bounding box, time-span, and frame annotations e) documentation of the research process.

WebAnno is one of the most popular annotation tools that supports generic annotation types and distributive annotation with multiple user roles. However, WebAnno focuses on annotating span-level mentions and relations among them, making document-level annotation complicated. When it comes to the annotation and analysis of social science materials, it usually involves the creation of codes to categorize a given document. The codes, which are known as codebooks, are typically hierarchical, which enables to code the document either with a general category or more fine-grained subcategories. CodeAnno is forked from WebAnno and designed to solve the coding problems faced by many social science researchers with the following main functionalities. 1) Creation of hierarchical codebooks, with functionality to move and sort categories in the hierarchy 2) an interactive UI for codebook annotation 3) import and export of annotations in CSV format, hence being compatible with existing annotations conducted using spreadsheet applications 4) integration of an external automation component to facilitate coding using machine learning 5) project templating that allows duplicating a project structure without copying the actual documents. We present different use-cases to demonstrate the capability of CodeAnno. A shot demonstration video of the system is available here: https://www.youtube.com/watch?v=RmCdTghBe-s

The objective of the SemEval-2023 Task 1: Visual Word Sense Disambiguation (VWSD) is to identify the image illustrating the indented meaning of a target word and some minimal additional context. The omnipresence of textual and visual data in the task strongly suggests the utilization of the recent advances in multi-modal machine learning, i.e., pretrained visiolinguistic models (VLMs). Often referred to as foundation models due to their strong performance on many vision-language downstream tasks, these models further demonstrate powerful zero-shot capabilities. In this work, we utilize various pertained VLMs in a zero-shot fashion for multiple approaches using external knowledge sources to enrich the contextual information. Further, we evaluate our methods on the final test data and extensively analyze the suitability of different knowledge sources, the influence of training data, model sizes, multi-linguality, and different textual prompting strategies. Although we are not among the best-performing systems (rank 20 of 56), our experiments described in this work prove competitive results. Moreover, we aim to contribute meaningful insights and propel multi-modal machine learning tasks like VWSD.

The multi-modal foundation model CLIP computes representations from texts and images that achieved unprecedented performance on tasks such as zero-shot image classification. However, CLIP was pretrained on public internet data. Thus it lacks highly domain-specific knowledge. We investigate the adaptation of CLIP-based models to the chest radiography domain using the MIMIC-CXR dataset. We show that the features of the pretrained CLIP models do not transfer to this domain. We adapt CLIP to the chest radiography domain using contrastive language supervision and show that this approach yields a model that outperforms supervised learning on labels on the MIMIC-CXR dataset while also generalizing to the CheXpert and RSNA Pneumonia datasets. Furthermore, we do a detailed ablation study of the batch and dataset size. Finally, we show that language supervision allows for better explainability by using the multi-modal model to generate images from texts such that experts can inspect what the model has learned.

MOTIF (MultimOdal ConTextualized Images For Language Learners) is a multimodal dataset that consists of 1125 comprehension texts retrieved from Wikipedia Simple Corpus. Allowing multimodal processing or enriching the context with multimodal information has proven imperative for many learning tasks, specifically for second language (L2) learning. In this respect, several traditional NLP approaches can assist L2 readers in text comprehension processes, such as simplifying text or giving dictionary descriptions for complex words. As nicely stated in the well-known proverb, sometimes “a picture is worth a thousand words” and an image can successfully complement the verbal message by enriching the representation, like in Pictionary books. This multimodal support can also assist on-the-fly text reading experience by providing a multimodal tool that chooses and displays the most relevant images for the difficult words, given the text context. This study mainly focuses on one of the key components to achieving this goal; collecting a multimodal dataset enriched with complex word annotation and validated image match.

In primary school, children’s books, as well as in modern language learning apps, multi-modal learning strategies like illustrations of terms and phrases are used to support reading comprehension. Also, several studies in educational psychology suggest that integrating cross-modal information will improve reading comprehension. We claim that state-of- he-art multi-modal transformers, which could be used in a language learner context to improve human reading, will perform poorly because of the short and relatively simple textual data those models are trained with. To prove our hypotheses, we collected a new multi-modal image-retrieval dataset based on data from Wikipedia. In an in-depth data analysis, we highlight the differences between our dataset and other popular datasets. Additionally, we evaluate several state-of-the-art multi-modal transformers on text-image retrieval on our dataset and analyze their meager results, which verify our claims.