mSTEB: Massively Multilingual Evaluation of LLMs on Speech and Text Tasks

Large Language models (LLMs) have demonstrated impressive performance on a wide range of tasks, including in multimodal settings such as speech. However, their evaluation is often limited to English and a few high-resource languages. For low-resource languages, there is no standardized evaluation benchmark. In this paper, we address this gap by introducing mSTEB, a new benchmark to evaluate the performance of LLMs on a wide range of tasks covering language identification, text classification, question answering, and translation tasks on both speech and text modalities. We evaluated the performance of leading LLMs such as Gemini 2.0 Flash and GPT-4o (Audio) and state-of-the-art open models such as Qwen 2 Audio and Gemma 3 27B. Our evaluation shows a wide gap in performance between high-resource and low-resource languages, especially for languages spoken in Africa and Americas/Oceania. Our findings show that more investment is needed to address their under-representation in LLMs coverage.

CTDGSI: A comprehensive exploitation of instance selection methods for automatic text classification. VII Concurso de Teses, Dissertações e Trabalhos de Graduação em SI -- XXI Simpósio Brasileiro de Sistemas de Informação

Progress in Natural Language Processing (NLP) has been dictated by the rule of more: more data, more computing power and more complexity, best exemplified by the Large Language Models. However, training (or fine-tuning) large dense models for specific applications usually requires significant amounts of computing resources. This \textbf{Ph.D. dissertation} focuses on an under-investi\-gated NLP data engineering technique, whose potential is enormous in the current scenario known as Instance Selection (IS). The IS goal is to reduce the training set size by removing noisy or redundant instances while maintaining the effectiveness of the trained models and reducing the training process cost. We provide a comprehensive and scientifically sound comparison of IS methods applied to an essential NLP task -- Automatic Text Classification (ATC), considering several classification solutions and many datasets. Our findings reveal a significant untapped potential for IS solutions. We also propose two novel IS solutions that are noise-oriented and redundancy-aware, specifically designed for large datasets and transformer architectures. Our final solution achieved an average reduction of 41\% in training sets, while maintaining the same levels of effectiveness in all datasets. Importantly, our solutions demonstrated speedup improvements of 1.67x (up to 2.46x), making them scalable for datasets with hundreds of thousands of documents.

Approximating Language Model Training Data from Weights

Modern language models often have open weights but closed training data. We formalize the problem of data approximation from model weights and propose several baselines and metrics. We develop a gradient-based approach that selects the highest-matching data from a large public text corpus and show its effectiveness at recovering useful data given only weights of the original and finetuned models. Even when none of the true training data is known, our method is able to locate a small subset of public Web documents can be used to train a model to close to the original model performance given models trained for both classification and supervised-finetuning. On the AG News classification task, our method improves performance from 65% (using randomly selected data) to 80%, approaching the expert benchmark of 88%. When applied to a model trained with SFT on MSMARCO web documents, our method reduces perplexity from 3.3 to 2.3, compared to an expert LLAMA model's perplexity of 2.0.

Unified Representation Space for 3D Visual Grounding

3D visual grounding (3DVG) is a critical task in scene understanding that aims to identify objects in 3D scenes based on text descriptions. However, existing methods rely on separately pre-trained vision and text encoders, resulting in a significant gap between the two modalities in terms of spatial geometry and semantic categories. This discrepancy often causes errors in object positioning and classification. The paper proposes UniSpace-3D, which innovatively introduces a unified representation space for 3DVG, effectively bridging the gap between visual and textual features. Specifically, UniSpace-3D incorporates three innovative designs: i) a unified representation encoder that leverages the pre-trained CLIP model to map visual and textual features into a unified representation space, effectively bridging the gap between the two modalities; ii) a multi-modal contrastive learning module that further reduces the modality gap; iii) a language-guided query selection module that utilizes the positional and semantic information to identify object candidate points aligned with textual descriptions. Extensive experiments demonstrate that UniSpace-3D outperforms baseline models by at least 2.24% on the ScanRefer and Nr3D/Sr3D datasets. The code will be made available upon acceptance of the paper.

CLIP the Landscape: Automated Tagging of Crowdsourced Landscape Images

We present a CLIP-based, multi-modal, multi-label classifier for predicting geographical context tags from landscape photos in the Geograph dataset--a crowdsourced image archive spanning the British Isles, including remote regions lacking POIs and street-level imagery. Our approach addresses a Kaggle competition\footnote{https://www.kaggle.com/competitions/predict-geographic-context-from-landscape-photos} task based on a subset of Geograph's 8M images, with strict evaluation: exact match accuracy is required across 49 possible tags. We show that combining location and title embeddings with image features improves accuracy over using image embeddings alone. We release a lightweight pipeline\footnote{https://github.com/SpaceTimeLab/ClipTheLandscape} that trains on a modest laptop, using pre-trained CLIP image and text embeddings and a simple classification head. Predicted tags can support downstream tasks such as building location embedders for GeoAI applications, enriching spatial understanding in data-sparse regions.

Self-Anchored Attention Model for Sample-Efficient Classification of Prosocial Text Chat

Millions of players engage daily in competitive online games, communicating through in-game chat. Prior research has focused on detecting relatively small volumes of toxic content using various Natural Language Processing (NLP) techniques for the purpose of moderation. However, recent studies emphasize the importance of detecting prosocial communication, which can be as crucial as identifying toxic interactions. Recognizing prosocial behavior allows for its analysis, rewarding, and promotion. Unlike toxicity, there are limited datasets, models, and resources for identifying prosocial behaviors in game-chat text. In this work, we employed unsupervised discovery combined with game domain expert collaboration to identify and categorize prosocial player behaviors from game chat. We further propose a novel Self-Anchored Attention Model (SAAM) which gives 7.9% improvement compared to the best existing technique. The approach utilizes the entire training set as "anchors" to help improve model performance under the scarcity of training data. This approach led to the development of the first automated system for classifying prosocial behaviors in in-game chats, particularly given the low-resource settings where large-scale labeled data is not available. Our methodology was applied to one of the most popular online gaming titles - Call of Duty(R): Modern Warfare(R)II, showcasing its effectiveness. This research is novel in applying NLP techniques to discover and classify prosocial behaviors in player in-game chat communication. It can help shift the focus of moderation from solely penalizing toxicity to actively encouraging positive interactions on online platforms.

TRIDENT: Temporally Restricted Inference via DFA-Enhanced Neural Traversal

Large Language Models (LLMs) and other neural architectures have achieved impressive results across a variety of generative and classification tasks. However, they remain fundamentally ill-equipped to ensure that their outputs satisfy temporal constraints, such as those expressible in Linear Temporal Logic over finite traces (LTLf). In this paper, we introduce TRIDENT: a general and model-agnostic inference-time algorithm that guarantees compliance with such constraints without requiring any retraining. TRIDENT compiles LTLf formulas into a Deterministic Finite Automaton (DFA), which is used to guide a constrained variant of beam search. At each decoding step, transitions that would lead to constraint violations are masked, while remaining paths are dynamically re-ranked based on both the model's probabilities and the DFA's acceptance structure. We formally prove that the resulting sequences are guaranteed to satisfy the given LTLf constraints, and we empirically demonstrate that TRIDENT also improves output quality. We validate our approach on two distinct tasks: temporally constrained image-stream classification and controlled text generation. In both settings, TRIDENT achieves perfect constraint satisfaction, while comparison with the state of the art shows improved efficiency and high standard quality metrics.

Improving Medical Visual Representation Learning with Pathological-level Cross-Modal Alignment and Correlation Exploration

Learning medical visual representations from image-report pairs through joint learning has garnered increasing research attention due to its potential to alleviate the data scarcity problem in the medical domain. The primary challenges stem from the lengthy reports that feature complex discourse relations and semantic pathologies. Previous works have predominantly focused on instance-wise or token-wise cross-modal alignment, often neglecting the importance of pathological-level consistency. This paper presents a novel framework PLACE that promotes the Pathological-Level Alignment and enriches the fine-grained details via Correlation Exploration without additional human annotations. Specifically, we propose a novel pathological-level cross-modal alignment (PCMA) approach to maximize the consistency of pathology observations from both images and reports. To facilitate this, a Visual Pathology Observation Extractor is introduced to extract visual pathological observation representations from localized tokens. The PCMA module operates independently of any external disease annotations, enhancing the generalizability and robustness of our methods. Furthermore, we design a proxy task that enforces the model to identify correlations among image patches, thereby enriching the fine-grained details crucial for various downstream tasks. Experimental results demonstrate that our proposed framework achieves new state-of-the-art performance on multiple downstream tasks, including classification, image-to-text retrieval, semantic segmentation, object detection and report generation.

A Semantically-Aware Relevance Measure for Content-Based Medical Image Retrieval Evaluation

Performance evaluation for Content-Based Image Retrieval (CBIR) remains a crucial but unsolved problem today especially in the medical domain. Various evaluation metrics have been discussed in the literature to solve this problem. Most of the existing metrics (e.g., precision, recall) are adapted from classification tasks which require manual labels as ground truth. However, such labels are often expensive and unavailable in specific thematic domains. Furthermore, medical images are usually associated with (radiological) case reports or annotated with descriptive captions in literature figures, such text contains information that can help to assess CBIR.Several researchers have argued that the medical concepts hidden in the text can serve as the basis for CBIR evaluation purpose. However, these works often consider these medical concepts as independent and isolated labels while in fact the subtle relationships between various concepts are neglected. In this work, we introduce the use of knowledge graphs to measure the distance between various medical concepts and propose a novel relevance measure for the evaluation of CBIR by defining an approximate matching-based relevance score between two sets of medical concepts which allows us to indirectly measure the similarity between medical images.We quantitatively demonstrate the effectiveness and feasibility of our relevance measure using a public dataset.

LTG at SemEval-2025 Task 10: Optimizing Context for Classification of Narrative Roles

Our contribution to the SemEval 2025 shared task 10, subtask 1 on entity framing, tackles the challenge of providing the necessary segments from longer documents as context for classification with a masked language model. We show that a simple entity-oriented heuristics for context selection can enable text classification using models with limited context window. Our context selection approach and the XLM-RoBERTa language model is on par with, or outperforms, Supervised Fine-Tuning with larger generative language models.