Dialectal data are characterized by linguistic variation that appears small to humans but has a significant impact on the performance of models. This dialect gap has been related to various factors (e.g., data size, economic and social factors) whose impact, however, turns out to be inconsistent. In this work, we investigate factors impacting the model performance more directly: we correlate Tokenization Parity (TP) and Information Parity (IP), as measures of representational biases in pre-trained multilingual models, with the downstream performance. We compare state-of-the-art decoder-only LLMs with encoder-based models across three tasks: dialect classification, topic classification, and extractive question answering, controlling for varying scripts (Latin vs. non-Latin) and resource availability (high vs. low). Our analysis reveals that TP is a better predictor of the performance on tasks reliant on syntactic and morphological cues (e.g., extractive QA), while IP better predicts performance in semantic tasks (e.g., topic classification). Complementary analyses, including tokenizer behavior, vocabulary coverage, and qualitative insights, reveal that the language support claims of LLMs often might mask deeper mismatches at the script or token level.

Recent papers show LLMs achieve near-random accuracy in causal relation classification, raising questions about whether such failures arise from limited pretraining exposure or deeper representational gaps. We investigate this under uncertainty-based evaluation, testing whether pretraining exposure to causal examples improves causal understanding using >18K PubMed sentences—half from The Pile corpus, half post-2024—across seven models (Pythia-1.4B/7B/12B, GPT-J-6B, Dolly-7B/12B, Qwen-7B). We analyze model behavior through: (i) causal classification, where the model identifies causal relationships in text, and (ii) verbatim memorization probing, where we assess whether the model prefers previously seen causal statements over their paraphrases. Models perform four-way classification (direct/conditional/correlational/no-relationship) and select between originals and their generated paraphrases. Results show almost identical accuracy on seen/unseen sentences (p>0.05), no memorization bias (24.8% original selection), output distribution over the possible options almost flat — with entropic values near the maximum (1.35/1.39), confirming random guessing. Instruction-tuned models show severe miscalibration (Qwen: >95% confidence, 32.8% accuracy, ECE=0.49). Conditional relations induce highest entropy (+11% vs direct). These findings suggest that failures in causal understanding arise from the lack of structured causal representation, rather than insufficient exposure to causal examples during pretraining.

Structured information extraction (IE) from scientific abstracts is increasingly leveraging large language models (LLMs). A crucial step in IE is relation extraction (RE), which becomes challenging when entity relations span sentences. Traditional path-based methods, such as shortest dependency paths, are often unable to handle cross-sentential relations effectively. Although LLMs have been utilized as zero-shot learners for IE tasks, they continue to struggle with capturing long-range dependencies and multi-hop reasoning. In this work, we propose using GPT as a zero-shot entity-guided summarizer to encapsulate cross-sentential context into a single-sentence summary for relation extraction. We perform intrinsic evaluations, comparing our approach against direct zero-shot prompting on biomedical scientific abstracts. On the Chemical-Disease Relation (CDR) dataset, our method achieves a 7-point improvement in overall F-score and 6 points for cross-sentential relations. On the Gene-Disease Association (GDA) dataset, we observe an 8-point gain for inter-sentential relations. These results demonstrate that entity-guided summarization with GPT can enhance zero-shot biomedical RE, supporting more effective structured information extraction from scientific texts.

We report the approaches submitted to the NADI 2024 Subtask 1: Multi-label country-level Dialect Identification (MLDID). The core part was to adapt the information from multi-class data for a multi-label dialect classification task. We experimented with supervised and unsupervised strategies to tackle the task in this challenging setting. Under the supervised setup, we used the model trained using NADI 2023 data and devised approaches to convert the multi-class predictions to multi-label by using information from the confusion matrix or using calibrated probabilities. Under unsupervised settings, we used the Arabic-based sentence encoders and multilingual cross-encoders to retrieve similar samples from the training set, considering each test input as a query. The associated labels are then assigned to the input query. We also tried different variations, such as co-occurring dialects derived from the provided development set. We obtained the best validation performance of 48.5% F-score using one of the variations with an unsupervised approach and the same approach yielded the best test result of 43.27% (Ranked 2).

Pre-trained models usually come with a pre-defined tokenization and little flexibility as to what subword tokens can be used in downstream tasks. This problem concerns especially multilingual NLP and low-resource languages, which are typically processed using cross-lingual transfer. In this paper, we aim to find out if the right granularity of tokenization is helpful for a text classification task, namely dialect classification. Aiming at generalizations beyond the studied cases, we look for the optimal granularity in four dialect datasets, two with relatively consistent writing (one Arabic and one Indo-Aryan set) and two with considerably inconsistent writing (one Arabic and one Swiss German set). To gain more control over subword tokenization and ensure direct comparability in the experimental settings, we train a CNN classifier from scratch comparing two subword tokenization methods (Unigram model and BPE). For reference, we compare the results obtained in our analysis to the state of the art achieved by fine-tuning pre-trained models. We show that models trained from scratch with an optimal tokenization level perform better than fine-tuned classifiers in the case of highly inconsistent writing. In the case of relatively consistent writing, fine-tuned models remain better regardless of the tokenization level.

This paper deals with the problem of incre-mental dialect identification. Our goal is toreliably determine the dialect before the fullutterance is given as input. The major partof the previous research on dialect identification has been model-centric, focusing on performance. We address a new question: How much input is needed to identify a dialect? Ourapproach is a data-centric analysis that resultsin general criteria for finding the shortest inputneeded to make a plausible guess. Workingwith three sets of language dialects (Swiss German, Indo-Aryan and Arabic languages), weshow that it is possible to generalize across dialects and datasets with two input shorteningcriteria: model confidence and minimal inputlength (adjusted for the input type). The sourcecode for experimental analysis can be found atGithub.

In this paper, we describe our systems submitted to the NADI Subtask 1: country-wise dialect classifications. We designed two types of solutions. The first type is convolutional neural network CNN) classifiers trained on subword segments of optimized lengths. The second type is fine-tuned classifiers with BERT-based language specific pre-trained models. To deal with the missing dialects in one of the test sets, we experimented with binary classifiers, analyzing the predicted probability distribution patterns and comparing them with the development set patterns. The better performing approach on the development set was fine-tuning language specific pre-trained model (best F-score 26.59%). On the test set, on the other hand, we obtained the best performance with the CNN model trained on subword tokens obtained with a Unigram model (the best F-score 26.12%). Re-training models on samples of training data simulating missing dialects gave the maximum performance on the test set version with a number of dialects lesser than the training set (F-score 16.44%)

Lexical semantic change detection (also known as semantic shift tracing) is a task of identifying words that have changed their meaning over time. Unsupervised semantic shift tracing, focal point of SemEval2020, is particularly challenging. Given the unsupervised setup, in this work, we propose to identify clusters among different occurrences of each target word, considering these as representatives of different word meanings. As such, disagreements in obtained clusters naturally allow to quantify the level of semantic shift per each target word in four target languages. To leverage this idea, clustering is performed on contextualized (BERT-based) embeddings of word occurrences. The obtained results show that our approach performs well both measured separately (per language) and overall, where we surpass all provided SemEval baselines.