This report presents the results of the shared tasks organized as part of the VarDial Evaluation Campaign 2024. The campaign is part of the eleventh workshop on Natural Language Processing (NLP) for Similar Languages, Varieties and Dialects (VarDial), co-located with NAACL 2024. Two shared tasks were included this year: dialectal causal commonsense reasoning (DIALECT-COPA), and Multi-label classification of similar languages (DSL-ML). Both tasks were organized for the first time this year, but DSL-ML partially overlaps with the DSL-TL task organized in 2023.

This study investigates the factors influencing the performance of multilingual large language models (MLLMs) across diverse languages. We study 6 MLLMs, including masked language models, autoregressive models, and instruction-tuned LLMs, on the SIB-200 dataset, a topic classification dataset encompassing 204 languages. Our analysis considers three scenarios: ALL languages, SEEN languages (present in the model’s pretraining data), and UNSEEN languages (not present or documented in the model’s pretraining data in any meaningful way). We examine the impact of factors such as pretraining data size, general resource availability, language family, and script type on model performance. Decision tree analysis reveals that pretraining data size is the most influential factor for SEEN languages. However, interestingly, script type and language family become more crucial for UNSEEN languages, highlighting the importance of cross-lingual transfer learning. Notably, model size and architecture do not significantly alter the most important features identified. Our findings provide valuable insights into the strengths and limitations of current MLLMs and hope to guide the development of more effective and equitable multilingual NLP systems.

Whisper is a state-of-the-art automatic speech recognition (ASR) model (Radford et al., 2022). Although Swiss German dialects are allegedly not part of Whisper’s training data, preliminary experiments showed Whisper can transcribe Swiss German quite well, with the output being a speech translation into Standard German. To gain a better understanding of Whisper’s performance on Swiss German, we systematically evaluate it using automatic, qualitative, and human evaluation. We test its performance on three existing test sets: SwissDial (Dogan-Schönberger et al., 2021), STT4SG-350 (Plüss et al., 2023), and Swiss Parliaments Corpus (Plüss et al., 2021). In addition, we create a new test set for this study based on short mock clinical interviews. To automatically evaluate performance, we used word error rate (WER) and BLEU. We also conducted a qualitative analysis of Whisper’s performance, discussing its strengths and weaknesses. Finally, 28 people participated in a survey evaluating Whisper’s performance. All of our evaluations showed that Whisper is a viable ASR system for Swiss German, so long as the Standard German output is desired.

How well does naturally-occurring digital text, such as Tweets, represent sub-dialects of Egyptian Arabic (EA)? This paper focuses on two EA sub-dialects: Cairene Egyptian Arabic (CEA) and Sa’idi Egyptian Arabic (SEA). We use morphological markers from ground-truth dialect surveys as a distance measure across four geo-referenced datasets. Results show that CEA markers are prevalent as expected in CEA geo-referenced tweets, while SEA markers are limited across SEA geo-referenced tweets. SEA tweets instead show a prevalence of CEA markers and higher usage of Modern Standard Arabic. We conclude that corpora intended to represent sub-dialects of EA do not accurately represent sub-dialects outside of the Cairene variety. This finding calls into question the validity of relying on tweets alone to represent dialectal differences.

Spanish is an official language in 20 countries; in 19 of them, it arrived by means of overseas colonisation. Its close contact with several coexistent languages and the rich regional and cultural diversity has produced varieties which divert from each other. We study these divergences in a data-based approach and according to their qualitative and quantitative effects in word embeddings. We generate embeddings for Spanish in 24 countries and examine the topology of the spaces. Due to the similarities between varieties —in contrast to what happens to different languages in bilingual topological studies— we first scrutinise the behaviour of three isomorphism measures in (quasi-)isomorphic settings: relational similarity, Eigenvalue similarity and Gromov-Hausdorff distance. We then use the most trustworthy measure to quantify the divergences among varieties. Finally, we use the departures from isomorphism to build relational trees for the Spanish varieties by hierarchical clustering.

Effectively normalizing spellings in textual data poses a considerable challenge, especially for low-resource languages lacking standardized writing systems. In this study, we fine-tuned a multilingual model with data from several Occitan dialects and conducted a series of experiments to assess the model’s representations of these dialects. For evaluation purposes, we compiled a parallel lexicon encompassing four Occitan dialects.Intrinsic evaluations of the model’s embeddings revealed that surface similarity between the dialects strengthened representations. When the model was further fine-tuned for part-of-speech tagging, its performance was robust to dialectical variation, even when trained solely on part-of-speech data from a single dialect. Our findings suggest that large multilingual models minimize the need for spelling normalization during pre-processing.

The paper presents new causal commonsense reasoning datasets for South Slavic dialects, based on the Choice of Plausible Alternatives (COPA) dataset. The dialectal datasets are built by translating by native dialect speakers from the English original and the corresponding standard translation. Three dialects are covered – the Cerkno dialect of Slovenian, the Chakavian dialect of Croatian and the Torlak dialect of Serbian. The datasets are the first resource for evaluation of large language models on South Slavic dialects, as well as among the first commonsense reasoning datasets on dialects overall. The paper describes specific challenges met during the translation process. A comparison of the dialectal datasets with their standard language counterparts shows a varying level of character-level, word-level and lexicon-level deviation of dialectal text from the standard datasets. The observed differences are well reproduced in initial zero-shot and 10-shot experiments, where the Slovenian Cerkno dialect and the Croatian Chakavian dialect show significantly lower results than the Torlak dialect. These results show also for the dialectal datasets to be significantly more challenging than the standard datasets. Finally, in-context learning on just 10 examples shows to improve the results dramatically, especially for the dialects with the lowest results.

This paper aims to assess the role of multiword compound adverbs in distinguishing Brazilian Portuguese (PT-BR) from European Portuguese (PT-PT). Two key factors underpin this focus: Firstly, multiword expressions often provide less ambiguity compared to single words, even when their meaning is idiomatic (non-compositional). Secondly, despite constituting a significant portion of lexicons in many languages, they are frequently overlooked in Natural Language Processing, possibly due to their heterogeneous nature and lexical range.For this study, a large lexicon of Portuguese multiword adverbs (3,665) annotated with diatopic information regarding language variety was utilized. The paper investigates the distribution of this category in a corpus consisting in excerpts from journalistic texts sourced from the DSL (Dialect and Similar Language) corpus, representing Brazilian (PT-BR) and European Portuguese (PT-PT), respectively, each partition containing 18,000 sentences.Results indicate a substantial similarity between the two varieties, with a considerable overlap in the lexicon of multiword adverbs. Additionally, specific adverbs unique to each language variety were identified. Lexical entries recognized in the corpus represent 18.2% (PT-BR) to 19.5% (PT-PT) of the lexicon, and approximately 5,700 matches in each partition. While many of the matches are spurious due to ambiguity with otherwise non-idiomatic, free strings, occurrences of adverbs marked as exclusive to one variety in texts from the other variety are rare.

This paper presents a new textual resource for Norwegian and its dialects. The NoMusic corpus contains Norwegian translations of the xSID dataset, an evaluation dataset for spoken language understanding (slot and intent detection). The translations cover Norwegian Bokmål, as well as eight dialects from three of the four major Norwegian dialect areas. To our knowledge, this is the first multi-parallel resource for written Norwegian dialects, and the first evaluation dataset for slot and intent detection focusing on non-standard Norwegian varieties. In this paper, we describe the annotation process and provide some analyses on the types of linguistic variation that can be found in the dataset.

Text summarization models have typically focused on optimizing aspects of quality such as fluency, relevance, and coherence, particularly in the context of news articles. However, summarization models are increasingly being used to summarize diverse sources of text, such as social media data, that encompass a wide demographic user base. It is thus crucial to assess not only the quality of the generated summaries, but also the extent to which they can fairly represent the opinions of diverse social groups. Position bias, a long-known issue in news summarization, has received limited attention in the context of social multi-document summarization. We deeply investigate this phenomenon by analyzing the effect of group ordering in input documents when summarizing tweets from three distinct linguistic communities: African-American English, Hispanic-aligned Language, and White-aligned Language. Our empirical analysis shows that although the textual quality of the summaries remains consistent regardless of the input document order, in terms of fairness, the results vary significantly depending on how the dialect groups are presented in the input data. Our results suggest that position bias manifests differently in social multi-document summarization, severely impacting the fairness of summarization models.

While Large Language Models (LLMs) have demonstrated considerable potential in advancing natural language processing in dialect-specific contexts, their effectiveness in these settings has yet to be thoroughly assessed. This study introduces a case study on Šariš, a dialect of Slovak, which is itself a language with fewer resources, focusing on Machine Translation and Common Sense Reasoning tasks. We employ LLMs in a zero-shot configuration and for data augmentation to refine Slovak-Šariš and Šariš-Slovak translation models. The accuracy of these models is then manually verified by native speakers. Additionally, we introduce ŠarišCOPA, a new dataset for causal common sense reasoning, which, alongside SlovakCOPA, serves to evaluate LLM’s performance in a zero-shot framework. Our findings highlight LLM’s capabilities in processing low-resource dialects and suggest a viable approach for initiating dialect-specific translation models in such contexts.

Linguistic variation is a complicating factor for digital language technologies. This is particularly true for languages that lack an official “standard” variety, including many regional and minoritized languages. In this paper, we describe a set of experiments focused on multivariant natural language processing for the Nahuatl, an indigenous Mexican language with a high level of linguistic variation and no single recognized standard variant. Using small (10k tokens), recently-published annotated datasets for two Nahuatl variants, we compare the performance of single-variant, cross-variant, and joint training, and explore how different models perform on a third Nahuatl variant, unseen in training. These results and the subsequent discussion contribute to efforts of developing low-resource NLP that is robust to diatopic variation. We share all code used to process the data and run the experiments.

We study highly granular dialect normalization and phonological dialect translation on Limburgish, a non-standardized low-resource language with a wide variation in spelling conventions and phonology. We find improvements to the traditional transformer by embedding the geographic coordinates of dialects in dialect normalization tasks and use these geographically-embedded transformers to translate words between the phonologies of different dialects. These results are found to be consistent with notions in traditional Limburgish dialectology.

Code-switching research depends on fine-grained language identification. In this work, we study existing corpora used to train token-level language identification systems. We aggregate these corpora with a consistent labelling scheme and train a system to identify English code-switching in multilingual text. We show that the system identifies code-switching in unseen language pairs with absolute measure 2.3-4.6% better than language-pair-specific SoTA. We also analyse the correlation between typological similarity of the languages and difficulty in recognizing code-switching.

This paper discusses the re-usibility of existing approaches, tools and automatic techniques for the annotation and detection of events in a challenging variant of centuries old Dutch written in the archives of the Dutch East India Company. We describe our annotation process and provide a thorough analysis of different versions of manually annotated data and the first automatic results from two fine-tuned Language Models. Through the analysis of this complete process, the paper studies two things: to what extent we can use NLP theories and tasks formulated for modern English to formulate an annotation task for Early Modern Dutch and to what extent we can use NLP models and tools built for modern Dutch (and other languages) on Early Modern Dutch. We believe these analyses give us insight into how to deal with the large variation language showcases in describing events, and how this variation may differ accross domains. We release the annotation guidelines, annotated data, and code.

Chavacano is a Spanish Creole widely spoken in the southern regions of the Philippines. It is one of the many Philippine languages yet to be studied computationally. This paper presents the development of a language identification model of Chavacano to distinguish it from languages that influence its creolization using character convolutional networks. Unlike studies that discriminated similar languages based on geographical proximity, this paper reports a similarity focused on the creolization of a language. We established the similarity of Chavacano and its related languages, Spanish, Portuguese, Cebuano, and Hiligaynon, from the number of common words in the corpus for all languages. We report an accuracy of 93% for the model generated using ten filters with a filter width of 5. The training experiments reveal that increasing the filter width, number of filters, or training epochs is unnecessary even if the accuracy increases because the generated models present irregular learning behavior or may have already been overfitted. This study also demonstrates that the character features extracted from convolutional neural networks, similar to n-grams, are sufficient in identifying Chavacano. Future work on the language identification of Chavacano includes improving classification accuracy for short or code-switched texts for practical applications such as social media sensors for disaster response and management.

This report presents gmnlp’s participation to the Dialect-Copa shared task at VarDial 2024 (Chifu et al., 2024), which focuses on evaluating the commonsense reasoning capabilities of large language models (LLMs) on South Slavic micro-dialects. The task aims to assess how well LLMs can handle non-standard dialectal varieties, as their performance on standard languages is already well-established. We propose an approach that combines the strengths of different types of language models and leverages data augmentation techniques to improve task performance on three South Slavic dialects: Chakavian, Cherkano, and Torlak. We conduct experiments using a language-family-focused encoder-based model (BERTić) and a domain-agnostic multilingual model (AYA-101). Our results demonstrate that the proposed data augmentation techniques lead to substantial performance gains across all three test datasets in the open-source model category. This work highlights the practical utility of data augmentation and the potential of LLMs in handling non-standard dialectal varieties, contributing to the broader goal of advancing natural language understanding in low-resource and dialectal settings.

The paper presents the JSI and WüNLP systems submitted to the DIALECT-COPA shared task on causal commonsense reasoning in dialectal texts. Jointly, we compare LLM-based zero-shot and few-shot in-context inference (JSI team), and task-specific few-shot fine-tuning, in English and respective standard language, with zero-shot cross-lingual transfer (ZS-XLT) to the test dialects (WüNLP team). Given the very strong zero-shot and especially few-shot in-context learning (ICL) performance, we further investigate whether task semantics, or language/dialect semantics explain the strong performance, showing that a significant part of the improvement indeed stems from learning the language or dialect semantics from the in-context examples, with only a minor contribution from understanding the nature of the task. The higher importance of the dialect semantics to the task semantics is further shown by the finding that the in-context learning with only a few dialectal instances achieves comparable results to the supervised fine-tuning approach on hundreds of instances in standard language.

The choice of plausible alternatives (COPA) task requires selecting the most plausible outcome from two choices based on understanding the causal relationships presented in a given text.This paper outlines several approaches and model adaptation strategies to the VarDial 2024 DIALECT-COPA shared task, focusing on causal commonsense reasoning in South-Slavic dialects. We utilize and evaluate the GPT-4 model in combination with various prompts engineering and the Retrieval-Augmented Generation (RAG) technique. Initially, we test and compare the performance of GPT-4 with simple and advanced prompts on the COPA task across three dialects: Cerkno, Chakavian and Torlak. Next, we enhance prompts using the RAG technique specifically for the Chakavian and Cerkno dialect. This involves creating an extended Chakavian-English and Cerkno-Slovene lexical dictionary and integrating it into the prompts. Our findings indicate that the most complex approach, which combines an advanced prompt with an injected dictionary, yields the highest performance on the DIALECT-COPA task.

We present a one-shot prompting approach to multi-class classification for similar language identification with off-the-shelf pre-trained large language model that is not particularly trained or tuned for the language identification task. Without post-training or fine-tuning the model, we simply include one example per class when prompting the model and surprisingly the model to generate the language andlocale labels accordingly.

The VLP team participated in the DSL-ML shared task of the VarDial 2024 workshop which aims to distinguish texts in similar languages. This paper presents our approach to solving the problem and discusses our experimental and official results. We propose to integrate semantics-aware word embeddings which are learned from ConceptNet into a bidirectional long short-term memory network. This approach achieves good performance – our sys- tem is ranked in the top two or three of the best performing teams for the task.

This paper describes the Brandeis University submission to VarDial 2024 DSL-ML Shared Task on multilabel classification for discriminating between similar languages. Our submission consists of three entries per language to the closed track, where no additional data was permitted. Our approach involves a set of simple non-neural baselines using logistic regression, random forests and support vector machines. We follow this by experimenting with finetuning multilingual BERT, either on a single language or all the languages concatenated together.In addition to benchmarking the model architectures against one another on the development set, we perform extensive hyperparameter tuning, which is afforded by the small size of the training data.Our experiments on the development set suggest that finetuned mBERT systems significantly benefit most languages compared to the baseline.However, on the test set, our results indicate that simple models based on scikit-learn can perform surprisingly well and even outperform pretrained language models, as we see with BCMS.Our submissions achieve the best performance on all languages as reported by the organizers. Except for Spanish and French, our non-neural baseline also ranks in the top 3 for all other languages.