This report presents the results of the shared tasks organized as part of the VarDial Evaluation Campaign 2022. The campaign is part of the ninth workshop on Natural Language Processing (NLP) for Similar Languages, Varieties and Dialects (VarDial), co-located with COLING 2022. Three separate shared tasks were included this year: Identification of Languages and Dialects of Italy (ITDI), French Cross-Domain Dialect Identification (FDI), and Dialectal Extractive Question Answering (DialQA). All three tasks were organized for the first time this year.

This paper presents a new tweet-based approach in geolinguistic analysis which combines geolocation, user IDs and textual features in order to identify patterns of linguistic variation on a sub-city scale. Sub-city variations can be connected to social drivers and thus open new opportunities for understanding the mechanisms of language variation and change. However, measuring linguistic variation on these scales is challenging due to the lack of highly-spatially-resolved data as well as to the daily movement or users’ “mobility” inside cities which can obscure the relation between the social context and linguistic variation. Here we demonstrate how combining geolocation with user IDs and textual analysis of tweets can yield information about the linguistic profiles of the users, the social context associated with specific locations and their connection to linguistic variation. We apply our methodology to analyze dialects in Buenos Aires and find evidence of socially-driven variation. Our methods will contribute to the identification of sociolinguistic patterns inside cities, which are valuable in social sciences and social services.

We present dialectR, an open-source R package for performing quantitative analyses of dialects based on categorical measures of difference and on variants of edit distance. dialectR stands as one of the first programmable toolkits that may freely be combined and extended by users with further statistical procedures. We describe implementational details of the package, and provide two examples of its use: one performing analyses based on multidimensional scaling and hierarchical clustering on a dataset of Dutch dialects, and another showing how an approximation of the acoustic vowel space may be achieved by performing an MFCC (Mel-Frequency Cepstral Coefficients)-based acoustic distance on audio recordings of vowels.

The development of Natural Language Processing (NLP) applications for Cantonese, a language with over 85 million speakers, is lagging compared to other languages with a similar number of speakers. In this paper, we present, to our best knowledge, the first benchmark of multiple neural machine translation (NMT) systems from Mandarin Chinese to Cantonese. Additionally, we performed parallel sentence mining (PSM) as data augmentation for the extremely low resource language pair and increased the number of sentence pairs from 1,002 to 35,877. Results show that with PSM, the best performing model (BPE-level bidirectional LSTM) scored 11.98 BLEU better than the vanilla baseline and 9.93 BLEU higher than our strong baseline. Our unsupervised NMT (UNMT) results also refuted previous assumption n (Rubino et al., 2020) that the poor performance was related to the lack of linguistic similarities between the target and source languages, particularly in the case of Cantonese and Mandarin. In the process of building the NMT system, we also created the first large-scale parallel training and evaluation datasets of the language pair. Codes and datasets are publicly available at https://github.com/evelynkyl/yue_nmt.

In this paper, we propose a method to detect if words in two similar languages, Assamese and Bengali, are cognates. We mix phonetic, semantic, and articulatory features and use the cognate detection task to analyze the relative informational contribution of each type of feature to distinguish words in the two similar languages. In addition, since support for low-resourced languages like Assamese can be weak or nonexistent in some multilingual language models, we create a monolingual Assamese Transformer model and explore augmenting multilingual models with monolingual models using affine transformation techniques between vector spaces.

Closely related languages are often mutually intelligible to various degrees. Therefore, speakers of closely related languages are usually capable of (partially) comprehending each other’s speech without explicitly learning the target, second language. The cross-linguistic intelligibility among closely related languages is mainly driven by linguistic factors such as lexical similarities. This paper presents a computational model of spoken-word recognition and investigates its ability to recognize word forms from different languages than its native, training language. Our model is based on a recurrent neural network that learns to map a word’s phonological sequence onto a semantic representation of the word. Furthermore, we present a case study on the related Slavic languages and demonstrate that the cross-lingual performance of our model not only predicts mutual intelligibility to a large extent but also reflects the genetic classification of the languages in our study.

Norwegian Twitter data poses an interesting challenge for Natural Language Processing (NLP) tasks. These texts are difficult for models trained on standardized text in one of the two Norwegian written forms (Bokmål and Nynorsk), as they contain both the typical variation of social media text, as well as a large amount of dialectal variety. In this paper we present a novel Norwegian Twitter dataset annotated with POS-tags. We show that models trained on Universal Dependency (UD) data perform worse when evaluated against this dataset, and that models trained on Bokmål generally perform better than those trained on Nynorsk. We also see that performance on dialectal tweets is comparable to the written standards for some models. Finally we perform a detailed analysis of the errors that models commonly make on this data.

This paper presents OcWikiDisc, a new freely available corpus in Occitan, as well as language identification experiments on Occitan done as part of the corpus building process. Occitan is a regional language spoken mainly in the south of France and in parts of Spain and Italy. It exhibits rich diatopic variation, it is not standardized, and it is still low-resourced, especially when it comes to large downloadable corpora. We introduce OcWikiDisc, a corpus extracted from the talk pages associated with the Occitan Wikipedia. The version of the corpus with the most restrictive language filtering contains 8K user messages for a total of 618K tokens. The language filtering is performed based on language identification experiments with five off-the-shelf tools, including the new fasttext’s language identification model from Meta AI’s No Language Left Behind initiative, released in July 2022.

We present an approach to multi-class classification using an encoder-decoder transformer model. We trained a network to identify French varieties using the same scripts we use to train an encoder-decoder machine translation model. With some slight modification to the data preparation and inference parameters, we showed that the same tools used for machine translation can be easily re-used to achieve competitive performance for classification. On the French Dialectal Identification (FDI) task, we scored 32.4 on weighted F1, but this is far from a simple naive bayes classifier that outperforms a neural encoder-decoder model at 41.27 weighted F1.

Automatic Language Identification represents an important task for improving many real-world applications such as opinion mining and machine translation. In the case of closely-related languages such as regional dialects, this task is often challenging. In this paper, we propose an extensive evaluation of different approaches for the identification of Italian dialects and languages, spanning from classical machine learning models to more complex neural architectures and state-of-the-art pre-trained language models. Surprisingly, shallow machine learning models managed to outperform huge pre-trained language models in this specific task. This work was developed in the context of the Identification of Languages and Dialects of Italy (ITDI) task organised at VarDial 2022 Evaluation Campaign. Our best submission managed to achieve a weighted F1-score of 0.6880, ranking 5th out of 9 final submissions.

We present our contribution to the Identification of Languages and Dialects of Italy shared task (ITDI) proposed in the VarDial Evaluation Campaign 2022, which asked participants to automatically identify the language of a text associated to one of the language varieties of Italy. The method that yielded the best results in our experiments was a Deep Feedforward Neural Network (DNN) trained on character ngram counts, which provided a better performance compared to Naive Bayes methods and Convolutional Neural Networks (CNN). The system was among the best methods proposed for the ITDI shared task. The analysis of the results suggests that simple DNNs could be more efficient than CNNs to perform language identification of close varieties.

We describe the systems developed by the National Research Council Canada for the French Cross-Domain Dialect Identification shared task at the 2022 VarDial evaluation campaign. We evaluated two different approaches to this task: SVM and probabilistic classifiers exploiting n-grams as features, and trained from scratch on the data provided; and a pre-trained French language model, CamemBERT, that we fine-tuned on the dialect identification task. The latter method turned out to improve the macro-F1 score on the test set from 0.344 to 0.430 (25% increase), which indicates that transfer learning can be helpful for dialect identification.

This article describes the language identification approach used by the SUKI team in the Identification of Languages and Dialects of Italy and the French Cross-Domain Dialect Identification shared tasks organized as part of the VarDial workshop 2022. We describe some experiments and the preprocessing techniques we used for the training data in preparation for the shared task submissions, which are also discussed. Our Naive Bayes-based adaptive system reached the first position in Italian language identification and came second in the French variety identification task.