Abstract
In this paper we take into account both social and linguistic aspects to perform demographic analysis by processing a large amount of tweets in Basque language. The study of demographic characteristics and social relationships are approached by applying machine learning and modern deep-learning Natural Language Processing (NLP) techniques, combining social sciences with automatic text processing. More specifically, our main objective is to combine demographic inference and social analysis in order to detect young Basque Twitter users and to identify the communities that arise from their relationships or shared content. This social and demographic analysis will be entirely based on the automatically collected tweets using NLP to convert unstructured textual information into interpretable knowledge.
Introduction
Basque is a "low-resource" language—while it has official status, its presence in traditional media is often overshadowed by Spanish. Social media, particularly Twitter, has emerged as a crucial "digital hearth" for Basque speakers, providing a space for linguistic expression and community building. However, performing demographic research on these users is challenging because metadata, such as age, is rarely public.
This research addresses two primary objectives:
- Demographic Inference: Identifying "young" users, who represent the future of the language, without relying on manual, small-scale surveys.
- Social Mapping: Identifying the specific communities that arise from the relationships and shared content of these young users.
Methodology
The authors propose a multi-stage approach that integrates traditional sociolinguistics with modern deep learning. A central challenge in demographic inference is the lack of labeled data. To overcome this, the researchers leveraged the sociolinguistic theory that youth language is typically more informal.
As shown in the pipeline above, an "informal/formal" classifier was used to rank users based on their "informal tweet concentration." The top and bottom of this ranked list were then manually verified to create a high-quality "gold standard" dataset for training. For the final classification task, the study compared three systems: IXA pipes (feature-engineered), mBERT (multilingual), and BERTeus (a monolingual BERT model specifically pre-trained on Basque).
Data
The study utilized the Heldugazte Corpus, a collection of 6 million Basque tweets. From this, the researchers derived the Heldugazte-Age dataset, containing 80,000 tweets from 1,000 users (500 young, 500 adult). This balanced and semi-automatically annotated dataset is a significant contribution to the field of Basque NLP.
The examples in the table highlight the distinct linguistic markers, such as the use of "jajaja" and non-standard spelling in the "young" category, versus the more formal, grammatically standard structures in the "adult" category.
Results and Analysis
The BERTeus model was the most effective, achieving an accuracy of 96.3% (F1-score of 0.963). This result underscores the importance of dedicated monolingual models for low-resource languages; despite its broader knowledge base, the multilingual mBERT lacked the specific nuance required for this demographic detection task.
Following identification, the researchers applied node2vec to a retweet graph of 1,635 young users to identify social clusters. The analysis revealed four distinct sub-communities:
- News (29.96%): Centered around regional media and journalists.
- Nationalist Left (26.98%): Focused on independentist politics and social organizations.
- Sports (22.58%): Dominated by football and cycling content.
- Music (20.49%): Centered around Basque-language bands and singers.
The most influential accounts were traditional media outlets like Berria and Argia, suggesting that Twitter serves as a primary news consumption tool for young Basque speakers.
Conclusion
This research demonstrates the power of combining NLP with sociolinguistic proxies to analyze digital communities. By successfully identifying young users and mapping their social interactions, the study provides valuable insights into the health and future of the Basque language online. Future work may expand this analysis to include more granular age spectrums and sentiment analysis within the identified communities.
Summary updated