Proceedings of the 9th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature (LaTeCH-CLfL 2025)

Anna Kazantseva, Stan Szpakowicz, Stefania Degaetano-Ortlieb, Yuri Bizzoni, Janis PagelNAACL 2025

This document is a collection of papers from a workshop, not a single research paper. Therefore, I cannot provide a summary of a specific paper's methods, datasets, results, or limitations.

However, I can provide a summary of the workshop's overall themes and contributions based on the provided table of contents and introduction.

TL;DR

The LaTeCH-CLfL 2025 workshop brings together researchers applying computational linguistics and machine learning to the humanities, cultural heritage, social sciences, and literature. Key themes include leveraging large language models (LLMs) for literary analysis, historical text processing, sociopolitical text analysis, and the computational study of literature. The workshop highlights the growing synergy between advanced AI techniques and traditional humanistic inquiry.

What Problem Does This Paper Address?

The workshop as a whole addresses the challenge of bridging the gap between computational linguistics and the humanities. This involves developing and applying advanced computational methods to understand and analyze complex humanistic data, such as historical texts, literary works, and social discourse. Specific problems tackled include:

  • Analyzing Literary Texts: Developing computational methods for tasks like scene segmentation, identifying relationships in fiction, and generating poetry.
  • Processing Historical and Low-Resource Languages: Overcoming challenges in working with historical documents and languages with limited digital resources, including tasks like transliteration, named entity recognition, and genre classification.
  • Sociopolitical Text Analysis: Using computational linguistics to detect and understand social issues like bias, propaganda, and hate speech in various forms of text.
  • Leveraging Large Language Models (LLMs): Exploring innovative ways to use LLMs for tasks ranging from literary analysis and sequence labeling to translation and conversational AI.
  • Bridging Disciplines: Fostering dialogue and collaboration between computer scientists and humanists to advance research in both fields.

Methods

The workshop features a diverse range of methods, reflecting the interdisciplinary nature of the field. Common approaches include:

  • Large Language Models (LLMs): Extensive use of LLMs for various tasks, including:
  • Prompting for sequence labeling and named entity recognition.
  • Zero-shot learning for historical text analysis.
  • Generating text (e.g., poetry).
  • Adapting LLMs for specific domains (e.g., historical texts, low-resource languages).
  • Using LLMs for conversational AI and surveys.
  • Deep Learning Architectures: Application of neural networks, including character-level tokenization for poetry generation and deep neural architectures for cultural studies.
  • Natural Language Processing (NLP) Techniques: Standard NLP tasks such as:
  • Named Entity Recognition (NER).
  • Text classification (e.g., bias detection, genre classification, hate speech classification).
  • Text segmentation.
  • Word embeddings and their analysis.
  • Machine translation.
  • Quantitative and Statistical Methods: Employing quantitative approaches for analyzing literary dynamics, psychological modeling, and sociolinguistic variation.
  • Dataset Creation and Annotation: Developing new datasets for specific tasks, such as artificial relationships in fiction and comparative statements.

Datasets

The papers in the workshop utilize a variety of datasets, including:

  • Historical Texts: Swedish encyclopedias, Comorian language data, Egyptian hieroglyphs, historical Luxembourgish texts, historical magazines (1875-1990), and ancient texts.
  • Literary Texts: Novels, song lyrics, and fictional works.
  • Social Media Data: Instagram posts.
  • News Media: Russian media.
  • General Language Datasets: Used for pretraining or as baselines.

The workshop also features the introduction of new benchmarks and datasets, such as:

  • A dataset for advancing NLP in literary domains focusing on artificial relationships in fiction.
  • A dataset for annotation and classification of comparative statements.

Key Insights and Ideas

  • LLMs as powerful tools for humanistic inquiry: LLMs are proving to be versatile tools that can be adapted to tackle complex questions in literature, history, and social sciences, offering new perspectives and analytical capabilities.
  • The importance of domain adaptation for LLMs: Pretrained LLMs often require significant adaptation and fine-tuning to perform well on specialized tasks within the humanities, especially for historical or low-resource languages.
  • Computational methods can reveal subtle social phenomena: Techniques like bias detection and propaganda analysis can uncover hidden patterns and strategies in text that might be missed by traditional methods.
  • Bridging the gap between quantitative and qualitative analysis: The workshop emphasizes the convergence of large-scale quantitative models with deep scholarly traditions, enriching our understanding of human language and culture.
  • The need for interpretable AI in the humanities: As computational methods become more sophisticated, there is a growing emphasis on developing interpretable models that can provide insights understandable to humanists.

Key Results

Specific quantitative and qualitative results are not available without summarizing individual papers. However, the workshop's contributions generally demonstrate:

  • Improved performance on various NLP tasks: Papers report advancements in tasks like named entity recognition, text classification, and language generation when applying novel computational methods to humanistic data.
  • Successful adaptation of LLMs: Demonstrations of LLMs being effectively used for tasks they were not explicitly trained for, particularly in specialized domains.
  • New insights into cultural and social phenomena: Computational analyses yielding new understandings of literary dynamics, historical language use, and sociopolitical discourse.

Limitations and Future Work

Limitations and future work are discussed within individual papers. However, overarching themes likely include:

  • Data scarcity for low-resource and historical languages: A persistent challenge is the lack of sufficient annotated data for many languages and historical periods.
  • Interpretability of complex models: Understanding why LLMs and deep learning models make certain predictions remains a significant research area.
  • Ethical considerations: Addressing issues of bias, fairness, and responsible use of AI in analyzing sensitive social and cultural data.
  • Developing robust evaluation metrics: Creating appropriate metrics to evaluate computational models on tasks that are inherently subjective or qualitative in the humanities.
  • Further integration of computational and humanistic expertise: Continued efforts to foster interdisciplinary collaboration and develop methodologies that truly bridge the two fields.

Summary updated

BibTeX
@inproceedings{kazantseva2025proceedings,
  author = {Anna Kazantseva and Stan Szpakowicz and Stefania Degaetano-Ortlieb and Yuri Bizzoni and Janis Pagel},
  title = {{Proceedings of the 9th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature (LaTeCH-CLfL 2025)}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
}
Figure