TL;DR
This document presents the proceedings of the 2nd International Workshop on Computational Approaches to Historical Language Change (LChange'21). It showcases research that uses computational methods to understand how languages evolve over time, addressing challenges in analyzing historical texts and modeling linguistic shifts. The workshop highlights the growing intersection of linguistics, computer science, and digital humanities in unraveling the complexities of language change.
What Problem Does This Paper Address?
Human language is a dynamic entity, constantly evolving. Understanding this evolution is crucial for various fields, from linguistics and history to computer science and digital humanities. However, studying historical language change presents significant challenges:
- Data Scarcity: Many languages lack extensive written records, making it difficult to trace their historical development. Even for languages with rich histories, the available texts are often fragmented or incomplete.
- Methodological Hurdles: Analyzing large historical corpora requires sophisticated computational tools. Traditional linguistic methods may not scale effectively, and new approaches are needed to handle the nuances of diachronic data.
- Representativeness of Corpora: Historical documents often survive by chance, leading to biased datasets. This makes it challenging to distinguish genuine language change from changes in the textual "habitat" or the way texts were produced and preserved.
- Quantifying Change: Accurately measuring and modeling language change is complex. Factors like sample size can significantly influence quantitative analyses, making it difficult to draw reliable conclusions.
This workshop aims to bring together researchers to address these challenges by developing and applying computational methodologies to the study of historical language change.
Methods
The papers presented at LChange'21 explore a variety of computational approaches to studying language change. While specific methods vary across individual papers, common themes include:
- Natural Language Processing (NLP) Techniques: Leveraging techniques like text translation, inference, semantic analysis, and topic modeling to extract information from historical texts.
- Corpus Linguistics: Building and analyzing specialized corpora of historical texts, often with a focus on specific languages, time periods, or domains (e.g., Italian newspapers, Old Literary Finnish, Ancient Greek).
- Data Visualization: Developing tools and methods to visualize linguistic data, such as dialect geography or the evolution of word senses.
- Machine Learning: Employing machine learning models for tasks like cognate prediction and event extraction from historical documents.
- Quantitative Analysis: Applying statistical methods and computational linguistics metrics to measure and model linguistic shifts.
- Dataset Creation: Developing new datasets and benchmarks for studying specific aspects of language change, such as diachronic semantic change datasets or datasets for historical text analysis.
Datasets
The workshop features research utilizing a range of datasets, including:
- Ancient Chinese Texts: For time-aware translation and inference.
- Russian Language Data: A three-part diachronic semantic change dataset for Russian is presented.
- Italian Newspapers: A case study on Italian newspapers is conducted.
- Old Literary Finnish: Research focuses on linguistic change and historical periodization.
- Scientific Topics: Datasets for studying the evolution of scientific topics and their relationships.
- Gender Asymmetry in Euphemism: Data for evaluating gender asymmetry in euphemism.
- Ancient Greek: The GLAUx corpus, designed for long-term, diverse, multi-layered analysis of Ancient Greek.
- South Asian Dialects: Data for visualizing dialect geography.
- Co-occurrence Networks: Used for modeling the evolution of word senses.
- English and Romance Languages: Cognate sets are analyzed for semantic change.
- Google Books Ngram Data: Used in analyses of diachronic corpora.
- German National Bibliography Data: Publicly available data used for corpus analysis.
- "Der Spiegel" Lexical Dynamics: A corpus of approximately 365,000 articles and 237,000,000 words from the German weekly news magazine published between 1947 and 2017.
- Black Rebellions: A new dataset for event extraction from historical texts related to Black Rebellions.
Key Insights and Ideas
- Linguistic Diversity as a Testing Ground: The workshop emphasizes that studying a wide range of languages, not just well-documented ones, is crucial for developing robust theories of semantic change and uncovering cross-linguistic regularities.
- Bridging Disciplines: The research highlights the growing synergy between computational linguistics, digital humanities, and traditional linguistic approaches, enabling new avenues of inquiry into language history.
- The Challenge of Diachronic Corpora: Researchers are grappling with fundamental issues in corpus representativeness, where the survival of historical documents is often accidental, making it difficult to isolate genuine language change from external factors.
- Sample Size Matters: Quantitative analyses of language change are highly sensitive to sample size, posing a significant challenge to accurate measurement and requiring novel approaches to mitigate this dependence.
- Visualizing Language Change: New visualization techniques are emerging to help researchers understand complex linguistic phenomena, such as the geographical distribution of dialects or the evolution of word meanings over time.
Key Results
The proceedings showcase a range of results, including:
- Development of new datasets for studying diachronic semantic change and historical event extraction.
- Demonstration of how computational methods can be used to analyze linguistic change in various languages and historical periods.
- Insights into the challenges of sample size dependence in quantitative linguistic analysis, with proposed methods to address it.
- Exploration of methods to disentangle actual language change from environmental changes in textual habitats.
- Presentations on the creation and methodological considerations of specialized corpora for historical language research.
Limitations and Future Work
- Data Availability: A recurring limitation is the scarcity of well-described historical data for many languages, hindering comprehensive diachronic studies.
- Corpus Representativeness: The inherent biases in historical document survival remain a significant challenge, requiring further research into methods for accounting for these biases.
- Methodological Refinements: Ongoing work is needed to refine computational methods for accurately quantifying and modeling language change, particularly in addressing sample size dependence and disentangling language evolution from external factors.
- Interdisciplinary Collaboration: Future work will likely involve deeper collaboration between computational linguists, domain experts in historical linguistics, and humanities scholars to tackle complex research questions.
- Development of New Tools and Resources: The creation of new datasets, software tools, and benchmarks will be crucial for advancing the field.
Summary updated