Proceedings of the 3rd Workshop on Computational Approaches to Historical Language Change

Nina Tahmasebi, Syrielle Montariol, Andrey Kutuzov, Simon Hengchen, Haim Dubossarsky, Lars BorinACL 2022

TL;DR

This document presents the proceedings of the 3rd International Workshop on Computational Approaches to Historical Language Change (LChange 2022). It showcases research on using computational methods to understand how language evolves over time, with a particular focus on semantic change detection and the development of new benchmarks and methodologies for this complex task.

What Problem Does This Paper Address?

The core problem this workshop addresses is the challenge of computationally modeling and understanding historical language change. Specifically, it focuses on:

  • Semantic Change Detection: How can we automatically identify and track changes in word meanings over time? This is crucial for understanding linguistic evolution, historical linguistics, and the impact of societal changes on language.
  • Methodological Gaps: While computational approaches to language change are growing, there's a need for robust evaluation methods, standardized benchmarks, and a deeper understanding of the interplay between computational models and linguistic theory.
  • Data Availability: The "digital turn" has made vast historical text corpora available, creating an opportunity but also a need for tools to effectively analyze them for diachronic linguistic insights.
  • Bridging Disciplines: The workshop aims to foster collaboration between computational linguists, computer scientists, and humanities scholars to tackle these challenges.

Methods

The workshop proceedings highlight a variety of computational methods applied to historical language change, including:

  • Word Embeddings: Various techniques using word embeddings (e.g., dynamic Bernoulli word embeddings) are employed to capture semantic shifts.
  • Neural Topic Models: Used to track context shifts of words, particularly relevant for analyzing recent changes (e.g., COVID-related terms).
  • BERT and other Language Models: BERT models are used for tasks like predicting publication years of texts and for lexical semantic change detection.
  • Shared Task Frameworks: The workshop features a significant "LSCDISCOVERY SHARED TASK" focused on semantic change discovery and detection in Spanish, with multiple teams developing and evaluating their approaches. This task includes phases for graded and binary change detection.
  • Phonological Reconstruction: Novel frameworks are proposed for automated phonological reconstruction using trimmed alignments and sound correspondence patterns.
  • Cognitive Linguistics Approaches: Some papers integrate cognitive linguistics perspectives into computational analyses of diachronic change.

Datasets

The proceedings mention or imply the use of various datasets:

  • Historical Text Corpora: Large collections of historical texts are implicitly used across many papers, enabling the study of language evolution over extended periods.
  • Eighteenth Century Texts: Specifically mentioned for the task of predicting publication years using BERT.
  • COVID-related Terms: Used for a case study on tracking context shifts of words.
  • Spanish Corpora: Central to the "LSCDISCOVERY SHARED TASK" for semantic change discovery and detection.
  • Multilingual Datasets: Used for tasks involving cognates and borrowings.
  • No new benchmarks are explicitly introduced as a primary contribution of the proceedings themselves, but the shared task serves as a significant evaluation platform.

Key Insights and Ideas

  • The "Digital Turn" is a Game-Changer: The availability of large digital historical corpora is democratizing the study of language change, allowing for more data-driven and less ad-hoc analyses.
  • Semantic Change is Multifaceted: Understanding semantic change requires considering various dimensions, including structural and functional features, as well as external societal phenomena.
  • Evaluation is Paramount: Automatic methods for detecting semantic change are powerful but must be rigorously evaluated to avoid empirically inadequate generalizations. The development of benchmark datasets and standardized evaluation protocols is crucial.
  • Interdisciplinary Collaboration is Key: Tackling the complexities of historical language change requires a synergy between computational linguistics, computer science, and traditional humanities fields like historical linguistics and lexicography.
  • Context is King (Even Historically): Tracking how word meanings shift within their contexts over time is fundamental to understanding language evolution.

Key Results

The proceedings showcase a range of results from the presented papers and the shared task:

  • Shared Task Performance: Multiple teams participated in the LSCDISCOVERY SHARED TASK, developing various models for semantic change detection in Spanish. Specific quantitative results for each team are detailed within their respective papers.
  • Methodological Advancements: Papers propose new frameworks for phonological reconstruction, methods for measuring semantic change using word embeddings, and approaches for explainable AI in historical text analysis.
  • Qualitative Insights: Several papers offer qualitative analyses of semantic change in specific languages or historical periods, providing linguistic insights informed by computational methods. For instance, one paper explores olfactory situations over time, and another examines the semantic change of "Vaderland, Volk and Natie" in Dutch literature.

Limitations and Future Work

  • Bias in Automatic Methods: A significant limitation acknowledged is the potential for automatic methods to be biased, leading to missed changes or incorrect labels, which can then skew linguistic generalizations.
  • Need for Robust Evaluation: The lack of comprehensive evaluation of proposed models was a driving force behind some of the research presented, highlighting the ongoing need for better benchmarks and evaluation protocols.
  • Interpretability: While models can detect changes, making these detections interpretable and linguistically meaningful remains an area for future work.
  • Cross-Lingual and Multilingual Challenges: Extending methods to a wider range of languages and handling multilingual data presents ongoing challenges.
  • Integration of Linguistic Theory: Further work is needed to better integrate computational findings with established linguistic theories of language change.
  • Societal Impact: Exploring the broader implications and applications of understanding language change, such as its potential to "save lives" as suggested by a keynote title, is a promising avenue for future research.

Summary updated

BibTeX
@inproceedings{tahmasebi2022proceedings,
  author = {Nina Tahmasebi and Syrielle Montariol and Andrey Kutuzov and Simon Hengchen and Haim Dubossarsky and Lars Borin},
  title = {{Proceedings of the 3rd Workshop on Computational Approaches to Historical Language Change}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2022},
}
Figure