TL;DR
This paper presents the proceedings of the 1st International Workshop on Computational Approaches to Historical Language Change (LChange'19). It highlights the growing intersection of computational linguistics, historical linguistics, and digital humanities, showcasing research on modeling and detecting language evolution using computational methods and large-scale corpora. The workshop aimed to foster interdisciplinary collaboration and address key challenges in the field.
What Problem Does This Paper Address?
The paper addresses the challenge of understanding and modeling how human language changes over time. While historical linguistics has long studied these phenomena, the advent of large-scale digital corpora has opened new avenues for computational approaches. These approaches aim to detect semantic and lexical changes, understand their drivers, and apply these insights to various fields like digital humanities and information retrieval. The workshop specifically sought to bridge the gap between computational methods and traditional linguistic theories, and to address methodological challenges such as data representativeness and evaluation of computational models.
Methods
The workshop proceedings cover a wide range of computational methods applied to historical language change. These include:
- Neural Embeddings: A significant portion of the research utilizes word embeddings (both standard and multi-sense) to capture semantic shifts over time.
- Diachronic Text Mining: Techniques for analyzing large historical corpora to identify patterns of change.
- Computational Linguistics Models: Development of models for detecting semantic change, lexical replacement, and even phonetic changes.
- Natural Language Processing (NLP) Tools: Application of NLP tools like part-of-speech taggers for syntactic change detection.
- Visualization Techniques: Methods for visualizing linguistic change as dimension interactions.
- Machine Learning Approaches: Including RNNs for proto-language reconstruction and Gaussian Process Models for dialectology.
Datasets
The paper does not explicitly list specific datasets used across all papers. However, it mentions the reliance on:
- Large-scale digital corpora: These are crucial for training and evaluating computational models of language change.
- Historical texts: Including poetry, prose, parliamentary debates, ancient devotional texts, and historical newspapers.
- Diachronic embedding projections: Used for tracing language evolution.
Key Insights and Ideas
- The power of embeddings: Neural embeddings have become a dominant tool for studying semantic change, but limitations exist regarding multi-sense representation and applicability to resource-poor languages.
- Interdisciplinary collaboration is key: Bridging the gap between computational linguists, historical linguists, sociolinguists, and digital humanities researchers is essential for advancing the field.
- Methodological challenges remain: Issues of data representativeness, evaluation of computational models, and the distinction between genuine language change and corpus artifacts are critical areas for future research.
- Language change has practical applications: Understanding historical language change is valuable for tasks like information retrieval, temporal analysis of documents, and even OCR algorithm design.
- The need for resource-poor language solutions: Developing methods that can handle languages with limited digital resources is an important future direction.
Key Results
The paper, being a collection of workshop proceedings, does not present a single set of quantitative results. Instead, it showcases a diverse range of research findings from individual papers. These include:
- Successful application of computational models to identify semantic bleaching of English intensifiers.
- Analysis of historical changes in poetry and prose using computational methods.
- Studies on semantic chain shifts and the evaluation of semantic change in psychology.
- Development of methods for automatically identifying diachronic variation in collocations.
- Computational evidence for the era of authorship of Old Thai prose.
- Investigations into gender bias in Dutch newspapers using word embeddings.
- Predictions of historical phonetic features using deep neural networks.
- Visualizations of linguistic change as dimension interactions.
Limitations and Future Work
The paper acknowledges several limitations and points towards future work:
- Data representativeness: The extent to which results derived from specific corpora (e.g., digital literature, newspapers) accurately represent language change as a whole is still an open question.
- Evaluation of models: Verifying language change solely through positive examples can lead to corpus bias. Developing robust evaluation test sets and standards is crucial.
- Resource-poor languages: Current embedding-based methods often require massive datasets, limiting their applicability to languages with fewer digital resources. Future work should focus on dynamic embeddings and methods that can discriminate and disambiguate word senses with less data.
- Understanding the drivers of change: While computational methods can detect change, understanding the underlying sociocultural and technological factors driving these changes remains a complex area.
- User interfaces for exploration: Developing user-friendly systems for exploring language changes in corpora is needed to facilitate serendipitous discovery.
Summary updated