TL;DR
This document comprises the proceedings of the Workshop on Computation and Written Language (CAWL) 2023, highlighting the growing importance of understanding written language beyond its spoken counterpart in NLP. It showcases research on diverse topics, from the historical evolution of writing systems to modern challenges in processing various forms of written text, emphasizing the need for dedicated research communities within NLP.
What Problem Does This Paper Address?
The workshop addresses a critical gap in Natural Language Processing (NLP): the tendency to conflate written and spoken language, or to treat written language phenomena as mere precursors to core NLP tasks. This oversight leads to a misunderstanding of language properties, such as confusing a language's inherent structure with its writing system (e.g., calling Chinese "logographic" without acknowledging its phonetic components). Furthermore, research on crucial aspects of written language, like normalization, OCR, and text entry, often resides outside mainstream NLP venues, hindering the development of a cohesive research community focused on these areas. CAWL 2023 aims to foster such a community by bringing together researchers working on these often-neglected but fundamental aspects of language processing.
Methods
This document does not detail specific methods as it is a collection of proceedings from a workshop. The individual papers within the proceedings would describe their respective methodologies. However, the overarching theme suggests a range of computational linguistic and NLP techniques applied to various aspects of written language. This would likely include:
- Linguistic analysis: Examining the structural properties of writing systems and their relationship to spoken language.
- Machine learning and deep learning: For tasks such as character recognition, transliteration, speech recognition evaluation, and decipherment.
- Data annotation and guideline development: For creating specialized datasets for learner language and historical scripts.
- Algorithmic approaches: Such as coupled simulated annealing for script decipherment.
- Information theory: For quantifying linguistic features like logography.
Datasets
This document does not list specific datasets used for training and/or evaluation. The individual papers within the proceedings would detail their respective datasets. However, based on the paper titles, it can be inferred that datasets would include:
- Learner language transcripts: For studying the authenticity of handwritten learner language.
- Low-resource language data: Specifically for Maltese and Arabic, to evaluate transliteration impact.
- Romanized Hindi and Urdu corpora: For distinguishing between the two.
- Japanese Kanji data: For analyzing pronunciation ambiguities.
- Japanese speech recognition data: To model spelling inconsistencies.
- Ancient accounting corpora: For deciphering numeral sequences.
- Undeciphered script data: For learning character inventories.
- Japanese and Sumerian corpora: For quantifying logography.
Key Insights and Ideas
- The "perfect" alphabet is a myth: The original Semitic alphabet, while aiming for a 1-1 sound-letter mapping, has diverged significantly over time, with modern alphabets exhibiting complex letter-to-phoneme relationships.
- Writing systems impose linguistic structure: Writing is not merely a passive representation of speech; it actively structures language by segmenting speech streams and imposing rules (e.g., syllable well-formedness) that may not exist in the spoken form.
- Writing systems can encode linguistic features absent in spoken language: This includes graphic classifiers and inflectional morphology, which can be present in writing even when the spoken language lacks them.
- The distinction between written and spoken language is crucial for NLP: Conflating them leads to inaccuracies and overlooks important linguistic phenomena.
- Dedicated research on written language is vital: Many NLP methods rely heavily on text representations, yet research on written language processing is often scattered across different fields, necessitating a focused community within NLP.
Key Results
This document, being a collection of workshop proceedings, does not present a single set of quantitative results. Each paper within the proceedings would have its own specific findings. However, the titles and abstracts suggest a range of successful applications of computational methods to written language challenges, including:
- Improved understanding of writing system evolution and properties.
- Development of annotation guidelines for specialized language data.
- Evaluation of transliteration's impact on NLP performance for low-resource languages.
- Methods for distinguishing between similar written forms (e.g., Romanized Hindi/Urdu).
- Techniques for handling pronunciation ambiguities in writing systems.
- Lenient evaluation metrics for speech recognition that account for spelling inconsistencies.
- Novel approaches to deciphering ancient scripts and learning character inventories.
- Quantification of logography in different writing systems.
Limitations and Future Work
As this is a collection of workshop proceedings, specific limitations and future work are detailed within each individual paper. However, the overarching theme of the workshop implies several general limitations and directions:
- Limitations:
- The historical tendency to overlook or conflate written and spoken language in NLP research.
- The fragmentation of research on written language across different academic fields.
- The challenges associated with processing diverse and often noisy written data (e.g., learner language, historical scripts).
- The difficulty in evaluating NLP systems for written language tasks when standard metrics may not be appropriate.
- Future Work:
- Establishing and nurturing a dedicated research community within NLP focused on computation and written language.
- Developing more robust and nuanced NLP methods that account for the unique properties of written language.
- Creating new datasets and benchmarks for various written language processing tasks.
- Investigating the historical evolution and cross-linguistic variations of writing systems from a computational perspective.
- Exploring the interplay between writing systems and cognitive processes.
Summary updated