TL;DR
This paper argues that the field of Natural Language Processing (NLP) is plagued by persistent myths about writing systems, particularly the conflation of language and writing, and the misconception that East Asian scripts are primarily ideographic. By debunking these myths, the authors aim to foster a more accurate understanding of language and improve NLP research.
What Problem Does This Paper Address?
The authors observe that much of NLP research, despite its focus on written text, operates under a set of misconceptions about writing systems. These myths, often implicitly endorsed, include:
- Conflating Language and Writing: Treating a language and its writing system as interchangeable entities. For example, describing Arabic as a "right-to-left language" when it's the script that is written right-to-left.
- The "Ideographic" Myth: The widespread belief that writing systems like Chinese, Japanese, and Korean are primarily "ideographic" (representing ideas directly) rather than being based on linguistic units like syllables or morphemes. This misunderstanding can lead to flawed assumptions in NLP models.
These misconceptions, the authors contend, stem from a lack of explicit training in linguistics and writing system typologies within the NLP community, leading to the propagation of inaccurate terminology and a potentially skewed understanding of language itself.
Methods
The paper employs a two-pronged approach:
- Conceptual Argumentation: The authors build a theoretical case by defining language as a cognitive ability and writing as a technology that records language. They then analyze the linguistic basis of various writing systems, including Arabic, Chinese, Japanese, and Korean, to demonstrate their phonological or morphological underpinnings.
- Empirical Survey: To quantify the prevalence of these myths, the authors conducted a survey of the ACL Anthology and Google Scholar. They searched for terms like "ideograph," "idiograph," and "ideographic" in speech and language processing papers published between 2003 and 2022. They manually coded 100 randomly selected papers to identify:
- The languages/writing systems being discussed.
- Whether language and writing were conflated.
- The apparent reason for mentioning "ideography."
Datasets
The paper does not introduce new datasets. Instead, it uses the following as sources for its analysis:
- ACL Anthology: A repository of papers in computational linguistics.
- Google Scholar: A broad academic search engine used to find speech processing research.
The authors selected 50 papers from each source for their survey.
Key Insights and Ideas
- Writing is a Technology, Not a Cognitive Ability: Unlike language, which is an innate human capacity, writing is a learned technology that represents spoken language.
- The "Right-to-Left" Myth is Script-Specific: The direction of writing (e.g., right-to-left for Arabic) is a property of the script, not the language itself.
- Chinese is Morpho-Syllabic, Not Ideographic: Chinese characters primarily represent syllables and morphemes. While some characters are semantic-phonetic compounds, they still rely on phonology for pronunciation hints.
- Japanese and Korean Writing Systems are Primarily Phonological/Morphological: While influenced by Chinese characters, Japanese and Korean writing systems (like hiragana, katakana, and hangul) are largely based on phonetic representation of syllables or morphemes.
- The Term "Ideographic" is Often Used as a Placeholder: The survey revealed that many NLP papers use "ideographic" simply to refer to East Asian characters without a deep understanding of their linguistic basis. It's often used as a way to describe or introduce these characters.
- Conflation of Language and Writing is Widespread: A significant portion of the surveyed papers incorrectly described languages (like Chinese) as "ideographic languages."
Key Results
The survey of 100 papers revealed:
- Chinese and Japanese were the most frequently described as "ideographic" (31 and 20 papers from the Anthology, respectively).
- 23 out of 100 papers explicitly conflated writing and language by describing languages as ideographic.
- 69 out of 100 papers used the term "ideography" to describe or introduce Han characters in Chinese, Japanese, and Korean.
- While some symbols like '$' and Arabic numerals were correctly identified as ideographic, the term was often misapplied to entire writing systems.
Limitations and Future Work
- Unicode Consortium's Role: The authors suggest that the Unicode Consortium should reconsider its use of the term "CJK Unified Ideographs," as it perpetuates the myth. They propose alternative terms like "sinographs."
- Editor/Reviewer Responsibility: The paper calls for editors and reviewers to be more vigilant in correcting the misuse of terminology related to writing systems.
- Need for Better Education: The authors acknowledge that many NLP researchers may lack the necessary background in writing systems and recommend specific texts (Rogers, 2005; Gnanadesikan, 2009) for further learning.
- Future Work: The paper implicitly suggests that NLP research should move towards a more accurate understanding of writing systems, potentially leading to more robust and linguistically informed models.
Summary updated