NT2Lex: A CEFR-Graded Lexical Resource for Dutch as a Foreign Language Linked to Open Dutch WordNet

Anaïs Tack, Thomas François, Piet Desmet, Cédrick FaironNAACL 2018

TL;DR

This paper introduces NT2Lex, a novel lexical resource for Dutch as a foreign language, graded by CEFR proficiency levels and linked to Open Dutch WordNet. This resource offers a unique insight into word sense complexity in L2 learning by providing graded frequency distributions for word senses, going beyond traditional vocabulary size measures.

What Problem Does This Paper Address?

Understanding vocabulary is crucial for second language (L2) learners to comprehend texts. While existing graded lexicons help identify vocabulary appropriate for different proficiency levels (e.g., CEFR A1-C2), they often focus on word frequency and breadth of knowledge. However, they largely overlook the complexity of word senses. A single word can have multiple meanings (polysemy), and different words can express similar meanings (synonymy). This "depth" of vocabulary knowledge, specifically understanding different word senses, is vital for true comprehension but has been under-researched in L2 learning resources. This paper aims to address this gap by creating a resource that accounts for word sense complexity in graded Dutch L2 vocabulary.

Methods

The authors developed NT2Lex by processing a large corpus of CEFR-graded Dutch L2 reading materials. Key methodological steps include:

  • Data Preprocessing: Manual correction of OCR errors, removal of tonic diacritics, lemmatization, and part-of-speech tagging. Multi-word units were automatically identified.
  • Word Sense Disambiguation (WSD): A WSD tool was applied to the tagged texts to assign specific word senses. This tool was trained on the DutchSemCor corpus and integrated with Open Dutch WordNet (ODWN) to leverage its synsets (sets of synonyms representing a concept).
  • Lexical Entry Definition: Non-alphanumeric entries, abbreviations, and certain compound word structures were simplified or split. The part-of-speech tagset was also simplified for the resource.
  • Graded Frequency Calculation: Raw frequencies of lexical entries were computed for each CEFR level. To account for lexical dispersion (how widely a word is used across documents), adjusted frequencies were calculated using formulas adapted from Carroll et al. (1971). This provides a more nuanced measure of word familiarity and difficulty.
  • Linkage to Open Dutch WordNet (ODWN): Lexical entries were linked to ODWN synset IDs, enabling the analysis of word sense complexity (polysemy and synonymy).

Datasets

  • Corpus: A 461,088-token corpus of CEFR-graded readers and textbooks for Dutch as a foreign language, spanning CEFR levels A1 to C1. This corpus includes both Netherlandic and Belgian Dutch texts.
  • Open Dutch WordNet (ODWN): Used for word sense disambiguation and linking lexical entries to semantic concepts.
  • DutchSemCor: Used to train the word-sense disambiguation tool.
  • Basiswoordenboek Nederlands: A list of the 2,000 most frequent Dutch words, used for comparing lexical sophistication.
  • Brysbaert et al. (2014) lexical database: Used for norms of age of acquisition and concreteness.

Key Insights and Ideas

  • Word Sense Complexity Matters: The paper strongly advocates for the inclusion of word sense information in L2 vocabulary resources, arguing it provides a deeper understanding of lexical complexity beyond simple frequency.
  • Linkage to WordNet is Key: Connecting graded lexical entries to a semantic network like WordNet is crucial for analyzing word sense complexity and ensuring the resource's semantic consistency.
  • Dispersion as a Proxy for Familiarity: The authors propose using lexical dispersion as a measure of theoretical word familiarity in L2 learners, as a lack of direct experimental data on Dutch L2 word familiarity exists.
  • Early Levels are More Polysemous: Lexical entries in the more basic CEFR levels (A1/A2) tend to be more polysemous and acquired earlier by learners, suggesting a foundational vocabulary with broader, less specialized meanings.
  • Advanced Levels are More Abstract: Higher CEFR levels (B2/C1) show a greater proportion of abstract words, indicating a shift towards more nuanced and less concrete vocabulary as proficiency increases.

Key Results

  • NT2Lex Resource: The resource contains 17,743 entries in its most complete version (NT2Lex-CGN+ODWN), with 11,999 unique word senses and 8,934 distinct synsets.
  • Frequency Coherence: The adjusted frequency distributions in NT2Lex show strong coherence with established frequency lists like Subtlex-NL (Pearson correlation r = .69), indicating the resource's reliability.
  • Dispersion and Familiarity: Theoretical word familiarity (measured by dispersion) accounts for approximately 83% of the distribution of adjusted frequencies.
  • Lexical Sophistication: A negative correlation exists between word dispersion and lexical sophistication, meaning more dispersed words (more familiar) are generally less sophisticated.
  • Polysemy and Synonymy: The degree of polysemy and synonymy in NT2Lex is strongly correlated with ODWN, suggesting semantic consistency with general language networks.
  • Age of Acquisition and Concreteness: Words acquired earlier by native speakers are generally more concrete and more dispersed in the corpus. Higher CEFR levels show a higher proportion of abstract words.

Limitations and Future Work

  • Expert Knowledge Reliance: The assumptions drawn about vocabulary understanding are primarily based on expert knowledge from pedagogical texts, rather than extensive experimental data from actual L2 learners.
  • Learner Data Contrast: The authors aim to contrast the knowledge gained from NT2Lex with actual receptive learner data in the future to validate the resource's assumptions.
  • C1 Subcorpus Size: The C1 subcorpus is noted as being restricted in size, potentially affecting the reliability of some statistics at that level.
  • Specialized Nature of L2 Lexicalization: The lower extent of onomasiological variation in NT2Lex compared to ODWN might indicate a specialized nature of the resource, reflecting L2-specific lexical choices.

Summary updated

BibTeX
@inproceedings{tack2018nt,
  author = {Anaïs Tack and Thomas François and Piet Desmet and Cédrick Fairon},
  title = {{NT2Lex: A CEFR-Graded Lexical Resource for Dutch as a Foreign Language Linked to Open Dutch WordNet}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2018},
}
Figure