Colexifications for Bootstrapping Cross-lingual Datasets: The Case of Phonology, Concreteness, and Affectiveness

Yiyi Chen, Johannes BjervaACL 2023

TL;DR

This paper introduces a novel method for automatically creating cross-lingual datasets by leveraging "colexifications" – instances where a single word form in different languages expresses related concepts. They build a large-scale dataset linking word meanings (concreteness, affectiveness) to their phonetic properties across 142 languages, enabling new interdisciplinary research.

What Problem Does This Paper Address?

Researchers have long been interested in how word meanings relate to their sounds (phonology) and how these relationships vary across languages. While datasets exist for semantic properties like concreteness and affectiveness, they are often limited in language coverage and expensive to create, especially for low-resource languages. Similarly, understanding cross-lingual phonetic patterns is challenging due to the sheer diversity of languages. This paper tackles the problem of efficiently creating rich, cross-lingual datasets that bridge these semantic and phonetic domains, making them accessible for broader research.

Methods

The core of this work lies in a clever data curation pipeline. They start by building a "synset graph" from BabelNet, which represents concepts and their relationships. Then, they leverage "colexifications" – instances where different languages use the same word form for related concepts (e.g., "DAD" and "POPE" might share a word in some languages) – to link these concepts across languages.

Crucially, they then integrate two key types of data:

  1. Semantic Ratings: They incorporate existing ratings for concreteness (how tangible a concept is) and affectiveness (valence, arousal, dominance – how pleasant, intense, or controlling a concept feels) from established datasets.
  2. Phonetic Information: They extract phonemes and their corresponding phonological features (like whether a sound is voiced, nasal, etc.) from the WikiPron and PanPhon resources.

By aligning these semantic and phonetic data through colexifications, they create a unified cross-lingual dataset. They also develop algorithms to construct the colexification graph and map synsets to concepts, ensuring the most prevalent sense of a word is used for consistency.

Datasets

  • BabelNet: Used as the source for WordNet synsets and lemmas to build the initial concept graph.
  • WikiPron: Provided phonemic pronunciation data for a large number of languages.
  • PanPhon: Used to extract phonological features from phonemes.
  • Warriner et al. (2013) and Brysbaert et al. (2014): These datasets provide concreteness and affectiveness ratings for English words, which are then leveraged for cross-lingual bootstrapping.
  • Newly Introduced Benchmark: The paper presents a novel cross-lingual dataset covering 142 languages, 21 language families, and 5 language macro areas, linking semantic ratings (concreteness, affectiveness) with phonemes and phonological features.

Key Insights and Ideas

  • Colexifications as Bridges: The central idea is that colexifications act as powerful bridges, allowing us to transfer semantic and phonetic information across languages, especially for low-resource ones.
  • Counter-Intuitive Colexification Patterns: Contrary to some previous theories suggesting that more dissimilar concepts are more likely to colexify, this study found that concepts closer in concreteness and affectiveness are more likely to colexify. This suggests a principle of semantic coherence in word meaning evolution.
  • Phonetic Influences on Meaning: The research hints at a connection between sound and meaning, with certain initial/last phonemes showing significant correlations with concreteness and valence across language families. For example, /k/ is linked to concreteness in Turkic and Tai-Kadai languages, while /p/ is linked to valence in Dravidian and Sino-Tibetan.
  • Phoneme Diversity and Concreteness: A higher "type-to-token ratio" (TTR) of phonemes (meaning more diverse phonemes used for a word) is positively correlated with concreteness. This suggests that more concrete concepts might be expressed with a richer phonetic palette.
  • Phonological Features and Concreteness: Several phonological features appear to be negatively correlated with concreteness, indicating a potential universal tendency in how sounds relate to abstractness.

Key Results

  • A dataset was created with colexification patterns for 142 languages across 21 language families.
  • Statistically significant negative correlations were found between colexifications and the distance of concreteness and affectiveness between concepts, meaning closer concepts colexify more.
  • Specific initial/last phonemes were found to be significantly correlated with concreteness and valence across language families (e.g., /k/ with concreteness, /p/ with valence).
  • The type-to-token ratio of phonemes was positively correlated with concreteness, while segment length was negatively correlated.
  • Phonological features were generally negatively correlated with concreteness.

Limitations and Future Work

  • Anglocentric Bias: The study acknowledges a potential anglocentric bias due to the reliance on English-based source datasets for concreteness and affectiveness ratings.
  • Correlation vs. Causation: The current analyses are primarily correlational. While they reveal interesting patterns, they don't establish causal relationships.
  • Future Work: The authors propose using this dataset for more in-depth multivariate modeling to explore the complex interplay between phonetics, semantics, and human conceptualization across diverse languages and cultures. They also aim to extend this approach to other linguistic domains and low-resource languages.

Summary updated

BibTeX
@inproceedings{chen2023colexifications,
  author = {Yiyi Chen and Johannes Bjerva},
  title = {{Colexifications for Bootstrapping Cross-lingual Datasets: The Case of Phonology, Concreteness, and Affectiveness}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2023},
}
Figure