Studying Laws of Semantic Divergence across Languages using Cognate Sets

Ana Uban, Alina Maria Ciobanu, Liviu P. DinuACL 2019

TL;DR

This paper introduces a novel method to study how word meanings diverge across related languages by comparing cognates. By analyzing semantic distances in multilingual word embeddings, they uncover statistical relationships between word properties like frequency and polysemy, and the degree of meaning shift, paving the way for cross-lingual semantic change laws.

What Problem Does This Paper Address?

Historical linguistics has long been interested in semantic change – how the meaning of words evolves over time. While computational linguistics has made strides in tracking this within a single language using diachronic corpora, this approach misses a crucial aspect: language evolution often happens in a multilingual context, with words being borrowed and inherited across languages. This paper addresses the gap of understanding semantic change across languages, specifically by examining how cognates (words with a common ancestor) in related languages diverge in meaning from their shared origin. They aim to move beyond monolingual studies to uncover cross-lingual laws of semantic change.

Methods

The core of the paper's methodology lies in leveraging multilingual word embeddings. Specifically, they use pre-trained FastText embeddings for six languages (five Romance languages and English). To compare cognates across these languages, they first align these individual language embedding spaces into a shared, common embedding space. This alignment is achieved using an algorithm that finds a linear transformation, preserving original distances as much as possible.

Once cognates are represented in this shared space, the semantic distance between them is computed using cosine distance. This distance serves as a proxy for the degree of semantic divergence. To identify "deceptive cognates" or "false friends" (cognates that have diverged in meaning), they employ a method similar to Uban et al. (2019), where a cognate is considered a false friend if another word in the target language is semantically closer to the source word than its cognate. The "falseness" of a false friend pair is quantified by the difference in semantic distances.

Finally, they analyze how word properties like frequency and polysemy correlate with this measure of semantic divergence (falseness). Frequency is represented by normalized rank, and polysemy is measured by the number of synsets a word belongs to in WordNet. They use Spearman correlation and partial correlations to isolate the effects of frequency and polysemy on semantic divergence.

Datasets

The primary dataset used is a list of cognate sets in Romance languages (Romanian, French, Italian, Spanish, Portuguese) along with their Latin ancestors, compiled by Ciobanu and Dinu (2014). This dataset was extracted from online etymology dictionaries. A subset of these sets also includes English cognates. For evaluating their false friend detection, they also refer to results obtained using multilingual WordNet as a gold standard.

Key Insights and Ideas

  • Cross-lingual perspective is crucial: Semantic change isn't solely an internal language phenomenon; it's deeply influenced by inter-lingual interactions.
  • Cognates as a window: Comparing cognates across languages offers a unique way to observe semantic divergence, treating present-day meanings as snapshots of different evolutionary paths from a common ancestor.
  • Frequency and polysemy matter cross-lingually: Similar to monolingual studies, word frequency and polysemy are found to be significant factors influencing how much a cognate's meaning shifts across languages.
  • Potential for "laws" of cross-lingual semantic change: The paper lays the groundwork for identifying statistical regularities in how meanings diverge between related languages, analogous to existing monolingual semantic change laws.
  • Romance language relationships: Stronger correlations between polysemy and falseness were observed for more closely related Romance languages, suggesting linguistic proximity might influence divergence patterns.

Key Results

  • On average, 37% of cognate pairs in their dataset were identified as deceptive cognates (false friends).
  • Correlations between word frequency and semantic divergence (falseness) were generally negative when using raw ranks, but when interpreted as positive relationships (due to rank inversion), they suggest that more frequent words have diverged more in meaning across languages. This is a notable difference from some monolingual findings.
  • Polysemy was found to be positively correlated with semantic divergence across most language pairs, indicating that more polysemantic words tend to undergo greater meaning shifts or that words with greater shifts tend to become more polysemantic.
  • The correlations for polysemy were stronger for more closely related Romance languages (e.g., Spanish-Portuguese) and weaker for English, which is not a Romance language.

Limitations and Future Work

The authors acknowledge that their approach relies on pre-aligned multilingual word embeddings, which can be noisy or biased. They also note that the use of WordNet for validation limits them to languages for which WordNet is available (excluding Romanian).

Future work suggested includes:

  • Incorporating the proto-word (e.g., Latin etymon) into the analysis where available to gain deeper insights into the divergence process.
  • Further investigating and explaining the observed correlations.
  • Studying other hypothesized laws of semantic change in a multilingual setting.

Summary updated

BibTeX
@inproceedings{uban2019studying,
  author = {Ana Uban and Alina Maria Ciobanu and Liviu P. Dinu},
  title = {{Studying Laws of Semantic Divergence across Languages using Cognate Sets}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
}
Figure