TL;DR
This paper tackles the challenge of tracking semantic change in related languages by introducing a novel method for analyzing cognate sets using word embeddings. They present a new, manually curated dataset of English and Romance language cognates, annotated with etymological information, and use it to uncover interesting patterns in semantic divergence, particularly for English borrowings.
What Problem Does This Paper Address?
Historical linguistics has long been interested in how the meanings of words change over time, especially within related languages that share a common ancestor. While computational linguistics has made strides in tracking semantic change within a single language using corpora, analyzing change across languages, particularly for words inherited or borrowed between them (cognates), has been less explored. A key challenge is distinguishing between true cognates (words with shared etymology and meaning) and "false friends" (words with shared etymology but divergent meanings). Furthermore, understanding the historical context of these words, such as when they entered a language and through what channel (e.g., direct inheritance vs. borrowing), is crucial for a comprehensive analysis of semantic divergence. This paper aims to bridge this gap by developing a computational framework to analyze semantic change in cognate sets across English and Romance languages, considering their historical transmission.
Methods
The core of the paper's methodology lies in leveraging cross-lingual word embeddings to measure semantic similarity between cognate words. They use pre-trained FastText embeddings for six languages (English, Romanian, French, Italian, Spanish, Portuguese). To enable cross-lingual comparison, they employ an alignment algorithm to create a common vector space, minimizing distances between known word pairs. Semantic similarity is then calculated using cosine similarity between the word vectors.
Beyond embeddings, the paper also utilizes word frequency scores (extracted using the multilingual Wordfreq library) and word polysemy (measured using Open Multilingual WordNet) as additional linguistic features.
A significant methodological contribution is the creation of a new, manually curated dataset of cognate sets. This dataset was built upon an existing list of Romance language cognates and augmented with English cognates. Crucially, the English cognates were manually annotated by experts to distinguish between "old borrowings" (entering English before the mid-15th century, often via Old French) and "recent borrowings" (entering English from the mid-15th century onwards, often directly from Latin). This annotation allows for a more nuanced analysis of semantic divergence based on historical transmission channels.
Datasets
- Original Cognate Dataset: A list of 3,218 complete cognate sets in Romanian, French, Italian, Spanish, and Portuguese, with their Latin ancestors, sourced from online etymology dictionaries.
- Augmented Cognate Dataset: The paper expands the original dataset by adding English cognates for a subset of 305 cognate sets. This augmented dataset is the primary focus for their analysis.
- New Curated Dataset: The authors manually curated the augmented dataset, including high-quality annotations for English cognates, categorizing them as "old borrowings" or "recent borrowings" based on etymology and period of entry into English. This dataset is made publicly available.
Key Insights and Ideas
- Bimodal Distribution of English Cognate Similarity: A striking finding is the bimodal distribution of semantic similarity scores for English cognates when compared to their Romance counterparts. This suggests the existence of at least two distinct groups of English borrowings with different semantic relationships.
- Hypothesis on Borrowing Strata: The authors hypothesize that this bimodal distribution might reflect different historical strata of borrowings: older borrowings (potentially more divergent) versus more recent borrowings (potentially more similar).
- Frequency and Polysemy Differences: While direct semantic similarity didn't show a statistically significant difference between old and recent English borrowings, analysis of word properties revealed significant patterns: old borrowings tend to have higher frequencies and polysemy than recent borrowings. This suggests that older, more established words have undergone more semantic evolution.
- Transmission Channel Matters: Qualitative analysis of specific examples (like "fork" and "powder") highlights how the transmission channel (direct borrowing vs. via another language) and the socio-cultural context of adoption significantly influence semantic divergence.
- Concrete vs. Abstract Terms: Terms denoting concrete, experientially grounded concepts (animals, materials, seasons) tend to show higher semantic similarity, possibly due to frequency effects or stable referents. Abstract terms with restricted domains or those that haven't experienced polysemic development also exhibit high similarity.
Key Results
- Overall Accuracy: The automatic extraction of cognate sets achieved an overall accuracy of 53% on the curated dataset, with per-language accuracies ranging from 70.8% (English) to 82.6% (Spanish).
- Average Similarity: The distribution of cross-lingual similarity scores between Romance cognates is unimodal, skewed to the right, with a mean similarity around 0.7.
- Frequency and Polysemy: Old English borrowings and their Romance cognates exhibit statistically significant higher average frequencies and polysemy scores compared to recent borrowings. For instance, average log-frequencies for old borrowings were around 4.12, while for recent borrowings, they were around 3.55.
- Qualitative Examples: Detailed analysis of specific cognate pairs like "fork" and "camp" illustrates significant semantic divergence, with similarity scores as low as 0.1-0.4, attributed to different historical trajectories and socio-cultural contexts.
Limitations and Future Work
- Limitations:
- The initial automatic extraction of cognates had a moderate accuracy (53%), necessitating extensive manual curation.
- The dataset size for English cognates was reduced due to difficulties in identifying etymologies and corresponding cognates.
- Romanian is not supported by the Open Multilingual WordNet for polysemy analysis.
- The bimodality in English cognate similarity remains somewhat unexplained, with the hypothesis about old vs. recent borrowings not being statistically supported by similarity scores alone.
- Future Work:
- Extend the study to cognate similarity based on phonetic transcription to complement the current orthographic analysis.
- Investigate more deeply the automatic identification of word entry dates into languages, potentially by adapting existing methods to approximate word "age."
- Further explore the reasons behind the observed bimodality in English cognate similarity.
Summary updated