TL;DR
This paper introduces a web-based tool that simulates how words evolve from one language to another, focusing on orthographic changes. By training a sequence labeling model on word pairs, the system can predict potential historical forms of borrowed words, aiding historical linguists in reconstructing language relationships and evolution.
What Problem Does This Paper Address?
Historical linguistics grapples with fundamental questions about how languages relate to each other and how they change over time and space. A key aspect of this is understanding how words are borrowed and evolve in new languages. Traditionally, this involved manual analysis and phonetic transcriptions. While modern approaches leverage computational methods, predicting word evolution often relies on complex phonetic transcriptions or extensive linguistic knowledge. This paper addresses the challenge of predicting word form evolution in a target language from a source language using a more accessible, orthography-based approach, aiming to assist historical linguists by automating and narrowing down the possibilities for investigation.
Methods
The core of the system is a sequence labeling model built using Conditional Random Fields (CRFs). The process begins with word alignment using the Needleman-Wunsch algorithm to pair characters from related words in source and target languages. This alignment is then used to train the CRF model. The model learns to predict the character sequence of a word in the target language based on its orthographic form in the source language. The system is designed to be language-agnostic, meaning it doesn't require pre-existing phonetic knowledge or linguistic rules beyond the training data. Features for the CRF model include character n-grams around the current token, capturing local orthographic patterns.
Datasets
The experiments were conducted using a dataset of word-etymon pairs (Ciobanu and Dinu, 2014). This dataset was compiled from machine-readable dictionaries and contains information about the etymology of words. The authors extracted Romanian words that had etymons in 10 different source languages. For each language, subsets of 800 word pairs were used for training, development, and testing, ensuring comparability across source languages.
Key Insights and Ideas
- Orthography as a Proxy for Sound Change: The paper's central idea is that orthographic changes in borrowed words often leave traceable patterns that can serve as a good proxy for underlying sound correspondences, making the task more accessible than relying solely on phonetic data.
- Language-Agnostic Approach: By focusing on orthography and using a sequence labeling model, the system avoids the need for deep linguistic expertise or language-specific phonetic inventories, making it broadly applicable.
- Assisting Historical Linguistics: The tool aims to significantly narrow down the search space for historical linguists, providing a list of plausible word forms rather than requiring them to manually hypothesize every possibility.
- Cultural Collaboration Correlates with Performance: The system performs best for source languages like English and French, which have had strong cultural ties with Romanian, suggesting that the intensity and recency of contact might influence the predictability of word evolution.
Key Results
The paper presents quantitative results using edit distance (both normalized and un-normalized) as an evaluation metric. The proposed system consistently outperforms a baseline "majority class" method across all tested source languages. For instance, the normalized edit distance for English as a source language is 0.15, significantly lower than the baseline's 0.23. The system also shows good performance for French and Italian. However, performance is lower for languages geographically more remote from Romania, like Portuguese and Spanish, indicating that distance might have historically hindered language contact and thus the predictability of word evolution. Table 2 provides qualitative examples of the system's 5-best productions for various source languages, highlighting the correct Romanian cognates in bold.
Limitations and Future Work
The authors acknowledge that the current system focuses on single word production and is primarily tested with Romanian as the target language. Future work includes:
- Expanding the system to support more target languages.
- Enhancing the user interface to handle blocks of text instead of just single words.
- Incorporating other types of linguistic relationships, such as cognate production and proto-word production, into the application.
Summary updated