TL;DR
This paper tackles the challenging problem of predicting cognates in a target language given a source word, moving beyond simple classification of existing pairs. By adapting information retrieval techniques, specifically positional shingling and graphical error modeling, the authors significantly improve both cognate classification and prediction accuracy.
What Problem Does This Paper Address?
Cognates are words in different languages that share a common origin, offering insights into historical linguistics and language evolution. Traditional cognate detection methods primarily focus on classifying whether a given pair of words are cognates or not. However, a more complex and practical problem is to predict a potential cognate in a target language when only a source word is provided. This is akin to building a specialized search engine where the query is a word, and the "documents" are the lexicon of another language, with the goal of finding its cognate counterpart. This paper aims to bridge the gap between information retrieval and approximate string similarity for this prediction task.
Methods
The authors propose a novel approach that combines several key techniques:
- Positional Character-based Shingling: Instead of treating words as simple bags of character n-grams (k-grams), this method incorporates the positional information of these n-grams. This is achieved by numbering the shingles from one end (1-ended) or both ends (2-ended) of the word. The 2-ended approach is more robust to mismatches, as it mitigates the issue of order disruption.
- Graphical Error Modelling: This technique focuses on understanding the "dissimilarity" between potential cognates by modeling the transformations that occur between them. It constructs a bipartite graph where nodes represent positional shingles from the source and target words. Edges represent potential mappings between these shingles, and their probabilities are derived from observed transformations in a training set. This allows for the modeling of specific phonetic or orthographic shifts.
- Information Retrieval-based Ranking Functions: The core of the proposed solution is to frame cognate detection as an information retrieval problem. They utilize ranking functions like BM25 and Dirichlet smoothing. These functions are adapted to score potential cognates based on the similarity of their shingle sets and the probabilities derived from the graphical error model. The final score is a weighted combination of a set-based similarity measure and the error model's output.
Datasets
The experiments were conducted on a dataset previously used by Ciobanu et al. (2014). This dataset comprises 400 pairs of cognates and non-cognates for Romanian-French (Ro-Fr), Romanian-Italian (Ro-It), Romanian-Spanish (Ro-Es), and Romanian-Portuguese (Ro-Pt). The dataset was split into a 3:1 ratio for training and testing.
Key Insights and Ideas
- Positional Information is Crucial: Simply treating words as unordered sets of n-grams is insufficient. Incorporating the sequential position of these n-grams significantly improves cognate detection.
- Modeling Transformations is Key to Prediction: Moving beyond similarity, explicitly modeling the errors or transformations between words (e.g., 's' becoming 'ss', or vowel changes) is vital for accurately predicting cognates. The graphical error model effectively captures these linguistic shifts.
- Information Retrieval Framework is Powerful: Adapting ranking functions from information retrieval provides a flexible and effective way to score and rank potential cognates, especially when dealing with large lexicons.
- Combining Similarity and Dissimilarity: The best performance is achieved by synergistically combining measures of similarity (set-based) and dissimilarity (graphical error model), with a weighted contribution that favors similarity.
Key Results
The proposed method, combining 2-ended positional shingling, graphical error modeling, and Dirichlet smoothing, achieved the best performance.
- Classification (Accuracy): The best model achieved an average accuracy of 86%.
- Prediction (MRR): For the more challenging prediction task, the model achieved an average Mean Reciprocal Rank (MRR) of 60%. This indicates a significant improvement over baselines, particularly in ranking the correct cognate highly.
- The paper highlights that the graphical error model significantly boosts test results, and Dirichlet smoothing is preferred for its ability to capture sequential data with fewer parameters.
Limitations and Future Work
- Dataset Size: While the dataset is standard, it's relatively small (400 pairs). Larger, more diverse datasets could further validate the approach.
- Computational Cost: The graphical error model and the process of evaluating the entire lexicon for prediction can be computationally intensive.
- Hyperparameter Tuning: The performance relies on tuning hyperparameters like the weight ($\lambda$) in the scoring function and the parameter ($q$) in the error model.
- Future Work: The authors suggest exploring more complex machine learning algorithms for scoring functions, but note that simpler IR-based approaches might be more robust for large output spaces like entire lexicons. Further investigation into the optimal balance between similarity and dissimilarity metrics is also implied.
Summary updated