TL;DR
This paper introduces a novel rule-based approach for morphological disambiguation in Turkish, a highly agglutinative language. By training separate decision lists for each of the 126 morphological features, the authors achieve state-of-the-art accuracy, outperforming previous statistical models. This method effectively tackles the data sparseness problem inherent in Turkish morphology.
What Problem Does This Paper Address?
Morphological disambiguation is the crucial task of selecting the correct interpretation of a word when a morphological analyzer generates multiple possibilities. This is particularly challenging for agglutinative languages like Turkish, where words can have an almost unlimited number of suffixes, leading to a vast number of potential "tags" (morphological analyses). Consequently, a significant portion of words in Turkish text are morphologically ambiguous. Traditional approaches often struggle with this data sparseness, making it difficult to build accurate models. The paper aims to address this by developing a robust and accurate method for Turkish morphological disambiguation.
Methods
The core of the paper's contribution lies in its novel approach to learning disambiguation rules. Instead of training a single model for the entire tag, the authors propose training a separate decision list for each of the 126 individual morphological features identified by the Turkish morphological analyzer.
- Decision Lists: These are ordered lists of rules, where each rule consists of a pattern (based on surface attributes of surrounding words) and a classification (predicting the presence or absence of a specific morphological feature).
- Greedy Prepend Algorithm (GPA): This is a novel decision list learning algorithm used to generate the rules. GPA starts with a default rule and iteratively prepends the rule that provides the maximum "gain" (increase in correctly classified instances) to the front of the list. This algorithm is chosen for its speed and robustness to irrelevant attributes.
- Feature-Specific Models: The key innovation is treating each morphological feature independently. For each feature, a subset of the training data containing that feature is selected. This subset is then split into positive and negative examples based on whether the correct parse includes the feature. GPA is then used to learn a decision list for that specific feature.
- Probabilistic Combination: To disambiguate a word, the system generates all possible parses. Then, for each candidate parse, the probabilities of its constituent features being present are calculated using the accuracies of the individual decision lists. The parse with the highest combined probability is selected.
Datasets
The training and testing data were derived from approximately 1 million words of semi-automatically disambiguated Turkish news text. The paper does not introduce a new benchmark dataset but utilizes existing data processed by the Oflazer morphological analyzer.
Key Insights and Ideas
- Decomposition is Key: Breaking down the complex problem of full tag disambiguation into smaller, manageable problems of individual feature disambiguation is highly effective for agglutinative languages. This directly addresses the data sparseness issue.
- Robustness to Noise: The GPA algorithm's ability to handle a large number of potentially irrelevant or redundant surface attributes (like surrounding words and their suffixes) is crucial for building accurate models without requiring extensive feature engineering.
- Surface Attributes Suffice: The model achieves high accuracy using only surface attributes (word strings, suffixes, character types) within a five-word window, without needing dictionary lookups or information about previous tags. This makes the approach more generalizable and less reliant on external resources.
- Human-Readable Rules: The decision lists generated by GPA are interpretable, offering insights into the disambiguation process, which is a valuable characteristic for understanding linguistic phenomena.
Key Results
The proposed model achieved a 96% disambiguation accuracy on a hand-tagged test set, which is slightly above the best previously reported results using statistical models. For comparison, training a single decision list on full tags yielded only 91% accuracy. The individual decision lists for features achieved an average test set accuracy of 95%, with some features being much harder to disambiguate (e.g., the six worst features had accuracies below 89%).
Limitations and Future Work
- Training Data Accuracy: The authors acknowledge that the semi-automatic generation of the training data may contain errors, which could impact the model's performance. They estimate the training data accuracy to be below 95%.
- Unambiguous Instances for Training: An experiment using only unambiguous instances for training resulted in significantly lower accuracy (82.57%). The authors suggest that a more careful selection of negative examples and a bootstrapping mechanism might improve this.
- Tagging vs. Disambiguation: The paper explored using the learned decision lists for direct "tagging" (predicting the full tag without a morphological analyzer) and achieved 91.23% accuracy for the top 1000 tags. This is identified as a promising avenue for future exploration.
- Future Work: The authors plan to experiment with using tags assigned to previous words as input attributes, which was not included in the current model.
Summary updated