TL;DR
This paper presents a supervised learning approach for morphological paradigm classification that significantly improves the accuracy of reconstructing complete inflection tables for unseen words. By combining a longest common subsequence (LCS) based generalization strategy with a discriminative classifier and careful feature selection, the system achieves state-of-the-art results across 11 languages, offering a more robust and efficient way to build morphological resources.
What Problem Does This Paper Address?
Building accurate morphological analyzers for new languages is a challenging and time-consuming task. While identifying word inflection patterns (morphological paradigms) is crucial for many natural language processing (NLP) tasks, creating comprehensive resources for every language is arduous. Previous work proposed using the longest common subsequence (LCS) of word forms within an inflection table to generalize these patterns. However, this approach, while simple, had limitations. This paper aims to improve upon this LCS-based generalization by integrating it with a more sophisticated discriminative classifier and feature selection to achieve higher accuracy in reconstructing complete inflection tables for unknown words.
Methods
The core of the proposed method involves two main stages:
- Paradigm Generalization (LCS-based): Similar to prior work, the system first generalizes concrete inflection tables into abstract paradigms. This is done by identifying the longest common subsequence (LCS) shared by word forms within a table. These LCSs are treated as variables, allowing for a reduction in the number of distinct paradigms. Figure 1 illustrates this process with an Icelandic example.
- Discriminative Classification: For classifying unseen base forms into these generalized paradigms, the authors employ a linear multi-class Support Vector Machine (SVM). Crucially, they extract binary features from the substrings at the left and right edges of the base form. This feature engineering is a key improvement, allowing the classifier to leverage information beyond just suffixes, which is particularly beneficial for languages with prefixation or other non-suffixal morphological processes.
- Feature Selection: To optimize the SVM's performance, a development set was used for feature selection. This involved tuning the SVM's regularization parameter (C) and determining the optimal length and type (prefix/suffix) of substrings to use as features.
Datasets
The paper evaluates the proposed method on two sets of datasets:
- Experiment 1: Utilizes existing datasets from Durrett and DeNero (2013) for Finnish (verbs, nouns+adjectives), German (nouns, verbs), and Spanish (verbs). These datasets contain complete inflection tables.
- Experiment 2: Introduces a new, more comprehensive, and complex dataset developed by the authors, covering 8 additional languages: Catalan (nouns, verbs), English (verbs), French (nouns, verbs), Galician (nouns, verbs), Italian (nouns, verbs), Portuguese (nouns, verbs), Russian (nouns), and Maltese (verbs). This dataset is noted for being more realistic, including incomplete or defective inflection tables, and containing alternate forms.
Key Insights and Ideas
- Prefixes Matter: The improved accuracy on German verbs highlights the importance of considering prefixes, not just suffixes, for morphological analysis. German verb prefixes can signal specific grammatical behaviors (e.g., inseparable prefixes affecting past participle formation) that are missed by purely suffix-based approaches.
- LCS as a Foundation: The LCS-based generalization strategy provides a robust and linguistically motivated foundation for capturing paradigmatic relationships, even across different languages.
- Discriminative Power of Substrings: Using both prefix and suffix substrings as features for the SVM classifier significantly enhances its ability to distinguish between similar paradigms, leading to higher accuracy.
- Efficiency for Resource-Scarce Languages: The method's success suggests a viable path for rapidly developing morphological resources for languages with limited existing tools, by leveraging a small set of well-chosen inflection tables.
Key Results
- Experiment 1: The proposed SVM-based approach consistently outperforms previous methods (AFH14 and D&DN13) on the three languages, showing significant accuracy improvements, particularly for German verbs. For instance, German verbs saw an increase in per-table accuracy from 68.0% (AFH14) to 85.0% (D&DN13) and then to 91.5% with the new SVM method.
- Experiment 2: Across the 8 new languages, the SVM classifier also demonstrates strong performance, often surpassing the AFH14 baseline. For example, on Italian verbs, per-table accuracy reached 89.68% with the SVM, compared to 83.20% for the baseline.
- Per-Table vs. Per-Form Accuracy: The paper notes that in some cases, per-table accuracy can be significantly lower than per-form accuracy. This is attributed to similar paradigms where a misclassification of the paradigm leads to an incorrect table, even if most individual forms are predicted correctly.
Limitations and Future Work
- Maltese Verbs Anomaly: The paper identifies Maltese verbs as an outlier, where despite a high oracle score, the per-form accuracy is relatively low. This suggests that the current features might not fully capture the complex Semitic interdigitation in Maltese morphology, warranting further analysis.
- Holistic Table Inspection: The authors suggest an extension where the reconstructed inflection table is evaluated holistically, rather than relying solely on base form features. This could involve using features from the generated forms to rate paradigm compatibility, potentially identifying incorrect reconstructions by their tendency to produce unseen phonotactic patterns.
- Semi-Supervised Extension: The paper acknowledges that the current method is supervised and suggests that extending it to a semi-supervised scenario (as explored in prior work) is a promising and interesting direction. This could further leverage unlabeled data to improve performance.
Summary updated