TL;DR
This paper explores training a neural network to mimic a hand-crafted finite-state transducer (FST) for Arapaho verb morphology. By leveraging the FST's output as training data, they achieve high accuracy in analyzing unseen verb forms, demonstrating a promising approach for building morphological analyzers for under-resourced, morphologically complex languages.
What Problem Does This Paper Address?
Building accurate morphological analyzers for languages with complex inflectional systems, like the polysynthetic Arapaho language, is a significant challenge. Traditional methods, such as Finite State Transducers (FSTs), require extensive linguistic expertise and manual effort, making them time-consuming to develop and maintain. While neural networks have shown promise in this area, training them from scratch for morphologically rich languages can be difficult due to limited data and the complexity of capturing all inflectional variations. This paper addresses the gap of efficiently bootstrapping a neural morphological analyzer by using an existing, albeit complex, FST as a source of supervised training data. This approach aims to leverage the FST's linguistic knowledge to train a more adaptable and generalizable neural model.
Methods
The core of this research involves training a recurrent neural network (RNN) encoder-decoder model. Specifically, they employ a bidirectional LSTM-based sequence-to-sequence architecture, a standard in neural machine translation and adapted for morphological reinflection tasks. A key aspect of their method is treating morphological parsing as a "translation" task: converting an input character sequence (the inflected verb form) into an output sequence of morphosyntactic tags and the verbal root.
To optimize the training process, they experimented with different orderings of input tags and the root morpheme. They found that the "Tags+Root+Tags" order for training, followed by evaluation on "Tags+Root," yielded the best results. To handle systematic ambiguities in the FST's output (where one form can have multiple valid parses), they employed a "disambiguation" strategy by altering the tagset. This involved creating more general tags to represent multiple possible interpretations, sacrificing some informativeness for improved generalization.
Datasets
The primary dataset for training and evaluation was generated by an existing FST morphological analyzer for Arapaho verbs. This FST produced pairs of inflected word forms and their corresponding parsed forms (morphosyntactic tags and root). The initial FST-extracted dataset contained 584,574 examples. After filtering out incorrectly stored forms, 60% of the remaining data was used for training, 20% for validation, and 20% for testing. The paper also mentions that the FST-generated data included both ambiguous and unambiguous verb forms, with the latter constituting a little over half of the total.
Key Insights and Ideas
- Bootstrapping with FSTs: Using a hand-crafted FST as a source of supervised data is a viable and effective strategy for training neural morphological analyzers, especially for morphologically complex languages where manual rule creation is arduous.
- Generalization Beyond Explicit Training: The neural model demonstrated a remarkable ability to generalize to unseen verb forms, even those not explicitly present in the training data. This highlights the power of neural networks to capture underlying linguistic patterns.
- Input Ordering Matters: The specific ordering of morphosyntactic tags and the root morpheme in the input sequence significantly impacts the model's performance, suggesting that the network learns differently based on how it receives this information.
- Systematic Ambiguity Handling: A pragmatic approach to handling systematic ambiguity in FST output (by creating more general tags) can improve overall accuracy, though it comes at the cost of some detailed information. This trade-off is crucial for practical applications.
- Polysynthetic Languages and Ambiguity: While polysynthetic languages are known for complex morphology, the paper suggests they might exhibit less unsystematic ambiguity in parsing compared to some other language types, which can be advantageous for neural modeling.
Key Results
The neural morphological analyzer achieved impressive accuracy:
- Unambiguous verbs: 98.68% accuracy on unseen forms.
- All verbs (including ambiguous): 92.90% accuracy after systematic ambiguity was addressed through tagset alteration.
- The model's predictions on individual tags and letters were also analyzed, showing high precision and recall for most elements.
Limitations and Future Work
The authors acknowledge several limitations and suggest avenues for future research:
- Loss of Information: The disambiguation strategy, while improving accuracy, leads to a loss of detailed morphosyntactic information. Future work could explore more sophisticated methods to retain this information while still achieving generalization.
- Handling Morphophonemic Changes: The current approach might not fully capture complex morphophonemic alternations, which can further complicate disambiguation.
- Further Disambiguation: The paper suggests that more advanced pre-processing could identify and re-place alternative parses with more accurate "super-tags" to better handle ambiguities, especially those complicated by morphophonemic changes.
- Incomplete Data Scenarios: While some experiments used random subsets of FST data, further investigation into how the neural model performs on truly incomplete hand-written descriptions of languages would be valuable.
Summary updated