ISI at the SIGMORPHON 2017 Shared Task on Morphological Reinflection

Abhisek Chakrabarty, Utpal GarainCoNLL Shared Task 2017 · OpenReview

Abstract: Abhisek Chakrabarty, Utpal Garain. Proceedings of the CoNLL SIGMORPHON 2017 Shared Task: Universal Morphological Reinflection.

  1. Instead of generating words character-by-character, this system treats morphological reinflection as a classification problem. By using an LSTM to predict the correct "edit tree"—a structural map of how to transform a source word into a target form—the researchers created a language-independent model that outperforms baselines in high-resource settings but struggles when data is scarce.

Introduction

Morphological reinflection is the task of taking a base word (lemma) and a set of target features (like "past tense, third person") and producing the correct inflected form (e.g., communicate + V;3;SG;PRS $\rightarrow$ communicates).

Traditionally, this challenge has been a tug-of-war between two approaches:

  1. Rule-based/Statistical methods: These often require heavy manual feature engineering, making them difficult to scale across the world's thousands of languages.
  2. Neural Sequence-to-Sequence (seq2seq) models: While powerful, these models generate characters one by one. This can lead to "hallucinations" where the model produces a word that looks plausible but is structurally unrelated to the input.

The authors aim to bridge this gap by creating a system that is neural and language-independent, yet constrained by the structural relationship between the input and output.

Methodology: Edit Tree Classification

The core innovation of this work is framing reinflection as a classification problem over Edit Trees. Rather than learning to spell the target word, the model learns the transformation required to reach it.

Edit Trees

An edit tree identifies the Longest Common Substring (LCS) between the source and target. It then recursively models the prefixes and suffixes that need to change. For example, to turn sang into sing, the edit tree identifies "ng" as the stable part and defines the transformation needed for the prefix "sa" $\rightarrow$ "si".

Figure 1: Structural representation of an edit tree transformation from 'sang' to 'sing'.

Neural Architecture

The model utilizes a multi-stage neural pipeline to predict the appropriate edit tree:

  • Word Encoding: A Long Short-Term Memory (LSTM) network processes the character sequence of the input word to create a fixed-length vector representation.
  • Feature Mapping: Morphosyntactic tags (like "plural" or "past") are mapped to a binary feature vector.
  • Applicability Filtering: Not every edit tree can be applied to every word (e.g., you cannot remove a suffix "-ing" from a word that ends in "-ed"). The model includes a "filter" layer that masks out edit trees that are invalid for the specific input string, simplifying the learning objective.
  • Joint Training: The LSTM output, the feature vector, and the applicability bit-mask are fed into a final layer that classifies the input into one of the distinct edit trees seen during training.
Figure 2: The proposed neural architecture for edit tree classification, featuring the embedding layer and applicability filter.

Experimental Results

The system was evaluated on the SIGMORPHON 2017 Shared Task dataset, covering 52 different languages across three training sizes: Low, Medium, and High.

The system's performance was highly dependent on the amount of training data available:

  • High-Resource: The model beat the baseline on 15 out of 52 languages, showing exceptional accuracy on languages like Hindi (99.1%), Urdu (98.2%), and Georgian (94.4%).
  • Medium/Low-Resource: Performance dropped significantly, beating the baseline in only 5 and 4 languages, respectively.

These results suggest that while the "edit tree" approach provides a powerful inductive bias by respecting the word stem, it requires a critical mass of examples to correctly identify which specific transformation tree applies to which morphological context.

Conclusion and Future Work

By predicting an edit tree rather than a character sequence, the model inherently respects the stem of the word, preventing it from "forgetting" the root. This language-independent approach requires zero linguistic expertise or manual rule-writing. However, the primary limitation is its "data hunger," as it struggles to generalize complex logic when only a few hundred examples are provided. Future work could involve using unsupervised pre-training on external monolingual corpora to improve performance in low-resource scenarios.

Summary updated

BibTeX
@inproceedings{chakrabarty2017isi,
  author = {Abhisek Chakrabarty and Utpal Garain},
  title = {{ISI at the SIGMORPHON 2017 Shared Task on Morphological Reinflection}},
  booktitle = {CoNLL Shared Task},
  year = {2017},
  url = {https://openreview.net/forum?id=9kKms9oig4c},
}
Figure