The NYU System for the CoNLL-SIGMORPHON 2018 Shared Task on Universal Morphological Reinflection

Katharina Kann, Stanislas Lauly, Kyunghyun Cho

Abstract

The NYU system for the CoNLL–SIGMORPHON 2018 Shared Task addresses the problem of morphological inflection in context—predicting the correct form of a lemma based on its surrounding sentence. We employ a triple-encoder LSTM architecture that independently processes the lemma, preceding context, and following context. This approach achieved the second-best results in the competition's low-resource track, significantly outperforming the official baseline and demonstrating the effectiveness of using raw context as a proxy for morphological tags.

Introduction

In morphologically rich languages such as Russian or Finnish, a single base word (lemma) can manifest in numerous forms to express grammatical features like gender, tense, or case. This "type sparsity" poses a significant challenge for NLP systems. While previous research often focused on "isolated" inflection (generating a form from a lemma and a set of explicit tags), real-world applications frequently lack these tags.

This paper addresses Task 2, Track 2 of the 2018 Shared Task: predicting the correct inflected form using only the lemma and the raw sentence context. For instance, given the sentence "The athletes [run] yesterday," a system must infer from "yesterday" and "athletes" that the correct form is "ran." This task represents a more realistic and difficult scenario for morphological modeling.

Methods

The NYU system extends the standard attention-based LSTM encoder-decoder architecture into a multi-encoder configuration designed to handle contextual dependencies.

Triple-Encoder Architecture

The system utilizes three distinct encoders to capture different information sources:

  1. Lemma Encoder: Processes the target lemma at the character level to capture fine-grained orthographic changes.
  2. Past Context Encoder: Processes words preceding the target.
  3. Future Context Encoder: Processes words following the target.
The triple-encoder architecture with separate attention mechanisms for lemma, past, and future contexts.

Hybrid Representations and Triple Attention

While the lemma is character-based, context words are processed using Byte Pair Encoding (BPE). This allows the model to handle sub-word units, enabling it to recognize morphological patterns in surrounding words even if the full word is out-of-vocabulary. The decoder employs three separate attention mechanisms, allowing it to simultaneously attend to the lemma and both context windows at each generation step.

Results

The system was evaluated on datasets derived from Universal Dependencies (UD) treebanks, focusing on the low-resource setting for seven languages: German (de), English (en), Spanish (es), Finnish (fi), French (fr), Russian (ru), and Swedish (sv).

Accuracy comparison of the NYU system against the baseline (BL) and other competitors in the low-resource track.

The NYU system was a top performer, securing the second-highest test accuracy among five participants. It significantly outperformed the official baseline; for example, in German, the baseline accuracy was 0.10%, whereas the NYU system reached 44.08%. Performance varied by language complexity, with high accuracy in English (66.57%) and lower results in Finnish (15.37%), which possesses more complex morphology.

Discussion and Conclusion

A key insight from this work is that surrounding words provide sufficient syntactic and semantic signals to serve as a proxy for explicit morphological tags. The use of BPE for context and separate encoders for past and future dependencies proved crucial for capturing these signals.

While highly effective in low-resource settings, the system's scalability to high-resource data remains to be explored. Additionally, the inherent ambiguity of the task—where multiple forms might be grammatically correct—remains a challenge for single-target evaluation metrics. Future work could adapt this architecture for tag-based tracks by incorporating a fourth encoder for morphological features.

Summary updated

BibTeX
@inproceedings{kann2018nyu,
  author = {Katharina Kann and Stanislas Lauly and Kyunghyun Cho},
  title = {{The NYU System for the CoNLL-SIGMORPHON 2018 Shared Task on Universal Morphological Reinflection}},
  booktitle = {CoNLL Shared Task},
  year = {2018},
  doi = {10.18653/v1/k18-3006},
}
Figure