Multilingual Deep Lexical Acquisition for HPSGs via Supertagging

Phil Blunsom, Timothy BaldwinEMNLP 2006 · OpenReview

Abstract

We propose a conditional random field-based method for supertagging, and apply it to the task of learning new lexical items for HPSG-based precision grammars of English and Japanese. Using a pseudo-likelihood approximation we are able to scale our model to hundreds of supertags and tens-of-thousands of training sentences. We show that it is possible to achieve start-of-the-art results for both languages using maximally language-independent lexical features. Further, we explore the performance of the models at the type- and token-level, demonstrating their superior performance when compared to a unigram-based baseline and a transformation-based learning approach.

Introduction

Precision grammars, such as those based on Head-driven Phrase Structure Grammar (HPSG), provide deep semantic abstractions and distinguish between grammatical and ungrammatical inputs. However, their reliance on highly detailed lexical entries—specifying not just parts of speech but exact argument structures—creates a significant "lexical gap." Deep Lexical Acquisition (DLA) aims to automate the process of adding new words to these hand-coded lexicons. This paper introduces a scalable sequence labeling approach using Conditional Random Fields (CRFs) to predict complex "supertags" for previously unseen words in English and Japanese.

Methodology

The core of this approach is a sequence labeling model using CRFs. To address the computational expense of training on hundreds of possible supertags across tens of thousands of sentences, the authors employ a pseudo-likelihood approximation. This optimizes the model based on the local neighborhood of each node rather than a globally normalized probability, ensuring scalability.

Feature Engineering

The model utilizes "maximally language-independent" features to ensure portability across different linguistic structures:

  • Word Context: Unigrams and bigrams of surrounding words.
  • Lexical Features: N-gram prefixes and suffixes (e.g., "-ing", "-ed").
  • Character Sets: Flags for specific character types, such as uppercase in English or Kanji/Hiragana in Japanese.

Additionally, the model handles continuous multiword expressions (MWEs) using a "ditto" tagging convention, allowing them to be treated as single lexical units during prediction.

Experimental Setup

The researchers evaluated the model on two major precision grammars: the English Resource Grammar (ERG) and JACY (Japanese Grammar). The scale of these datasets and the complexity of the grammars are detailed in the table below.

Table 1: Grammar and Treebank statistics for the English (ERG) and Japanese (JACY) datasets.

The CRF model was compared against a simple Unigram Baseline and fnTBL (a transformation-based learner).

Results and Discussion

The CRF model outperformed all baselines across both languages, particularly in identifying "unknown" tokens (words not seen during training). In English, the CRF achieved an overall token accuracy of 91.7%, while for unknown tokens, it reached 48.9%—a massive improvement over the 5.3% achieved by the unigram baseline.

Table 2: Performance comparison of the CRF model against the Baseline and fnTBL across ERG and JACY.

Key Insights

  • Lexical Features are Crucial: For unknown words, internal word structure (prefixes/suffixes) is more informative than surrounding context. Adding these features nearly doubled accuracy for unknown English tokens.
  • Precision for Lexicography: High precision is more valuable than high recall in DLA, as it provides human lexicographers with accurate suggestions rather than noisy data. The CRF model naturally favored high precision.
  • Language Independence: By focusing on character-set features rather than deep morphological analysis, the system proved easily portable between English and Japanese with minimal tuning.
  • The Difficulty of Verbs: The model struggled most with verbs and adverbs, likely due to their complex subcategorization frames which are difficult to infer from local context alone.

Conclusion

This work demonstrates that CRF-based supertagging is a powerful tool for Deep Lexical Acquisition. While type-level performance suggests that fully autonomous lexicon expansion remains a challenge, the model's high precision makes it an effective aid for grammar developers. Future work will focus on incorporating more sophisticated morphological features to handle verbal complexity and exploring "on-line" DLA to guess word types during live parsing.

Summary updated

Abstract

We propose a conditional random field-based method for supertagging, and apply it to the task of learning new lexical items for HPSG-based precision grammars of English and Japanese. Using a pseudo-likelihood approximation we are able to scale our model to hundreds of supertags and tens-of-thousands of training sentences. We show that it is possible to achieve start-of-the-art results for both languages using maximally language-independent lexical features. Further, we explore the performance of the models at the type- and token-level, demonstrating their superior performance when compared to a unigram-based baseline and a transformation-based learning approach.
BibTeX
@inproceedings{blunsom2006multilingual,
  author = {Phil Blunsom and Timothy Baldwin},
  title = {{Multilingual Deep Lexical Acquisition for HPSGs via Supertagging}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2006},
  url = {https://openreview.net/forum?id=By-1WzMuZr},
}
Figure