[Paper Review] Probabilistic Parsing Using Left Corner Language Models

Christopher D. Manning, Bob Carpenter|arXiv (Cornell University)|Nov 17, 1997

Natural Language Processing Techniques16 references5 citations

TL;DR

This paper introduces a probabilistic parsing framework that leverages left-corner parsing to improve accuracy by conditioning rule probabilities on both top-down goals and bottom-up derivations. Evaluated on the Wall Street Journal using Penn Treebank data, the method outperforms standard top-down probabilistic context-free grammars, though the dataset's flat bracketings and grammar over/undergeneration limit its effectiveness as a benchmark.

ABSTRACT

We introduce a novel parser based on a probabilistic version of a left-corner parser. The left-corner strategy is attractive because rule probabilities can be conditioned on both top-down goals and bottom-up derivations. We develop the underlying theory and explain how a grammar can be induced from analyzed data. We show that the left-corner approach provides an advantage over simple top-down probabilistic context-free grammars in parsing the Wall Street Journal using a grammar induced from the Penn Treebank. We also conclude that the Penn Treebank provides a fairly weak testbed due to the flatness of its bracketings and to the obvious overgeneration and undergeneration of its induced grammar.

Motivation & Objective

• To develop a probabilistic parsing approach that combines top-down and bottom-up reasoning for better language modeling.
• To investigate whether conditioning rule probabilities on both goals and derivations improves parsing accuracy.
• To evaluate the performance of the proposed parser on real-world syntactic data from the Penn Treebank.
• To assess the limitations of the Penn Treebank as a testbed for parsing models.

Proposed method

• The parser uses a left-corner strategy, where each rule application is guided by a top-down goal (the expected category) and a bottom-up derivation (the already-parsed constituents).
• Rule probabilities are conditioned on both the predicted category and the actual constituents derived so far, enabling richer context in probability estimation.
• A grammar is induced from parsed data using maximum-likelihood estimation on the observed left-corner derivations.
• The model integrates top-down hypothesis generation with bottom-up verification, reducing ambiguity and improving parsing decisions.
• The approach extends traditional probabilistic context-free grammars by incorporating structural constraints from both directions of parsing.

Experimental results

Research questions

• RQ1 Can conditioning rule probabilities on both top-down goals and bottom-up derivations lead to improved parsing accuracy?
• RQ2 How does the left-corner parsing strategy compare to standard top-down probabilistic context-free grammars on real-world text?
• RQ3 To what extent does the Penn Treebank's bracketing structure constrain the evaluation of parsing models?
• RQ4 Does the induced grammar from the Penn Treebank exhibit significant overgeneration or undergeneration?

Key findings

• The left-corner parser outperforms standard top-down probabilistic context-free grammars on the Wall Street Journal corpus.
• The improvement is attributed to the richer conditioning of rule probabilities using both top-down and bottom-up information.
• The Penn Treebank is found to be a weak testbed due to its flat bracketings, which limit the ability to evaluate structural precision.
• The induced grammar exhibits both overgeneration and undergeneration, suggesting limitations in the training data's representativeness.