Applying and Improving the Restriction Grammar Approach for Dutch Patient Discharge Summaries

Peter Spyns, Geert AdriaensCOLING 1992

TL;DR

This paper adapts the Restriction Grammar (RG) formalism, previously successful for English medical text, to process Dutch patient discharge summaries (PDS). They introduce a novel control structure for the parser to handle ambiguity more efficiently and outline a semi-automatic method for building a Dutch medical lexicon, aiming to improve information extraction from PDSs.

What Problem Does This Paper Address?

Patient discharge summaries (PDSs) contain crucial medical information, but extracting this data from free-text reports is challenging. Traditional methods often rely on pattern matching, which can miss the nuances of medical language. Linguistically-based analysis offers a more sophisticated approach, but developing such systems for specific languages and domains requires significant effort. This paper tackles the problem of building an NLP system for Dutch PDSs, leveraging existing successful frameworks like the Linguistic String Project (LSP) and its Restriction Grammar formalism. The goal is to move beyond simple retrieval to a more intelligent analysis that can aid in classifying and understanding medical information.

Methods

The core of the approach is the Restriction Grammar (RG) formalism, a Prolog-based system derived from String Grammar. RG uses context-free production rules augmented with Prolog predicates as restrictions to prune the search space and handle linguistic phenomena.

Key methodological contributions include:

  • Adaptation of RG for Dutch: The paper describes the process of developing an RG grammar for Dutch, starting from an existing English grammar and adapting it to the specificities of Dutch medical terminology.
  • Novel Control Structure: A significant innovation is a new control structure for the parser. Unlike standard depth-first parsers that backtrack, this structure records options that lead to failure or are realized as empty strings. This aims to prevent infinite loops and prune unsuccessful search paths more efficiently, especially when dealing with ambiguity.
  • Semi-Automatic Dictionary Buildup: To address the lack of a comprehensive Dutch medical lexicon, the authors propose a semi-automatic method. This involves ordering technical terms alphabetically by their endings to identify suffixes and their associated grammatical categories. This process is interactive, allowing for corrections and the addition of semantic features.
  • Abstract Data Type for Parse Trees: The paper adopts the data structure for parse trees proposed by Hirschman and Puder (1986), defining it as an abstract data type with specific representation invariants.

Datasets

  • Dutch Patient Discharge Summaries (PDSs): The grammar was initially built and tested on a set of 6 Dutch PDSs.
  • Medical Terminology: A set of technical medical terms was used for the semi-automatic dictionary buildup.

The paper does not explicitly mention a newly introduced benchmark dataset for evaluation, but rather focuses on the development and initial testing of the grammar and parser on a small corpus.

Key Insights and Ideas

  • Effectiveness of RG for Medical Text: The paper reinforces the idea that Restriction Grammar, despite its theoretical roots, can lead to large-scale, useful NLP systems, as demonstrated by the success of the LSP-MLP.
  • Importance of Control Structure for Ambiguity: The proposed control structure is a crucial insight for handling the inherent structural ambiguity in natural language parsing, particularly in specialized domains like medicine. By intelligently recording and utilizing information about failed parsing attempts, it promises more efficient and robust parsing.
  • Regularity in Medical Terminology: The semi-automatic dictionary building highlights the significant regularity in medical terminology, making it amenable to automated or semi-automated lexicon construction. This regularity is a key enabler for developing domain-specific NLP tools.
  • Need for Domain-Specific Analysis: The paper emphasizes that while syntactic analysis is important, a deeper understanding requires accounting for medical sublanguage concepts and potentially deductive reasoning, pointing towards the need for semantic and domain-specific modules.

Key Results

  • Grammar Development: A limited RG grammar for Dutch, consisting of approximately 157 clauses and 41 restrictions, was developed.
  • Parser Implementation: An RG translator-interpreter was implemented, handling parameter bindings differently from DCG.
  • Dictionary Buildup: A semi-automatic method for building a Dutch medical lexicon was developed and implemented, proving to be effective and time-saving due to the regularity of medical terms.
  • Qualitative Findings: The paper demonstrates the application of the RG formalism and the control structure to parse Dutch sentences, showing how the system can navigate and structure linguistic information.

Quantitative results (e.g., parsing accuracy, speedup from the control structure) are not explicitly detailed in the provided text.

Limitations and Future Work

  • Limited Grammar Coverage: The current Dutch grammar is still limited, not yet covering conjunctions, relative clauses, or interrogative sentences.
  • Structural Ambiguity: Structural ambiguity remains a significant challenge, and the paper acknowledges that selecting the correct parse tree from multiple possibilities is an open question.
  • Dictionary Completeness: The semi-automatic dictionary is still a primitive version and requires completion with semantic features. Its current functionality is too limited for full integration into an NLP system.
  • Future Research Directions:
    • Integration of Conjunction Handling: Using a meta-grammar approach.
  • Grammar Extension: Continuously extending the grammar with theoretically sound rules and checking them against a corpus.
  • Refined Restrictions: Developing more and refined restrictions to eliminate dead ends in parsing.
  • Sublanguage Module: Adding a module to handle the syntactic specificities of medical language and rearrange parse trees.
  • Large-Scale Dictionary and Morphological Analyzer: Developing a comprehensive dictionary and a morphological analyzer.
  • Semantic Analysis: Completing syntactic analysis with a semantic counterpart, distinguishing general and medical sublanguage concepts.
  • Deductive Reasoning: Incorporating deductive reasoning for a more intelligent information retrieval system.

Summary updated

BibTeX
@inproceedings{spyns1992applying,
  author = {Peter Spyns and Geert Adriaens},
  title = {{Applying and Improving the Restriction Grammar Approach for Dutch Patient Discharge Summaries}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1992},
}
Figure