A Corpus-based Analysis for the Ordering of Clause Aggregation Operators

James ShawCOLING 2002

TL;DR

This paper investigates the optimal ordering of clause aggregation operators in natural language generation (NLG) by analyzing manually annotated text. They propose a specific ordering, prioritizing hypotactic (subordinating) operators before paratactic (coordinating) ones, and find that this order effectively reconstructs human-written sentences, suggesting its utility for NLG systems.

What Problem Does This Paper Address?

Generating grammatically correct and natural-sounding sentences often involves combining multiple simpler clauses into a more complex sentence. This process, known as clause aggregation, relies on various "operators" (like conjunctions, adjectives, and relative clauses) that transform these clauses. The challenge lies in determining the correct order in which these operators should be applied. Applying them in the wrong sequence can lead to ungrammatical sentences or alter the intended meaning. Previous work suggested heuristics for this ordering, but this paper aims to provide empirical evidence from a corpus-based analysis to validate a specific ordering.

Methods

The core of this research is a corpus-based analysis. The authors manually annotated two corpora (medical domain and Wall Street Journal) to evaluate their proposed ordering of clause aggregation operators. The annotation process involved:

  1. De-aggregating original sentences: Breaking down complex sentences into their constituent propositions.
  2. Specifying rhetorical relations: Identifying the logical or pragmatic relationships between these propositions (e.g., concession, elaboration).
  3. Specifying transformation sequences: Determining the sequence of aggregation operators that would reconstruct the original sentence from the de-aggregated propositions.

A novel aspect introduced is the "proposition set" (propset) concept. This hierarchical grouping of propositions helps clarify the scope of modifying clauses, addressing ambiguity that arises when a single proposition modifies multiple others. This concept proved crucial for accurate annotation and for simplifying the analysis. The authors then evaluated their proposed ordering by checking if the annotated transformation sequences adhered to it.

Datasets

The study used two corpora:

  • A corpus from a medical domain.
  • A corpus from the Wall Street Journal.

For each domain, 100 sentences were annotated, totaling 200 sentences for the evaluation. These sentences were specifically selected because they contained the conjunction "and," increasing the likelihood of encountering both paratactic and hypotactic transformations.

Key Insights and Ideas

  • Ordering Matters for Meaning: The paper clearly demonstrates with examples (like John drinking cider and eating oranges) how the order of aggregation operators directly impacts the meaning and scope of modification in the resulting sentence.
  • Hypotactic Before Paratactic: The proposed ordering prioritizes hypotactic operators (like relative clauses) over paratactic ones (like simple conjunctions). This seems to align with producing more concise and natural surface expressions.
  • The Power of "Propset": The introduction of the "proposition set" (propset) is a significant contribution. It elegantly handles scope ambiguity in complex sentence structures, making the annotation process more robust and the analysis of rhetorical relations more precise. This hierarchical grouping allows for clearer representation of how propositions relate to each other.
  • Corpus Validation is Key: Relying on human-annotated data provides strong empirical backing for the proposed ordering, moving beyond theoretical heuristics.

Key Results

The evaluation showed that the proposed ordering of clause aggregation operators was highly effective. Out of 200 annotated sentences, 195 could be resynthesized using the proposed ordering. This high success rate, particularly after incorporating the propset concept, provides strong evidence that the proposed ordering is effective for reconstructing human-written sentences. The authors also noted that unsupported cases were rare, further validating their approach.

Limitations and Future Work

  • Limited Scope of Operators: The study focused on specific types of clause aggregation operators (conjunction, adjective, relative clause transformations). Transformations involving extraction or certain rhetorical relations were not fully handled by their system, though they were mapped to similar types for evaluation.
  • Bias in Corpus Selection: The selection of sentences containing "and" might introduce a bias, potentially favoring sentences with operators that the system handles well.
  • Future Work: The authors suggest that further research could explore the ordering of operators when the same operator is applied multiple times (e.g., sequences of adjective transformations), an issue they acknowledge is outside the scope of this particular analysis. They also highlight the potential for integrating this ordering into NLG systems to produce more grammatical and concise outputs.

Summary updated

BibTeX
@inproceedings{shaw2002corpus,
  author = {James Shaw},
  title = {{A Corpus-based Analysis for the Ordering of Clause Aggregation Operators}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2002},
}
Figure