Language choice models for microplanning and readability

Sandra WilliamsNAACL 2003

TL;DR

This paper introduces a novel approach to building language choice models for Natural Language Generation (NLG) systems, specifically for tailoring text to users with varying reading abilities. By analyzing corpus data and adapting it based on psycholinguistic evidence, the authors create constraint satisfaction problem (CSP) graphs that allow NLG systems to make more readable microplanning decisions, like choosing discourse cues and sentence structure.

What Problem Does This Paper Address?

Generating text that is not only grammatically correct and coherent but also readable by a specific audience is a significant challenge in Natural Language Generation (NLG). This is particularly true when aiming to produce texts for individuals with lower literacy levels, such as adults who may have missed schooling or have learning disabilities. Existing NLG systems often generate text optimized for "good readers," which can be difficult for those with poorer reading skills to comprehend. The authors tackle this by focusing on the "microplanning" stage of NLG, where decisions about sentence structure, discourse relations, and the use of cue phrases (like "for example" or "however") are made. They aim to create models that can adapt these choices to improve readability for "poor readers."

Methods

The core of the paper describes the construction of "language choice models" for the microplanner. This process involves several key stages:

  • Corpus Analysis: The researchers analyzed seven discourse relations from the RST Discourse Treebank corpus, focusing on features that influence readability. These features include the length of text spans, the ordering of spans, the presence and position of discourse cue phrases, and punctuation.
  • Model Representation: They represent these linguistic choices using Constraint Satisfaction Problem (CSP) graphs. Each node in the graph represents a feature (e.g., length, punctuation), and the connections between nodes represent constraints derived from the corpus analysis. This allows for modeling the interdependencies between different linguistic choices.
  • Adaptation for Poor Readers: Since the initial corpus was written for good readers, the models were adapted for poor readers. This was achieved by "tightening" the constraints within the CSP graphs, based on psycholinguistic evidence and preliminary experiments. This tightening favors choices known to improve readability for less skilled readers, such as using more common discourse cue phrases and specific punctuation patterns.
  • CSP Solver: A Java Constraint Library is used to find solutions within the CSP graphs. The solver finds all possible valid combinations of features that satisfy the constraints.
  • Filtering: A filter module then selects the most appropriate solution from the CSP solver's output, prioritizing shorter sentences for poor readers.

Datasets

  • RST Discourse Treebank corpus: Used for the initial corpus analysis to identify linguistic features and their patterns in text written for good readers.
  • British National Corpus (BNC): Used to determine the frequency of discourse cue phrases.
  • Dolch lists: Used as a reference for common function words that beginner readers are taught to sight-read, influencing the selection of cue phrases for poor readers.

Key Insights and Ideas

  • Discourse-level choices are crucial for readability: The paper highlights that decisions made at the discourse level (how sentences are connected, what cue phrases are used) have a significant impact on text comprehension, especially for readers with lower literacy.
  • CSP graphs are a powerful representation for modeling linguistic choices: The use of CSP graphs effectively captures the complex interdependencies between various linguistic features, allowing for a more nuanced modeling of language generation.
  • "Tightening" constraints can adapt models for specific audiences: The core innovation lies in adapting existing models for "poor readers" by making the constraints within the CSP graphs more restrictive, favoring features known to enhance readability.
  • Trade-offs between conciseness and comprehensibility: The authors acknowledge that strategies like aggregation (combining multiple ideas into a single sentence) for conciseness, which are generally favored in NLG, might actually hinder comprehension for poor readers. This suggests a need to balance conciseness with clarity.
  • Psycholinguistic evidence is key for adaptation: The adaptation process relies heavily on existing research and new experiments to understand how different linguistic features affect readers with varying skill levels.

Key Results

  • The paper successfully constructed language choice models for seven discourse relations using CSP graphs.
  • These models were adapted to favor features that improve readability for poor readers, such as selecting common discourse cue phrases and specific punctuation.
  • The adaptation process involves "tightening" constraints in the CSP graphs, leading to solutions that prioritize features beneficial for less skilled readers.
  • Preliminary experiments suggest that these adaptations can lead to more readable texts for poor readers.

Limitations and Future Work

  • Limited corpus size: The authors acknowledge that the corpus analysis was based on a limited number of instances and plan to expand its size.
  • Manual data reconfiguration: The current process of reconfiguring data for CSP graphs is manual, and they are investigating ways to automate it.
  • Preventing incoherent solutions: The current microplanner can sometimes produce incoherent solutions, and future work will focus on developing mechanisms to prevent this, potentially through backtracking and considering coherence factors like focus and reference.
  • Deeper processing: Generating embedded discourse text spans and mid-text-span cue phrases requires deeper syntactic and semantic processing, which is not yet implemented.
  • Further reading experiments: The authors plan to conduct more extensive reading experiments with both good and bad readers to empirically validate the effectiveness of their adapted models.
  • User satisfaction evaluations: They also plan to conduct user satisfaction evaluations and gather feedback from professional tutors.
  • Automating analysis and reconfiguration: Future work will focus on automating parts of the corpus analysis and data reconfiguration process.

Summary updated

BibTeX
@inproceedings{williams2003language,
  author = {Sandra Williams},
  title = {{Language choice models for microplanning and readability}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2003},
}
Figure