Abstract
Most Natural Language Generation (NLG) systems are designed for highly literate audiences, often leaving low-skilled readers behind. This paper introduces an "Enhanced Readability" (ER) model for microplanning—the stage of text generation that decides how to express relationships between ideas. By optimizing cue phrases, sentence structure, and ordering, the researchers demonstrated that specific linguistic choices can make generated reports significantly easier and faster to read for those with lower literacy levels.
Introduction: The Microplanning Challenge
While we often focus on what a computer says, how it says it is just as critical—especially for the millions of adults with below-average literacy. Most existing NLG systems assume a "proficient reader" baseline. When these systems generate complex discourse structures—how ideas like "cause and effect" or "concessions" are linked—they often produce text that is difficult for low-skilled readers to parse.
The authors address the "microplanning" gap: the middle step in the NLG pipeline where abstract content plans are turned into specific linguistic structures. Specifically, they look at how to express discourse relations (e.g., "Although X, Y" or "If A, then B") so that the cognitive load on the reader is minimized.
Methodology: ER vs. Control Models
The researchers developed two competing models for a microplanner:
- The Enhanced Readability (ER) Model: A set of rules hypothesized to help low-skill readers. These rules govern cue phrases (e.g., "but" vs. "although"), ordering of logical components, and sentence structure (splitting complex ideas into separate sentences).
- The Control Model: This model followed "standard" behavior, using the most common choices found in the RST-DTC (a large corpus of human-annotated discourse relations).
The system takes a "content plan"—a tree-like structure where messages are linked by relations like Concession or Condition—and uses a constraint-based approach to select the best linguistic expression based on the active model.
The evaluation utilized feedback reports generated for the GIRL and SkillSum projects, which assess adult basic literacy skills. The RST-DTC (Rhetorical Structure Theory Discourse Treebank) was used to derive the Control Model by identifying the most common ways humans naturally express discourse relations in general text.
Key Insights: "Common" is not "Accessible"
The research highlights that the most frequent ways people write (captured in the Control Model) are not necessarily the easiest to read. Standard writing often uses complex, nested structures that can trip up readers with lower literacy.
The "glue" matters: It isn't just about using simple words (lexical choice); it’s about the "discourse glue." How you connect two simple sentences can drastically change how easily a reader understands the logical link between them. Furthermore, structural simplicity is key; the ER model generally favored splitting complex relations into separate sentences and using clear, explicit cue phrases.
Results and Evaluation
The researchers conducted two main evaluations. In the GIRL Experiment, while low-skilled readers generally preferred the ER texts and read them faster with fewer errors, the results were not statistically significant due to high variance in literacy levels.
However, in the SkillSum Experiment, by narrowing the focus to a more homogenous group with moderate literacy problems, the results became clear. Subjects read the ER version 9% faster than the control version (p=0.04). There was also a "weakly significant" reduction in reading errors (p=0.058).
The researchers also found that silent reading tests are unreliable for this demographic because poor readers often "skim" to hide their difficulties. Measuring "reading aloud" time and errors provided a much more accurate picture of text accessibility.
Conclusion and Future Work
While the 9% speed increase is statistically significant, the authors note they have only "scratched the surface" of how microplanning affects comprehension. Future work includes:
- Lexical Choice: Integrating simpler synonyms (vocabulary) into the models.
- Personalization: Tailoring text even more precisely to a specific individual’s literacy profile.
- Scaling: Testing more complex discourse relations beyond Concession and Condition.
This study demonstrates that intentional microplanning can significantly improve information accessibility for low-skilled readers, moving beyond simple word-level simplification to structural optimization.
Summary updated