Introduction
In the field of Natural Language Generation (NLG), the primary goal of Referring Expression Generation (REG) has traditionally been "distinction"—uniquely identifying a target referent. For instance, in a park full of animals, a system might identify a specific dog as "the brown dog" using the minimal set of attributes required for identification. However, human communication often transcends this functional minimum. Authors frequently employ "descriptive" expressions, such as "the rich merchant" instead of "the man," where the additional information ("rich") is not strictly necessary for identification but serves to provide context, set a tone, or introduce information efficiently.
This research addresses the lack of empirical data on these descriptive expressions. By quantifying their prevalence, the authors argue that achieving human-like fluency in AI requires moving beyond mere identification toward more expressive, informative writing.
Methodology
The study employed a rigorous corpus analysis to quantify descriptive expressions. The methodology involved:
- Taxonomy Definition: Referring expressions were divided into Nuclei (the head noun, e.g., king) and Modifiers (variable or removable parts, e.g., the, old).
- Functional Tagging: Each constituent was categorized as either Distinctive (essential for identification) or Descriptive (providing non-essential, extra information).
- Contextual Evaluation: Annotators determined descriptiveness by asking if the referent remained identifiable after removing the constituent, or if the word choice was intentionally informative.
- Annotation Pipeline: Using the "Story Workbench" tool, two native English speakers annotated the text, followed by an adjudication phase to establish a "gold standard" dataset.
Corpus Analysis
The study utilized a corpus of 62 texts, split between narrative (30 folktales) and news (32 Wall Street Journal articles). This selection allowed for a comparison between "lyrical" fiction and professional journalistic prose.
As shown in Figure 1, the corpus comprised over 24,000 words and 7,207 referring expressions. Notably, the prevalence of descriptive expressions was remarkably consistent across genres: 17% in tales and 19% in news articles, resulting in an overall average of 18%.
Key Results and Insights
The analysis revealed that description is a pervasive feature of high-quality human communication. A significant finding is that description is not limited to adjectives; the choice of the head noun (nucleus) itself can be descriptive. For example, referring to a previously mentioned character as "the villain" identifies the referent while simultaneously adding characterization.
Detailed statistics in Figure 3 highlight that modifiers are the primary vehicle for description, with 20% of all modifiers serving a descriptive function. While less frequent, 4% of nuclei also performed a descriptive role. The data also shows that while the maximum number of nuclei per reference is one, modifiers are more varied, with an average of 0.8 per reference and up to 6 in some cases.
The inter-annotator agreement yielded a Kappa score of 0.59 for the function of referring expressions. This moderate agreement reflects the inherent subjectivity of language, where the distinction between "necessary context" and "extra flavor" can vary between readers.
Conclusion and Future Work
The prevalence of descriptive expressions suggests that current REG algorithms, which focus solely on distinction, ignore nearly one-fifth of human referring behavior. Future work in NLG should focus on Aggregation—the process of folding multiple ideas into a single sentence. Descriptive referring expressions are a vital tool for aggregation, allowing facts to be tucked into references rather than requiring separate sentences. Improving "coherence planning" in AI will be essential for generating text that matches the density and narrative flow of human-authored prose.
Summary updated