Statistics-Based Lexical Choice for NLG from Quantitative Information

Xiao Li, Kees van Deemter, Chenghua Lin

Introduction

In Natural Language Generation (NLG), one of the most persistent challenges is the "lexical choice" problem: determining the most appropriate natural language descriptor for quantitative data. For instance, deciding whether a wind speed of 5 knots should be described as "light," "calm," or "mild." This paper, presented at the 9th International Natural Language Generation conference, proposes a statistical framework to automate this process, moving beyond traditional rule-based systems.

The Problem: Mapping Numbers to Words

Historically, lexical choice has relied on manual coding, where experts define strict "if-then" boundaries for words. However, this approach is brittle and fails to capture the inherent vagueness and subjectivity of human language. Furthermore, many descriptors are multi-dimensional; a word like "mild" might depend on both temperature and wind speed. Traditional models often struggle to learn these overlapping, multi-dimensional relationships from data alone.

Methodology: Shared Vector Spaces

Researchers at the University of Aberdeen developed a methodology inspired by "contour stylization" to map both raw data and linguistic expressions into a shared vector space. The process involves three key steps:

  1. Data Discretization: Continuous data (e.g., wind speed) is mapped to "key-points." A specific value is then converted into a vector of weights using linear interpolation.
  2. Word Representation: Words are mapped into the same vector space. A word's vector reflects its association with specific data ranges based on a training corpus.
  3. Lexical Choice: To select a word for a new data point, the system calculates the Cosine Similarity between the data vector and all candidate word vectors. The word with the highest similarity score is selected.
Individual prediction probabilities based on single dimensions: wind direction, wind speed, and wind direction variance.

Experimental Setup

The system was trained and tested using the SumTime-Meteo Corpus, which contains human-written weather forecasts paired with meteorological data. The study focused on 144 records, summarizing data into three dimensions: Wind Speed, Wind Direction, and Wind Variance.

Results and Discussion

The authors conducted a "sanity check" by comparing the model's predictions against actual word usage in the test set. The model was evaluated on its ability to assign higher occurrence probabilities to words in "Group 1" (where the word was actually used by humans) versus "Group 2" (where it was not).

Occurrence probabilities for the combined multi-dimensional model, showing significantly improved accuracy in distinguishing word usage groups.

The results demonstrated that while individual dimensions provided some predictive power, the model performed significantly better when considering all dimensions simultaneously. This multi-dimensional approach was particularly effective for vague modifiers like "variable" or "mainly," which rely on the interaction between multiple data points to be used correctly. As shown in the comparison between single-dimension and combined models, the multi-dimensional integration provides much clearer separation between actual usage and non-usage groups.

Conclusion

This research offers a robust statistical bridge between quantitative information and natural language. By treating data and language as part of the same vector space, the system naturally handles vagueness and multi-dimensional dependencies. While currently limited by a small sample size and a specific domain (meteorology), the approach shows great promise for scaling to more complex NLG tasks such as financial or medical reporting.

Summary updated

BibTeX
@inproceedings{li2016statistics,
  author = {Xiao Li and Kees van Deemter and Chenghua Lin},
  title = {{Statistics-Based Lexical Choice for NLG from Quantitative Information}},
  booktitle = {INLG},
  year = {2016},
  doi = {10.18653/v1/w16-6618},
}
Figure