Estimating Numerical Attributes by Bringing Together Fragmentary Clues

Hiroya Takamura, Jun'ichi Tsujii

Abstract

Artificial intelligence systems often lack basic physical grounding, such as the typical dimensions of everyday objects. This paper presents a method for automatically estimating the physical size of objects by mining "fragmentary clues" from large-scale text corpora. By integrating absolute measurements (regression) and relative comparisons (ranking) into a joint mathematical framework, the proposed model successfully estimates dimensions across a wide range of scales. Experimental results demonstrate that combining these disparate data sources significantly improves estimation accuracy, particularly when direct measurements are scarce.

Introduction

A fundamental gap in current artificial intelligence is the lack of "common sense" regarding the physical world. While humans intuitively understand that a surfboard cannot fit inside a jacket pocket, computers often lack this basic grounding. This deficiency hinders performance in tasks such as natural language processing (NLP) and computer vision, where numerical knowledge can serve as a vital constraint for disambiguation and reasoning.

Comprehensive numerical data is rarely available in structured formats. While Wikipedia might provide the height of the Eiffel Tower, it seldom lists the average size of a hand towel. To bridge this gap, we propose mining "fragmentary clues"—scattered hints in massive text corpora that, when aggregated, reveal the dimensions of the physical world.

Methodology: Combined Regression and Ranking

The core of our approach is a Combined Regression and Ranking framework. We represent physical objects (synsets from WordNet) as feature vectors and model their size as a linear function of these features. The model learns from two primary types of data:

  1. Absolute Clues (Regression): Explicit measurements found in text (e.g., "a 10cm blade"). The model uses a squared loss function to minimize the discrepancy between predicted and observed values.
  2. Relative Clues (Ranking): Comparisons between objects, which can be explicit ("X is larger than Y"), implicit ("I put X into Y"), or structural (WordNet holonymy relations). A hinge loss function ensures the predicted size of X exceeds that of Y.

A trade-off parameter, $\alpha$, balances these two objectives. When $\alpha = 0$, the model relies solely on regression; as $\alpha$ increases, the influence of relative ranking constraints grows.

Results and Discussion

The experimental results confirm that the synergy of absolute and relative clues significantly enhances model performance.

Accuracy of relative size estimation as a function of the trade-off parameter $\alpha$. (a) shows the full model, while (b) excludes certain WordNet features.

As shown in the accuracy plots, incorporating ranking data ($\alpha > 0$) consistently improves the system's ability to correctly order object pairs. The system achieved an accuracy range of 70% to 80% when all features were utilized, as seen in Figure (a) above.

Average difference (log scale) between predicted and actual sizes as a function of $\alpha$. Lower values indicate better regression performance.

Furthermore, the regression error (average difference) decreases as relative clues are introduced. This "Alpha effect" is most pronounced when the number of training samples (indicated by the legend 300, 500, 800, 1000) is small. This suggests that the web of comparisons acts as a scaffolding, guiding the model toward correct numerical values even when direct measurements are sparse.

Conclusion

By stitching together fragmentary clues from text, we can effectively teach machines the typical sizes of physical objects. The combination of regression on absolute values and ranking on relative comparisons proves robust across various scales, from microscopic cells to entire airports. Future work will explore applying this framework to other numerical attributes such as weight, temperature, and price.

Summary updated

Abstract

This work is an attempt to automatically obtain numerical attributes of physical objects. We propose representing each physical object as a feature vector and representing sizes as linear functions of feature vectors. We train the function in the framework of the combined regression and ranking with many types of fragmentary clues including absolute clues (e.g., A is 30cm long) and relative clues (e.g., A is larger than B).
BibTeX
@inproceedings{takamura2015estimating,
  author = {Hiroya Takamura and Jun'ichi Tsujii},
  title = {{Estimating Numerical Attributes by Bringing Together Fragmentary Clues}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2015},
  doi = {10.3115/v1/n15-1143},
}
Figure