Estimating Numerical Attributes by Bringing Together Fragmentary Clues

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 1305–1310, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Estimating Numerical Attributes by Bringing Together Fragmentary Clues Hiroya Takamura and Jun’ichi Tsujii Tokyo Institute of Technology Microsoft Research Asia takamura@pi.titech.ac.jp jtsujii@microsoft.com Abstract This work is an attempt to automatically ob- tain 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 com- bined 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). 1 Introduction We know how large surfboards usually are and also that an inner pocket of any jacket is much smaller than a surfboard. Since we know about these nu- merical attributes, nobody of sound mind has prob- ably ever tried to vainly put a surfboard into an in- ner pocket of a jacket. However, computers do not have comprehensive knowledge of this sort. This lack of comprehensive knowledge of the numerical attributes is one obstacle to flexible and natural man- machine communication. This work is an attempt to automatically obtain knowledge on numerical at- tributes so that computers can use it. The knowledge on numerical attributes is also very useful on many other occasions. For exam- ple, it enables computers to alert their users when the users input incorrect numbers that are outside of the normal range of the attribute. In image recogni- tion, a large red object will unlikely be recognized as a strawberry if the computer knows its normal size. In natural language processing, QA systems can use numerical knowledge to eliminate the out-of-range answer candidates to numerical questions. A number of attempts similar to the current work have been made in some other fields such as psy- chology or fuzzy theory. However, such attempts heavily rely on costly experiments such as giving questionnaires to human subjects and have a prob- lem in their scalability. In contrast, the current work attempts to use NLP techniques on large text data both online and offline in order to obtain such knowledge without relying on costly experiments such as questionnaires. A possible criticism of this project is that simply examining an existing knowl- edge source such as Wikipedia might accomplish this purpose without much effort. Indeed Wikipedia provides numerical information of some physical objects, but not all. For example, the Wikipedia page for watches provides descriptions on their function and their history, but no information on their size. Clues to the numerical attributes are rather scat- tered over corpora and other linguistic resources. In a corpus, we can find informative descriptions such as “X is 35cm tall”. We can also find text fragments suggesting an order relation between two physical objects with regard to the size as in “X is larger than Y”, as well as implicit …

Content truncated to 500 words