Exploring the Language of Data

Proceedings of the 28th International Conference on Computational Linguistics, pages 6638–6648 Barcelona, Spain (Online), December 8-13, 2020 6638 Exploring the Language of Data G´abor Bella University of Trento via Sommarive, 5 38123 Trento, Italy gabor.bella@unitn.it Linda Gremes University of Trento via Sommarive, 5 38123 Trento, Italy linda.gremes@alumni.unitn.it Fausto Giunchiglia University of Trento via Sommarive, 5 38123 Trento, Italy fausto.giunchiglia@unitn.it Abstract We set out to uncover the unique grammatical properties of an important yet so far under- researched type of natural language text: that of short labels typically found within structured datasets. We show that such labels obey a specific type of abbreviated grammar that we call the Language of Data, with properties significantly different from the kinds of text typically addressed in computational linguistics and NLP, such as ‘standard’ written language or social media messages. We analyse orthography, parts of speech, and syntax over a large, bilingual, hand-annotated corpus of data labels collected from a variety of domains. We perform exper- iments on tokenisation, part-of-speech tagging, and named entity recognition over real-world structured data, demonstrating that models adapted to the Language of Data outperform those trained on standard text. These observations point in a new direction to be explored as future research, in order to develop new NLP tools and models dedicated to the Language of Data. 1 Introduction Structured data, such as database tables or XML trees, frequently contain short natural language labels that either describe the data structure itself (attribute names) or provide content (attribute values). In high- precision data integration and analysis applications, where mistakes in interpretation pose a significant risk, such as in health data analytics or emergency response, an unambiguous parsing of such pieces of text is crucial. Conventional NLP tools, such as supervised sequence labellers or embeddings trained on full sentences, do not, however, perform well on structured data. Firstly, data labels may be very short and thus devoid of context necessary for adequate NLP performance. Secondly, they are not self- contained: their interpretation relies on the data structure surrounding them, such as other values of the same attribute or record. And thirdly, their grammar shows significant difference with respect to the language on which NLP tools are typically trained. This paper explores the grammar of structured data labels—that we succinctly designate as the Lan- guage of Data or LoD for short—according to three main hypotheses: (1) uniqueness: it is markedly different from other forms of language typically studied in computational linguistics, such as short so- cial media messages, encyclopaedic articles, literature, or newswire; (2) uniformity: it displays coherent grammatical characteristics across domains, data sources, and even languages; and (3) usefulness: its distinctive features can be exploited to improve the performance of computational language analysis. Following an empirical approach, our paper verifies the first two hypotheses through an extensive cor- pus analysis, and the third one through actually building and evaluating an initial set of NLP tools tailored to the grammar of short text labels. Accordingly, the contributions of this paper are (1) a …

Content truncated to 500 words