Simple Neologism Based Domain Independent Models to Predict Year of Authorship

Proceedings of the 27th International Conference on Computational Linguistics, pages 202–212 Santa Fe, New Mexico, USA, August 20-26, 2018. 202 Simple Neologism Based Domain Independent Models to Predict Year of Authorship Vivek Kulkarni Department of Computer Science University of California, Santa Barbara vvkulkarni@cs.ucsb.edu Yingtao Tian Department of Computer Science Stony Brook University yittian@cs.stonybrook.edu Parth Dandiwala Department of Computer Science Stony Brook University pdandiwala@cs.stonybrook.edu Steven Skiena Department of Computer Science Stony Brook University skiena@cs.stonybrook.edu Abstract We present domain independent models to date documents based only on neologism usage pat- terns. Our models capture patterns of neologism usage over time to date texts, provide insights into temporal locality of word usage over a span of 150 years, and generalize to various domains like News, Fiction, and Non-Fiction with competitive performance. Quite intriguingly, we show that by modeling only the distribution of usage counts over neologisms (the model being agnostic of the particular words themselves), we achieve competitive performance using several orders of magnitude fewer features (only 200 input features) compared to state of the art models some of which use 200K features. 1 Introduction Determining when a document is written is an important task in historical linguistics and temporal infor- mation retrieval. For instance, several works attempt to date historical biblical texts like the The Book of Isiah (Rooker, 1996; Ehrensv¨ard, 1997; Hurvitz, 2000; Young and Rezetko, 2016) or ancient texts like Beowoulf (Chase, 1997). Likewise, in the field of information retrieval, establishing the dates of docu- ments is an important pre-requisite to returning temporally relevant documents and provides important information for a large number of search tasks (Ostroumova Prokhorenkova et al., 2016; Efron, 2013). Most efforts to automatically date texts adopt a learning based approach and rely on several linguis- tic features that are time-relevant (Garcia-Fernandez et al., 2011; Jatowt and Tanaka, 2012; Zampieri et al., 2015; Ostroumova Prokhorenkova et al., 2016). Such time-relevant features include neolo- gisms/archaisms, political events, spelling variations, and the presence of named entities as well as ex- ternal knowledge bases. In addition to the large input feature dimensionality (of the order of the input vocabulary), these approaches all focus on particular domains (for example, primarily News articles). While such feature rich models which are arguably domain specific perform very well, these features are rarely generalizable across domains (for example. models that use political events can perform very well when evaluated in the News domain, but do not necessarily generalize to Fiction). Developing models that generalize across domains without further fine tuning generally entails modeling linguistic cues that are stable across domains.1 Consequently, in this work, we propose the first set of models that effectively generalize across do- mains by leveraging insights into the temporal usage of language. Our models rely on a key insight which we term the temporal locality of neologisms: Documents written at time t tend to use neologisms invented shortly before t. By effectively modeling this usage of neologisms (as a feature class cumula- tively) we propose models that not only reduce our …

Content truncated to 500 words