Building Literary Corpora for Computational Literary Analysis - A Prototype to Bridge the Gap between CL and DH

LitText: Building Literary Corpora for Computational Literary Analysis A Prototype to Bridge the Gap between CL and DH Andrew U. Frank, Christine Ivanovic Geoinformation TU Wien, Comparative Literature University Vienna Gusshausstr. 27-29 1040 Wien Austria, Sensengasse 3a 1090 Wien Austria frank@geoinfo.tuwien.ac.at, christine.ivanovic@univie.ac.at {frank, christine}@gerastree.at Abstract The design of LitText follows the traditional research approach in digital humanities (DH): collecting texts for critical reading and underlining parts of interest. Texts, in multiple languages, are prepared with a minimal markup language, and processed by NLP services. The result is converted to RDF (a.k.a. semantic-web, linked-data) triples. Additional data available as linked data on the web (e.g. Wikipedia data) can be added. The DH researcher can then harvest the corpus with SPARQL queries. The approach is demonstrated with the construction of a 20 million word corpus from English, German, Spanish, French and Italian texts and an example query to identify texts where animals behave like humans as it is the case in fables. Keywords: linked-data, semantic-web, RDF, SPARQL, digital humanities 1. Introduction Literary analysis, like all other fields of Digital Humani- ties (DH) based on the evaluation of texts in natural lan- guages, can benefit from advances in computational lin- guistics (CL) (Pustejovsky et al., 2017a). Many language tools for linguistic analysis are available, and produce re- sults which could be useful for corpus-based literary analy- sis. Moretti (2005) and others have demonstrated viable ap- proaches to computational literary analysis, and have pro- duced valuable results (e.g. (Fischer et al., 2017) ). Projects like DARIAH (Bl¨umm et al., 2016), LAPSS (Ide et al., 2015), and others have brought together CL tools in uni- fied frameworks to make the results of CL easier to use. Despite this, the uptake of CL in literary studies has been slow so far, and very few of the immense number of literary publications show use of CL. More efforts in bridging the gap between CL and DH are required. It could be beneficial to take a step back from technical issues, and instead focus on the research practices of literary studies researchers. These usually consist of (1) selecting texts which will be analyzed; and (2) reading the texts and identifying parts with properties relevant to the research goal by underlining them. The approach is proba- bly similar for most text-based DH. A tool for DH should therefore be presented following this model of (1) building a corpus and (2) selecting text pieces of interest. CL can identify parts with specific text properties in a cor- pus quicker than a human reader, it can process texts sys- tematically, and it can browse a corpus larger than any hu- man could peruse. The DH researcher has only to describe what properties the relevant text parts should have. Lord et al. (2006) followed an approach where the researcher gives some example texts she is interested in, and the search en- gine then finds all text similar to the given one, with the risk that the automatic detection misses the intention of …

Content truncated to 500 words