Julia Nantke & Nils Reiter
The project CompAnno develops a comparative annotation framework for the exploration and explanation of text similarities regarding the representation of character properties. We focus our approach of comparative annotation on the category of literary character, because it is central to the composition of literary texts in general as well as for establishing, describing and interpreting intertextual relations. In contrast to existing and frequently used annotation approaches, comparative annotation rests on the simultaneous observation and comparison of multiple text snippets. Character properties are initially assigned manually in comparison of the snippets. Collecting annotations for a large number of pairs allows to establish a ranking of similarity, which in turn will be used to train a machine learning system. The comparative annotation process results in a partial ordering for each property. The main benefit of this approach is that besides supporting intertextual annotation it generally allows the annotation of abstract, interpretation-related textual features like the description and characterization of literary characters. We employ various technical methods that offer different perspectives on the task at hand. Basically, three perspectives can be differentiated: 1. If cast as a document retrieval task, a technical system can provide a user with snippets that are similar to the one they are investigating. This reveals (potentially new) intertextual relations, but requires an efficiently computable similarity function, which is constructed from assignments of property values to individual snippets. 2. If the problem is cast as a ranking task, a system can provide a ranking of provided snippets according to some property’s value. This has many other benefits and use-cases, besides comparison, as it provides access to (some aspects of) characterization in general. 3. If similarity itself is taken into focus, a ranking system can be applied to pairs of snippets, and provide, over a set of snippets, a ranking of all the pairs of snippets according to their similarity. In this case, similarity is also caused by the absence or small value of a property (e.g., two snippets can be similar because the characters are not malicious). In all cases, we support the re-integration of these quantitative results into literary studies discourse by providing guiding information for their interpretation. The project thus aims at two concrete outcomes: On the one hand, we develop the first framework for computer-based investigation of intertextuality and text similarity, that goes beyond text re-use and does not assume a fixed corpus. On the other hand, we provide a new way of working with interpretative categories, which are common in computational literary studies.