another story, and iterate the process (see Figure 6). The system keeps the user-selected landmarks as aggregated counts that can be used by authorized story “curators” to update the existing hypergraph for future analysis. This incorporation of scholarly feedback is essential for the system’s performance, as it allows for the aggregation of accumulated expert knowledge, which in turn helps guide researchers as they navigate the hypergraph.

The graph-theoretic techniques we applied to generate the final hierarchical graph map involve four major tasks:

˲Define similarity measures in stories;

˲ Construct a weighted graph among stories, where the weight of the connection between two stories is obtained through order-preserving transformations of their similarity measures;

˲ Decompose the graph(s); and

˲ Compute a “hierarchy tree” for the obtained graph(s).

The similarities measures we used are based on several algorithms, including: Hellinger distance, 23 weighted Jaccard coefficient, 19 cosine similarity, 18 and scholar-defined weights. The measures are transformed through Gaussian kernels and other standard kernels20, 30 that assign to every pair of stories that share at least a minimum set of attributes a corresponding similarity weight. In general, the graphs obtained through these transformations vary depending on the type of similarity measure used. Consequently, we obtain a final weight between two stories by computing a norm of their similarity-weights vector. Because the obtained graph has different density “regions,” we follow the approach described in earlier work by Abello et al., 2 decomposing the collection of stories into maximal sub-graphs according to their inherent k-connectivity. Each obtained sub-graph is then clustered using an adapted version of Markov Clustering. 2 The entire process is encoded in a data structure called a “hierarchy tree” visually represented in the user interface as a hierarchical graph map (see Figure 3), 2 constituting one of the central modes of interaction between user and story corpus. Worth noting is that each cluster includes a unique associated label string that encodes its “semantic” placement in the hierarchy

the trouble
with house elves
begins with
their unpredictable
nature, which,
for folklorists,
makes them
difficult to track
through the
landscape of
the story corpus.

tree. These cluster labels can be used to help provide and receive feedback from scholars studying the corpus, as well as judge the effectiveness of the suggested classification.

Performance on the
two Folklore tasks

The first task we included in the system was to place the target story in a neighborhood of closely related ghost stories. The system succeeded admirably, moving the story from a group of stories about manor lords in the original classification scheme to a neighborhood of stories about ghosts and other supernatural beings that haunt workers in farm buildings and barns— a new implicit category that did not exist in Tang Kristensen’s original indices. Although we left success criteria for this work fuzzy, we placed the target story in the same part of the organized meta-graph as several ghost stories exhibiting certain topical similarities; for example, another story where a haunt makes it impossible to use part of the farm appears nearby in the hypergraph: “A woman died in a farm in Dokkedal… and she showed herself every noon at a specific place in one of the rooms and always as a black shape. Because of that, the room stood empty for a long time, since nobody dared go in there...” 33

Perhaps more interesting is the system’s ability to meet the second task— suggest candidate stories, given the researcher’s interest in a target story. In this case, the visual representation of the meta-network reveals a close affiliation of the target story with a story not classified by Tang Kristensen as a story about ghosts but as a story about house elves (nisse), a category of supernatural beings not usually considered thematically related to ghosts: “When they got home, the farmhand was happy because now he’d gotten something to use for feed, and afterward nis [a house elf] could go and feed the animals just as much as he wanted to. Then they got another farmhand, and he didn’t want to let him go on like that. But he got lifted up in his bed and all the way up to the rafters, so he lay there dead when people got up the next morning.” 33

The intersection between the target story and this latter story, likewise not