A topic modelling tool was adapted to requirements for a collection of Swedish folk legends. To offer an overview of a list of folk legend texts, which had been automatically extracted by the topic modelling tool, snippet text versions of the folk legends were displayed. The snippets were constructed from the full-text versions of the legends using the sentences most relevant to the topics extracted by the topic modelling algorithm. In addition, collection-adapted data was constructed for performing a pre-processing of the folk legend texts, before they were submitted to the topic modelling algorithm. This data consisted of a collection-adapted stop word list and word lists for improving the quality of clusters of semantically similar words.
Topic modelling has been shown to be a useful text mining technique for automatically
identifying recurring information in large text collections. The technique has been
applied on text collections belonging to a wide range of genres, e.g. news paper text [
The topic modelling algorithm automatically analyses a text collection in search for
topics that often recur in the text. There are a number of topic modelling visualisation
tools, for instance, visualisations that focus on relations between the automatically
detected topics and the terms that represent these topics [
Copyright © 2021 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
Topics2Themes has previously been applied to posts in discussion forum threads
[
Texts extracted by the topic modelling algorithm as most closely associated with the topics detected are displayed by Topics2Themes in a scrollable list. By scrolling through this text list, the user is meant to be given an overview of the content extracted. However, despite most of the folk legend texts being relatively short, many of them are longer than the very short types of texts for which the tool was originally developed. This has the effect that the overview, meant to be provided by the scrollable list, is lost. To achieve the same overview in those cases, we therefore added the functionality to – instead of displaying the full texts in the scrollable list – display a short snippet of the text. In addition to this technical adaption of the tool, we also (i) further adapted the stop word list to the folk legends collection by adding additional collection-specific words, and (ii) performed a manual correction of the automatically constructed concept clusters. 2
Topics2Themes was applied on a text collection containing around 10,000 records of folklore from the archive at the Institute for Language and Folklore in Gothenburg. The material mainly consists of folk legends, i.e. typically narratives about our perceptions of the world around us, about events in the past, and about everyday life manifestations of that, which we perceive as supernatural. The legends are usually relatively short and monoepisodic, have a fixed structure, and have been transmitted verbally from one generation to another.
Most of the records have been collected during the Interwar Period, through interviews with older people in the countryside in western Sweden. The records either contain (manual or HTR-based) transcriptions of handwritten text, or text obtained from printed material using OCR. Most of the material is written in Standard Swedish, although occasional dialectal expressions, as well as words written with older spelling variants, occur in the texts.
The Topics2Themes tool uses the topic modelling algorithm Non-negative matrix
factorisation [
Figure 1 shows the graphical user interface of Topics2Themes, when applied on the folk legend collection. The tool displays information in four different panels. The Topics panel, i.e. the second panel from the left, lists the topics that have been automatically extracted by the topic modelling algorithm. A topic element in the Topics panel is represented by a list element that contains the three terms most closely associated with the topic, e.g. “kyrka – bygga – stannade” (“church – build – stayed”). The Terms and Texts panels list the terms and texts, respectively, that are associated with the extracted topics. We configured the tool to retrieve 10 terms and 40 texts for each topic. That is, for each topic detected, the Terms panel displays 10 terms, and the Texts panel displays 40 texts. Term-topic and topic-text associations are visualised by lines that connect the list elements. Finally, to the Themes panel (only shown partly in the figure), the user can add recurring themes that are identified when the extracted texts are manually analysed. That is, while the first three panels stem from an automatic process, the content of the Themes panel is to be manually added by a user.
The stop word list that was used for pre-processing the texts in the folk legends collection has been iteratively expanded. This process consists of first using Topics2Themes to construct a topic model, and thereafter of using the Terms panel to examine terms extracted (as well as the Texts panel to examine their textual contexts), in order to determine which terms are suitable as stop words. The process is then started over, using a list expanded with the newly detected stop words when constructing the new model. This iterative process has so far resulted in a stop word list consisting of 542 words.
For creating clusters of semantically similar words, we used a pre-trained word2vec continuous skipgram model2 and the automatic clustering functionality of the Topics2Themes tool. We then used the functionality provided by the tool for manually correcting the automatically constructed clusters. This is carried out by providing the tool with two types of word lists, (i) a list of words that are to be removed from the automatic clusters (since the automatic process assigns them to a semantic cluster to which they do not belong), and (ii) a list of manually constructed word clusters. Our manually constructed clusters were often automatic clusters that we expanded with additional, semantically similar words, or automatic clusters that we divided into smaller, more semantically coherent groups of words. The construction of the word lists for improving the clustering is also an iterative process. We have so far detected 571 words to exclude from the automatic clustering, and constructed 118 manual clusters based on the automatic ones. A total of 369 concept clusters were used when pre-processing the texts before sending them to the topic modelling algorithm. Examples of concept clusters are shown in the Terms panel. For instance, the first element consists of different inflections of “church”, and the tenth element consists of a large cluster of words referring to body parts. The advantage of using semantic clustering instead of methods based on morphology – e.g. lemmatisation – is that semantically similar words can be conflated into one concept, regardless of their lexical instantiations.3 2 The model has been trained on the Swedish CoNLL17 corpus – which contains 3,010,472 tokens – by the Language Technology Group at the University of Oslo.
The model is available at: http://vectors.nlpl.eu/repository/. 3 The topic model configuration, as well as the manually constructed word list and clusters is available at: https://github.com/mariask2/swedish-folk-legends. Fig. 1. Topics2Themes applied on a collection of Swedish folk legends. The topic “kyrka – bygga – stannade” (“church – build – stayed”) is displayed with a blue background in the center panel, which indicates that it has been selected by the user. The snippet versions of the texts most closely associated with the selected topic are shown to the right, while associated terms are indicated with a bold border to the left.
The snippet text is produced by only retaining those sentences that contain terms that the topic modelling algorithm has extracted as typical to the topics. A sentence – or a sequence of sentences – that does not contain any of these terms is replaced by ellipsis, as shown in Figure 1. The Topics2Themes configuration file lets the user specify the maximum number of sentences to be included in the snippet text. This has the effect that only the first sentences in the original text that contain a term will be added to the text snippet, until the snippet has reached the maximum length specified. However, it is also made sure that the snippet text includes at least one context sentence for each one of the terms occurring in a text. Therefore, when more terms occur in the text than the maximum snippet length specified, the length of the snippet will instead be determined by the number of unique topic model-extracted terms occurring in the text.
The interface of Topics2Themes shows the snippet version of the text by default in the Texts panel. There are two methods by which the user can make the tool display the full-text version, (i) either by double-clicking on one specific list element to make the tool show the full-text version for this specific text, or (ii) by clicking on the arrow button in the header of the Texts panel to make the tool show the full-text versions for all texts. 4
We have now carried out the technical developments required for adapting the Topics2Themes tool to our collection of folk legends. We have also provided the tool with collection-adapted data in the form of an adapted stop word list and word lists for improving the word2vec-based word clusters. As the next step, we will therefore use Topics2Themes for carrying out a computer-assisted analysis of the folk legend collection. That is, we will use the tool for identifying recurring themes in Swedish folk legends, i.e. themes that might (i) reflect already established classification systems, as well as (ii) provide new dimensions on how legends can be categorised in folkloristics.