Natural Language Processing (NLP) has been one of the fastest-growing research fields in the last decade. Innovations such as pre-trained large language models based on transformer neural networks have not only led to the popularization of AI and NLP in the general public, but also to interdisciplinary research projects in the humanities and social sciences facilitated by the scalability of these methods. In this post, we present two tools that aim to facilitate said interdisciplinary research: Styloscope and Toposcope. The tools were developed in Python and can be used from the command line or from a user interface. The code, detailed installation instructions, and user guidelines can be found on GitHub:
Styloscope
Styloscope is a tool for automatic writing style analysis. It can be used to test hypotheses about large-scale corpora, parse documents, or detect outliers. Users can provide data by either uploading a local file or by using a publicly available Huggingface dataset. When uploading a corpus, the tool accepts CSV files with one document per row, and ZIP folders in which documents are stored in individual text files. The output contains the parsed documents, raw statistics on various writing style features such as syntactic dependencies, lexical richness, readability, etc., and visualizations of aggregated results. An example for syntactic dependencies is provided below:
Toposcope
Toposcope can be used to detect topics in unstructured text data. It provides annotations and visualizations of the detected topics, including (changes in) topic frequency over time. The tool features four algorithms: BERTopic (Grootendorst, 2022), Top2Vec (Angelov, 2020), Non-negative Matrix Factorization (Choo et al., 2013), and Latent Dirichlet Allocation (Blei et al., 2003). Users can modify a selection of topic model parameters, and apply a number of built-in preprocessing steps, such as lemmatization and stopword removal. The input format is identical to the Styloscope format: users can upload a local corpus (CSV/ZIP), or use a Huggingface dataset. The output includes visualizations of the topic-document clusters (as shown below) and the most important keywords per topic. The raw results, among other things, consist of annotations, a topic-document matrix, and a topic-term matrix. Topic diversity and topic coherence are also computed in order to support the user during the evaluation of the tool.
How to cite
Jens Lemmens and Walter Daelemans. 2024. Styloscope and Toposcope: Towards user-friendly digital text analysis. CLiPS Technical Report Series (CTRS): 10. https://www.uantwerpen.be/en/research-groups/clips/research/computational-linguistics/compling-resources/clips-technical-repo/
References
- Dimo Angelov. 2020. Top2Vec: Distributed representation of topics. arXiv:2008.09470.
- David M. Blei, Andrew Y. Ng, and Michael I. Jordan. 2003. Latent dirichlet allocation. Journal of Machine Learning Research, vol. 3, pp. 993—1022.
- Jaegul Choo, Changhyun Lee, Chandan K. Reddy and Haesun Park. 2013. Utopian: User-driven topic modeling based on interactive nonnegative matrix factorization. IEEE Transactions on Visualization and Computer Graphics, vol. 19 (12), pp. 1992—2001. Institute of Electrical and Electronics Engineers (IEEE).
- Maarten Grootendorst. 2022. BERTopic: Neural topic modeling with a class-based TF-IDF procedure. arXiv:2203.05794.
CLARIAH-VL SIC 5 tool descriptions
- Styloscope: https://clariahvl.hypotheses.org/styloscope
- Toposcope: https://clariahvl.hypotheses.org/toposcope
