This is code that was used in our recent work on clustering of diachronic word embeddings:
Marjanen, J., Kurunmäki, J. A., Pivovarova, L., & Zosa, E. (2020). The expansion of isms, 1820-1917: Data-driven analysis of political language in digitized newspaper collections. Journal of Data Mining and Digital Humanities. https://doi.org/10.46298/jdmdh.6159
Pivovarova, L., Marjanen, J., & Zosa, E. (2019). Word Clustering for Historical Newspapers Analysis. In Ranlp Workshop on Language technology for Digital Humanities.
Marjanen, J., Pivovarova, L., Zosa, E., & Kurunmäki, J. (2019). Clustering ideological terms in historical newspaper data with diachronic word embeddings. In 5th International Workshop on Computational History, HistoInformatics 2019. CEUR-WS.
Diachronic embeddings built on the National Library of Finland newspaper collection could be downloaded from here.
We used an incremental training method, closely following (Kim et al., 2014) and previously applied by (Hengchen et. al, 2019). More explanations, code and several embeddings model check could be found here.
Once you obtained embeddings you can apply clustering using clustering.py (clustering of selected words) or cluster_all.py (enriched clustering).
Currently the code uses hard-coded links to models and hardcoded list of words.
The clustering outputs json files that can be used to make Sankey chart using diachronic_shift_sankey.py.
Selected embeddings could be also vizualized using embeddings_drift_tsne.py.
The code currently uses hard-coded paths.