Topic modeling guide

[dselivanov]

It will be useful to create a comprehensive practical guide for topic modeling. Now we have all components in place:

• POS tags and lemmatization - thanks to udpipe package
• coherence measures - thanks to Manuel work
• fast LDA, thanks to WarpLDA in text2vec
• fast non-negative matrix factorization, thanks to rsparse package
• multi-word phrase extraction - several approaches text2vec::Collocations, udpipe::as_phrasemachine

Steps

• [ ] find interesting non-trivial corpus with large number of documents
• [ ] demonstrate how to create tokenizer which only use particular POS
• [ ] create collocation model on top of that
• [ ] create document-term matrix using tokens with multi-word expression
• [ ] fit several topic models (text2vec::LDA, rsparse::WRMF) with different hyper parameters
• [ ] cross-validate / compare them using different coherence metrics
  • [ ] demonstrate usage of external corpus for tcm calculation
  • [ ] check on how coherence metrics are correlated (is perplexity correlated with them? )

There are already good vignettes in udpipe package topic modeling and phrase extraction. They can be used as inspiration.

@manuelbickel @jwijffels anything we can add to the plan above?

[sjankin]

Re non-trivial corpus with large number of documents, how about UN General Debate corpus? It's publicly available from Harvard Dataverse: "UNGDC 1970-2017.zip". Direct link here. It covers country statements in the UN General Debate (presidents, prime ministers etc), once per year at the opening of each UN session from 1970 to 2017. Total 7,897 speeches.

[manuelbickel]

We might add some aspects regarding downstream analysis (and maybe visualization depending on the target audience or format of publication).

Regarding downstream analysis we might do (feel free to change/adapt/add):

• [ ] use LDAvis for PCA on topics to understand the main differences between the topics in the corpus
• [ ] apply cluster analysis on doc topic matrix (or topic term matrix) to highlight commonalities - either with probabilities or with classified probabilities, in the extreme case 0 and 1
• [ ] apply network analysis on doc topic matrix (or topic term matrix) to identify relations/co-occurrence between documents or topics (setting a threshold seems reasonable, e.g., only use the three strongest topics per document and set others to zero)
• [ ] do trend analysis over time on prevalence of topics (e..g by calculating mean of topic probability per year and using AICC loess smoothing); this procedure is not perfect, since standard LDA neglects the timely order of documents; more sophisticated algorithms like dynamic topic models exist (see, e.g. Blei) that do this, which, however, are more restrictive - not sure what the best practice of trend analysis is at the moment

[jwijffels]

Nice points. I have some time from May 13 onwards to work on this. I would be interested in having a corpus which has the same text in several languages to show that the flow works for all languages with limited manual intervention. In Belgium we have some open data (20000 records if I recall) for all question/answers in parliament for the last years but that is only Dutch and French. It would be nice to have a corpus with also English in it + some more languages (maybe europarl?) FYI. I've also added more docs on multi-word phrase extraction at https://bnosac.github.io/udpipe/docs/doc7.html