tmtoolkit: Text mining and topic modeling toolkit

tmtoolkit is a set of tools for text mining and topic modeling with Python developed especially for the use in the social sciences, linguistics, journalism or related disciplines. It aims for easy installation, extensive documentation and a clear programming interface while offering good performance on large datasets by the means of vectorized operations (via NumPy) and parallel computation (using Python’s multiprocessing module and the loky package). The basis of tmtoolkit’s text mining capabilities are built around SpaCy, which offers many language models. Currently, the following languages are supported for text mining:

Catalan
Chinese
Croatian
Danish
Dutch
English
Finnish
French
German
Greek
Italian
Japanese
Korean
Lithuanian
Macedonian
Norwegian Bokmål
Polish
Portuguese
Romanian
Russian
Spanish
Swedish
Ukrainian

The documentation for tmtoolkit is available on tmtoolkit.readthedocs.org and the GitHub code repository is on github.com/internaut/tmtoolkit.

Features

Text preprocessing and text mining

The tmtoolkit package offers several text preprocessing and text mining methods, including:

tokenization, sentence segmentation, part-of-speech (POS) tagging, named-entity recognition (NER) (via SpaCy)
lemmatization and token normalization
extensive pattern matching capabilities (exact matching, regular expressions or “glob” patterns) to be used in many methods of the package, e.g. for filtering on token or document level, or for keywords-in-context (KWIC)
adding and managing custom document and token attributes
accessing text corpora along with their document and token attributes as dataframes
calculating and visualizing corpus summary statistics
finding out and joining collocations
calculating token cooccurrences
splitting and sampling corpora
generating n-grams and using N-gram models
generating sparse document-term matrices

Wherever possible and useful, these methods can operate in parallel to speed up computations with large datasets.

Topic modeling

model computation in parallel for different copora and/or parameter sets
support for lda, scikit-learn and gensim topic modeling backends
evaluation of topic models (e.g. in order to an optimal number of topics for a given dataset) using several implemented metrics:
- model coherence (Mimno et al. 2011) or with metrics implemented in Gensim)
- KL divergence method (Arun et al. 2010)
- probability of held-out documents (Wallach et al. 2009)
- pair-wise cosine distance method (Cao Juan et al. 2009)
- harmonic mean method (Griffiths, Steyvers 2004)
- the loglikelihood or perplexity methods natively implemented in lda, sklearn or gensim
plotting of evaluation results
common statistics for topic models such as word saliency and distinctiveness (Chuang et al. 2012), topic-word relevance (Sievert and Shirley 2014)
finding / filtering topics with pattern matching
export estimated document-topic and topic-word distributions to Excel
visualize topic-word distributions and document-topic distributions as word clouds or heatmaps
model coherence (Mimno et al. 2011) for individual topics
integrate PyLDAVis to visualize results

Other features

loading and cleaning of raw text from text files, tabular files (CSV or Excel), ZIP files or folders
splitting and joining documents
common statistics and transformations for document-term matrices like word cooccurrence and tf-idf
interoperability with R

Limits

only languages are supported, for which SpaCy language models are available
all data must reside in memory, i.e. no streaming of large data from the hard disk (which for example Gensim supports)

Built-in datasets

Currently tmtoolkit comes with the following built-in datasets which can be loaded via from_builtin_corpus:

“en-NewsArticles”: News Articles (Dai, Tianru, 2017, “News Articles”, https://doi.org/10.7910/DVN/GMFCTR, Harvard Dataverse, V1)
random samples from ParlSpeech V2 (Rauh, Christian; Schwalbach, Jan, 2020, “The ParlSpeech V2 data set: Full-text corpora of 6.3 million parliamentary speeches in the key legislative chambers of nine representative democracies”, https://doi.org/10.7910/DVN/L4OAKN, Harvard Dataverse) for different languages:
- “de-parlspeech-v2-sample-bundestag”
- “en-parlspeech-v2-sample-houseofcommons”
- “es-parlspeech-v2-sample-congreso”
- “nl-parlspeech-v2-sample-tweedekamer”
“en-healthtweets”: Health News in Twitter Data Set

About this documentation

This documentation guides you in several chapters from installing tmtoolkit to its specific use cases and shows some examples with built-in corpora and other datasets. All “hands on” chapters from Getting started to Topic modeling are generated from Jupyter Notebooks. If you want to follow along using these notebooks, you can download them from the GitHub repository.

There are also a few other examples as plain Python scripts available in the examples folder of the GitHub repository.

License

Code licensed under Apache License 2.0. See LICENSE file.

Contents: