README.md

In taylorgrant/twAccount: Twitter Timelines, Images, and Topic Models

twAccount

This package includes a series of functions that allow the user to extract a brand’s Twitter timeline, visualize the brand identity through the imagery posted and estimate topic models for both the timeline as well as the mentions of that brand.

Twitter extraction capabilities are courtesy of rtweet. Image processing and conversion into collages is thanks to the magick package. The topic modeling process relies on the BTM package, which is designed for short texts as opposed to standard LDA, which relies on a larger corpus. The BTM output is then processed using several functions borrowed from the LDAvis package.

Installation

You can install the development version of twAccount from GitHub with:

# install.packages("devtools")
devtools::install_github("taylorgrant/twAccount")
remotes::install_github("taylorgrant/twAccount")

Usage

# Full Usage 
twitter_tm("BMWUSA", search_query = "bmw", topic_ct = 40, term_ct = 30)

# If you have a timeline from `rtweet` with/without mentions, and just want topic models 
run_btm("BMWUSA", data_source = 'timeline', n_topics = 40, n_terms = 30)

# If you have a timeline and want images 
get_pics(dir_location, timeline)
# If you want to make final collages
unique(lubridate::year(timeline$created_at)) %>% 
  purrr::walk(make_collage)

The package provides an overall wrapper function twitter_tm() that will extract (1) a brand’s Twitter timeline, (2) recent, public tweets matching a boolean search, and (3) all images that a brand has included in its publicly available tweets.

Once the data is extracted, a model is run to estimate topic models based upon the Biterm Topic Models algorithm in the BTM package. Currently, the model is specified to run based on the number of topics selected by the user, rather than running through a series of topic sizes and maximizing the log-likelihood (maybe later). The topic modeling process yields estimates of where the topics lie in 2d space, as well as the top N most salient terms for each topic (N terms specified by the user).

Estimation of the 2d space relies upon the Jensen-Shannon Divergence, the estimates of which are then scaled via PCA. This allows for easy visualization of the topics on an X,Y axis. For example, if we took the last 3,200 tweets from the BMWUSA account, a version of the 2d topic model representation would look like the below. This can be presented in a static graph, or in an html rmarkdown file using the crosstalk package, or you can simply use the d3 visualization from LDAvis.

Using the package

Ideally, the user would create a new project within RStudio, but this package will build out a new directory in the current working directory of the user.

After running the twitter_tm() function, the new directory will follow this structure:

├── [HANDLE]
│   └── [RUN DATE]
│       ├── [FIRST YEAR OF TWEETS]
│       ├── [NEXT YEAR OF TWEETS]
│       ├── [ANOTHER YEAR OF TWEETS]
│       ├── {handle}_tm_timeline.rds
│       ├── {handle}_tm_mentions.rds
│       ├── {handle}_twitter_info.rds
│       └── local_cols
└── [udpipe MODEL]

taylorgrant/twAccount documentation built on March 20, 2022, 7:33 p.m.