Diego García

Below is the data we shared in the version from 20211201. The methodology/output is very similar, but we use a more stringent criteria for inclusion in the 2022 version (smaller dictionaries, avoiding overfitting as much as possible). This is a link to the 2021 paper version, which has the robust MNIR discussion in Section 6, and it has a simpler implementation of the MNIR model in the body of the draft (slightly easier to replicate). README file describing the contents of the shared data. The file dictionaries data (RData file) contains the dictionaries, with associated code (part 1), which among other things, includes a disambiguation function (as in Table 6 of the paper). The metadata and document-term-matrices (RData file) contain public returns (from Kaggle) that allow for a partial replication of our main results. The associated code (part 2) presents some simple commands to both manipulate dtms and create sentiment scores, as well as showing how we fit the MNIR model. If you are only interested in the lists of words, here are the ML+LM dictionaries from the paper (zip file with txt files).