Replication code for "Robust Machine Learning Algorithms for Text Analysis". If you use this code, please cite Ke, Montiel Olea, Nesbit. Quantitiative Economics Forthcoming.
The paper has two modules of programming exercises. One is a simulation exercise that illustrates the main results with stylized examples with V = K = D = 2. The other is an empirical exercise that runs our algorithm on FOMC transcripts.
Compile language: Matlab, Python
Python dependencies: tika, topicmodels, nltk, gensim, scipy
The replication codes for simulation (Section 6.1 of main text and Section E of online appendix) are in \simulation folder. They are self-contained scripts that could be run individually. Below describes each folder and which plot the script replicates:
-
1Sensitivity: Run Sensitivity.m for Figure 3 in main text
-
2Range: Run Range_main.m for Figure 4, 5 and 6 in main text
-
3MonteCarlo: Run Main_MonteCarlo.m for Figure 7 in main text and Figure 4 in online appendix
-
4Approximation: Run Main_approx.m for Figure 3 in online appendix
-
5AnchorWord: Run Main_AnchorWord.m for Figure 5 and 6 in online appendix
The rest of the codebase is for replicating the empirical exercise on FOMC transcript (Section 6.2 of main and Section F of online appendix).
The raw data used are in folder
- FOMC_pdf: raw PDF files for FOMC transcripts from Aug-1987 to Jan-2006
- util_files: covariates and separation rules for FOMC 1 and 2
The replication consists of following steps:
-
run preprocessing.MAIN_preprocess.py, which parses raw pdf files, clean up corpus, and generate term-document frequency matrices. The final term-document frequency matrices are stored in \term-document matrix folder
-
run preprocessing.MAIN_generate_NMF_draws.py, which generates NMF draws of prior and posterior estimates of P_hat with different prior values used in the paper. The final outputs are .mat files stored in "NMF_draws_folder" defined in preprocessing.Constant.py
Note: for full replication of all graphs, this step will probably take a long time (6-8 hours) and the output data takes significant storage space. It is highly recommended to change the NMF_draws_folder to point to a directory with >150G spare space. Alternatively, user can run single block in Main_generate_NMF_draws.py to replicate individual plot
-
In empirical_analysis, run Main_empirical.m to generate all graphs in the empirical section. Change NMF_draw_parent_folder in Main_empirical.m to point to NMF_draws_folder as specified in preprocessing.Constant.py.
Running the full script will also take some time. User can also run single block to generate individual graph. The plots are stored in empirical_analysis/Figures. The file names for figures in the paper can be found in figure_names.md.