Measuring Corporate Culture Using Machine Learning

Introduction

The repository implements the method described in the paper

Kai Li, Feng Mai, Rui Shen, Xinyan Yan, Measuring Corporate Culture Using Machine Learning, The Review of Financial Studies, 2020; DOI:10.1093/rfs/hhaa079 [Available at SSRN]

The code is tested on Ubuntu 18.04 and macOS Catalina, with limited testing on Windows 10.

Requirement

The code requres

os.environ["CORENLP_HOME"] = "/home/user/stanford-corenlp-full-2018-10-05/"

  • If you are using Windows, use "/" instead of "\" to separate directories.

  • Make sure requirements for CoreNLP are met. For example, you need to have Java installed (if you are using Windows, install Windows Offline (64-bit) version). To check if CoreNLP is set up correctly, use command line (terminal) to navigate to the project root folder and run python -m culture.preprocess. You should see parsed outputs from a single sentence printed after a moment:

    (['when[pos:WRB] I[pos:PRP] be[pos:VBD] a[pos:DT]....

Data

We included some example data in the data/input/ folder. The three files are

  • documents.txt: Each line is a document (e.g., each earnings call). Each document needs to have line breaks remvoed. The file has no header row.
  • document_ids.txt: Each line is document ID (e.g., unique identifier for each earnings call). A document ID cannot have _ or whitespaces. The file has no header row.
  • (Optional) id2firms.csv: A csv file with three columns (document_id:str, firm_id:str, time:int). The file has a header row.

Before running the code

You can config global options in the global_options.py. The most important options are perhaps:

  • The RAM allocated for CoreNLP
  • The number of CPU cores for CoreNLP parsing and model training
  • The seed words
  • The max number of words to include in each dimension. Note that after filtering and deduplication (each word can only be loaded under a single dimension), the number of words will be smaller.

Running the code

  1. Use python parse.py to use Stanford CoreNLP to parse the raw documents. This step is relatvely slow so multiple CPU cores is recommended. The parsed files are output in the data/processed/parsed/ folder:

    • documents.txt: Each line is a sentence.
    • document_sent_ids.txt: Each line is a id in the format of docID_sentenceID (e.g. doc0_0, doc0_1, ..., doc1_0, doc1_1, doc1_2, ...). Each line in the file corresponds to documents.txt.

    Note about performance: This step is time-consuming (~10 min for 100 calls). Using python parse_parallel.py can speed up the process considerably (~2 min with 8 cores for 100 calls) but it is not well-tested on all platforms. To not break things, the two implementations are separated.

  2. Use python clean_and_train.py to clean, remove stopwords, and named entities in parsed documents.txt. The program then learns corpus specific phrases using gensim and concatenate them. Finally, the program trains the word2vec model.

    The options can be configured in the global_options.py file. The program outputs the following 3 output files:

    • data/processed/unigram/documents_cleaned.txt: Each line is a sentence. NERs are replaced by tags. Stopwords, 1-letter words, punctuation marks, and pure numeric tokens are removed. MWEs and compound words are concatenated.
    • data/processed/bigram/documents_cleaned.txt: Each line is a sentence. 2-word phrases are concatenated.
    • data/processed/trigram/documents_cleaned.txt: Each line is a sentence. 3-word phrases are concatenated. This is the final corpus for training the word2vec model and scoring.

    The program also saves the following gensim models:

    • models/phrases/bigram.mod: phrase model for 2-word phrases
    • models/phrases/trigram.mod: phrase model for 3-word phrases
    • models/w2v/w2v.mod: word2vec model
  3. Use python create_dict.py to create the expanded dictionary. The program outputs the following files:

    • outputs/dict/expanded_dict.csv: A csv file with the number of columns equal to the number of dimensions in the dictionary (five in the paper). The row headers are the dimension names.

    (Optional): It is possible to manually remove or add items to the expanded_dict.csv before scoring the documents.

  4. Use python score.py to score the documents. Note that the output scores for the documents are not adjusted by the document length. The program outputs three sets of scores:

    • outputs/scores/scores_TF.csv: using raw term counts or term frequency (TF),
    • outputs/scores/scores_TFIDF.csv: using TF-IDF weights,
    • outputs/scores/scores_WFIDF.csv: TF-IDF with Log normalization (WFIDF).

    (Optional): It is possible to use additional weights on the words (see score.score_tf_idf() for detail).

  5. (Optional): Use python aggregate_firms.py to aggregate the scores to the firm-time level. The final scores are adjusted by the document lengths.