Awesome

3x8emotions

Tobias Widmann & Maximilian Wich June 2022

Repo containing code and models for 3 different tools to measure appeals to 8 discrete emotions in German political text, as described and validated in the following article:

Widmann, Tobias, and Maximilian Wich. "Creating and Comparing Dictionary, Word Embedding, and Transformer-Based Models to Measure Discrete Emotions in German Political Text." Political Analysis, June 29, 2022, 1--16. https://doi.org/10.1017/pan.2022.15

Please start by reading this article which contains information about the creation and performance of the different tools. These tools are free to use for academic research. In case you use one or multiple of these, please always cite the article above.

Important: Download the files by clicking on the latest release on the right side. Two files need to be downloaded additionally due to size limitations: the ELECTRA model (pytorch_model.bin) and the locally trained word embeddings (vec_ed_preprocessed.txt). The folder contains all scripts to apply the (1) ed8 dictionary, (2) the neural network models based on locally trained word embeddings and (3) the ELECTRA model.

(1) ed8

The ed8 dictionary is provided in YAML format and can be applied via the quanteda package. The dictionary and the R script apply_ed8.R to apply the dictionary to a data frame with a ‘text’ column can be found in the folder “./01_ed8”.

# First, load quanteda package
library(quanteda)

#  Load in dictionary
ed8 <- dictionary(file = "./ed8.yml",
                  format = "YAML")

# Create the function
get_ed8_emotions <- function(data){
  #Create a corpus from your data frame
  corp <- corpus(data)
  
  #Tokenize corpus and pre-process (remove punctuations, numbers, and urls)
  toks <- tokens(corp, remove_punct = TRUE, remove_numbers = TRUE, remove_url = TRUE)
  
  #Create DFM just to measure number of terms before removing stopwords
  terms_dfm <- dfm(toks)
  
  #Create bigram-compounds to include negation control
  toks_neg_bigram <- tokens_compound(toks, pattern = phrase("nicht *"))
  toks_neg_bigram <- tokens_compound(toks_neg_bigram, pattern = phrase("nichts *"))
  toks_neg_bigram <- tokens_compound(toks_neg_bigram, pattern = phrase("kein *"))
  toks_neg_bigram <- tokens_compound(toks_neg_bigram, pattern = phrase("keine *"))
  toks_neg_bigram <- tokens_compound(toks_neg_bigram, pattern = phrase("keinen *"))
  
  #Turn tokens into DFM, remove stopwords
  emo_dfm <- dfm(toks_neg_bigram, remove = stopwords("de"))
  
  #Apply dictionary
  dict_dfm_results <- dfm_lookup(emo_dfm,ed8)
  
  #Convert results back to data frame
  results_df <- cbind(data, convert(dict_dfm_results, to = 'data.frame'))
  
  #Assign length to each documents
  results_df$terms_raw <- ntoken(terms_dfm)
  results_df$terms <- ntoken(emo_dfm)
  
  return(results_df)
}

# Now you can use the function on your data; simply enter a data frame with a column called "text" including the text data
results <- get_ed8_emotions(data)

# Finally, you can create normalized emotional scores by dividing the ed8-scores by document length
results$anger.norm <- results$ed8.ANGER / results$terms
results$fear.norm <- results$ed8.FEAR / results$terms
results$disgust.norm <- results$ed8.DISGUST / results$terms
results$sadness.norm <- results$ed8.SADNESS / results$terms
results$joy.norm <- results$ed8.JOY / results$terms
results$enthusiasm.norm <- results$ed8.ENTHUSIASM / results$terms
results$pride.norm <- results$ed8.PRIDE / results$terms
results$hope.norm <- results$ed8.HOPE / results$terms

(2) Neural Network Classifiers

The neural network classifiers and locally trained word embedding model are provided in the folder “./02_neuralnet”. The code for turning text into numerical vectors and subsequently applying the neural network classifiers can be found in the R script apply_neuralnet.R. Remember, the machine learning models were trained on sentences, so you need to bring your text data on sentence level first.

# Load necessary packages
library(quanteda)
library(corpus)
library(keras)
library(tidytext)

# Set working directory
setwd("./neuralnet")


# First, you need to turn your text into sentences
data <- data %>% 
  unnest_tokens(sentences, text, "sentences")

# Now, you can turn your text documents in a corpus
corp <- corpus(data$sentences)

# Create a document feature matrix and conduct pre-processing
text_dfm <- dfm(corp, remove=stopwords("german"), verbose=TRUE, tolower = TRUE)

# Stemming
text_dfm <- dfm_wordstem(text_dfm, language = "german")

# Now, we will convert the word embeddings into a data frame 
# and match the features from each document with their corresponding embeddings

#F irst, we load the locally trained word embeddings into R
w2v <- readr::read_delim("./vec_ed_preprocessed.txt", 
                         skip=1, delim=" ", quote="",
                         col_names=c("word", paste0("V", 1:100)))

# Stem the terms included in the embeddings to increase matches
w2v$word <- text_tokens(w2v$word, stemmer = "de")

# creating new feature matrix for embeddings
embed <- matrix(NA, nrow=ndoc(text_dfm), ncol=100)
for (i in 1:ndoc(cgdfm)){
  if (i %% 100 == 0) message(i, '/', ndoc(text_dfm))
  # extract word counts
  vec <- as.numeric(text_dfm[i,])
  # keep words with counts of 1 or more
  doc_words <- featnames(text_dfm)[vec>0]
  # extract embeddings for those words
  embed_vec <- w2v[w2v$word %in% doc_words, 2:101]
  # aggregate from word- to document-level embeddings by taking AVG
  embed[i,] <- colMeans(embed_vec, na.rm=TRUE)
  # if no words in embeddings, simply set to 0
  if (nrow(embed_vec)==0) embed[i,] <- 0
}

# After you created the sentence embeddings, you can apply the trained machine learning models for each emotion
# The machine learning models are provided in the folder "./neuralnet/models"
# for example, anger:

model <- load_model_hdf5("./models/keras_anger", custom_objects = NULL, compile = TRUE)
wb.anger <- model %>% predict_classes(embed)
data <- cbind(data, wb.anger)

(3) ELECTRA Model

The ELECTRA files are provided in the folder ./03_electra. The model can be applied to text data using the Python code as shown in the Python notebook apply_electra.ipynb. The ELECTRA model was trained on sentences, so you need to bring your text data on sentence level first.

# Set working directory
%cd /03_electra/

# load necessary modules
import transformers
import pandas as pd

# text documents the model will be applied to
df = pd.read_csv('./data.csv')
documents = data.text

# load inferencer
from helper.inferencing import Inferencer

# predicting
predictor = Inferencer()
df_results = predictor.predict_dataframe(documents)

# show results
df_results