Predict numeric variable from a text variable using word embeddings in R-CodePudding

I have a text variable with reviews of movies and another variables with ratings – I want to try to use the text reviews to predict the ratings.

Here are some example data:

movie_reviews <- c("I really loved the movie plot", "This movie really sucked", "I really found this movie thought provoking", "ahh what a boring movie", "A wonderful movie, with a wonderful end", "Great action movie: Very thrilling", "Worst movie ever, it never stopped being cheesy", "Enjoying, feelgood movie for the entire family", "I will definitely watch this movie again")

movie_ratings <- c(8, 2, 6, 3, 9, 8.5, 3.5, 9.5, 7.5)  
  
movie_df <- tibble(movie_reviews, movie_ratings)

Thank you.

CodePudding user response：

For this you can use the text-package

# Create word embedding representations of your text
help(textEmbed)
reviews_embeddings <- textEmbed(movie_df, 
                                model = "bert-base-uncased", # Select model you want from huggingface
                                layers = 11:12) # Select which layers you want to use

# Train the word embeddings to the numeric variable using ridge regression 
reviews_rating_model <- textTrain(reviews_embeddings$movie_reviews, 
                                  movie_df$movie_ratings) 
# See the results
reviews_rating_model

Result

$results

    Pearson's product-moment correlation

data:  predy_y$predictions and predy_y$y
t = 5.621, df = 7, p-value = 0.0003991
alternative hypothesis: true correlation is greater than 0
95 percent confidence interval:
 0.6785761 1.0000000
sample estimates:
      cor 
0.9047823