This homework assignment is intended as an opportunity to practice working with converting between tidy and non-tidy formats for text data and also for using the non-tidy formats as input for modeling functionality. As a reminder, the tidy format is very useful for EDA and data visualization, but it not the shape expected for machine learning algorithms. It is often the case where we want to predict genre, authorship, recommend titles, etc. – it these cases, the document-term matrix can be a more appropriate representation of our text data. This is because, in all of these scenarios, the observation is not the particular token, it is the full-text which we wish to classify or recommend.
This week we will work with movie descriptions, each labeled by their primary genres. We’ll start with an exploratory analysis and do some comparison of tidy and untidy organizations of our data. Next we’ll use what was covered in Chapter 6 to create a topic model based on the movie descriptions. We will see how well these models correspond to the primary genres for our movies.
Open up a new R Script or RMarkdown file and load the tidyverse, tidytext, tm, and topicmodels libraries. Remember to install any packages you’ve never used before by running install.packages("PACKAGE_NAME").
You’ll read in the labeled data from our GitHub Repo using the following commands. Note that mutate() is adding a new column called movie_id and select() is just reordering the columns.
movies <- read_csv("https://raw.githubusercontent.com/agmath/FacultyUpskilling/main/2021_NLP/data/train.csv")
movies <- movies %>%
mutate(movie_id = row_number()) %>%
select(movie_id, genre, overview)Now that we have our data, let’s use our familiar workflow of unnesting tokens to create a tidy version of our data frame. Use unnest_tokens() to unnest the overview column.
You should see that the resulting data frame looks quite familiar. We have columns tracking the genre and movie_id and built a column containing each individual word from each review. This is our familiar tidy text data frame. Let’s do something we’ve done before and remove stop_words by using an anti_join(). You can overwrite your movies_tidy data frame here.
That step cut over 120,000 rows of very common words from our data frame. This can be an important step because it will reduce the number of columns present in a document-term-matrix when we construct one. Now let’s build a new data frame that includes word counts for each word within each movie description. Call your new object movies_word_count, start with movies_tidy, group by movie_id, count the word occurrences, and then ungroup().
Now let’s cast our tidy data frame as a Document Term Matrix, which is a sparse matrix (meaning that it contains mostly 0’s) in which each row corresponds to a movie description and each column corresponds to a word present in our corpus of reviews. The entries of this matrix will denote the number of times the word corresponding to the column appears in the description corresponding to the row. Call your new object movies_dtm and create it using the cast_dtm() function. As a reminder, we can pipe a data frame containing our word counts into the cast_dtm() function, and pass three additional arguments which indicate the document id, the column of words/tokens, and the column containing the counts.
Notice that R won’t print out the sparse matrix for us even if we ask for it by name. Instead, we are provided some information about the object. In particular, we can see that our matrix is almost 100% sparse, there are only 100,721 non-zero entries among the over 87 Million entries in the matrix. If you would like to see the terms represented by the first ten columns in our document-term matrix, you can do so by running Terms(movies_dtm)[1:10]. Try it. Change the indices to see terms represented by other columns as well.
Okay, now that we’ve got our data into a format where every row represents a movie description and every column represents a feature of that movie description (ie. a word it may contain), we are in a scenario where we can try building a topic model using the LDA() function from R’s topicmodels package. As a reminder, when we began, we knew the primary genre of each movie in this dataset of descriptions. There are 20 primary genres present in our movie descriptions dataset, which you can verify by running length(unique(movies$genre)). Use LDA() on our document-term-matrix representing the movie descriptions and use 20 topics – it may take about five minutes or so to run. Be sure to store the result in movies_lda.
Create an object called movie_genre_est by using the tidy() function with the arguments movies_lda and matrix = "beta". View the result – what do the columns in the output represent?
Similar to what is done in the text, let’s use slice_max() to find the 10 terms from each topic which are most surely associated with that topic. Adapt the code from the textbook to produce a faceted set of plots showing your results. If you are working in R Markdown, you may want to set the fig.height chunk parameter to something like 20 – you set your chunk arguments within the curly braces at the beginning of an R chunk.
After seeing your plot panel, do you think LDA learned topics based on the film genres? Let’s see which topic each film was assigned to using tidy() again, but setting the matrix argument to "gamma". Store your result in an object called movie_assignments. View the result.
Let’s try to extract the topic each movie has been assigned to by adapting our code from 10. We still want to use slice_max() here, but let’s group by document and then take only the row corresponding to the top gamma value. After slicing, don’t forget to ungroup() the data frame. The document column will be converted to a string by default, so you’ll want to use mutate() and as.numeric() to convert this column back to a numeric column. Store your result in movie_assignment.
Did our Latent Dirichlet model create topics corresponding to the primary genres? Let’s see. We’ll start by joining the known primary genre of each movie onto our movie_assignment data frame. Then we will build and plot a confusion matrix to see our results!
movies and movie_assignment data frames have columns representing the movie id. In the original movies data frame, that column is called movie_id while it is called document in the movie_assignment data frame. Start with the movie_assignment data frame and left_join() the information from movies onto it. You’ll need to supply the argument by = c("document" = "movie_id") to left_join() because the key columns don’t share the same name.genre and the assigned topic contained in the movie_assignment data frame, follow and adapt the code from the book to create the tile plot representing the confusion matrix.Final Thoughts: It looks like our model didn’t choose its topics according to the primary genre. The foreign films and made for TV movies were each split across very few topics, but films from the other genres were shared widely across the 20 topics learned by our LDA model. There are a few reasons for this. First, as we saw in the barplots for each topic, the model learned storylines and settings rather than genres. Secondly, in the original dataset, most films were assigned several genre tags, so genres can overlap quite a bit and my choice to assume that the first genre listed was the primary genre may have been a poor one. If you want to dig a bit deeper into how our topic model grouped movies, you can run code blocks similar to the one appearing below:
movie_assignment %>%
filter(topic == 1) %>%
select(genre, overview)In any case, I hope you found this assignment fun and that it showed you a bit about why untidy formats are sometimes preferred over tidy data. Additionally, I hope you’ll consider applying topic models to different contexts!