Data scrapping

For the purpose of this exercise I’ve accessed Trustpilot reviews of Monoprix. This is what the page looks like:

As you can see, the reviews are spread through several pages, so I write a small script that goes through all fo them, collecting data. To access the elements that are of interest, I look at the souce code of the page first.

For instance, I want to collect the review title, text and the rating. That’s what the code for the rating looks like:

    rank_data <- webpage %>%
                                      html_nodes('.star-rating img') %>%
                                      html_attr("alt") 

As a result we have the following dataset with 234 observations (reviews):

Obviously, the description text is rough, has lots of special characters and needs a through cleaning before sentiment analysis can be run on it.

Text mining and preliminary analysis

I convert the review text into a special object called text corpus and run transformations on it. The basic ones are removing numbers, punctuation, extra white space and special characters like apostrophes and semicolons, followed by transforming the text to lower case.

Next, the stopwords are removed (the words that are common to a given language but don’t give out any sentiment, like, for instance, ‘is’, ‘are’ and ‘the’ in English or ‘de’, ‘la’ and ‘sont’ in French).

This is what the stripped text looks like:

## <<SimpleCorpus>>
                                ## Metadata:  corpus specific: 1, document level (indexed): 0
                                ## Content:  documents: 4
                                ##
                                ## [1]  moitié produits commandés ligne disponibles donc livrés livraison a lieux minutes plus tard heure indiqué fruit légumes sûrement pires stock plus tout produits date expiration lendemain jours suivants!plus jamais
                                ## [2]  pourquoi monoprix embauche seniors cv jamais entretien
                                ## [3]  bonjour souhaite souligner efficacité service client ainsi gentillesse disponibilité personnel aime acheter articles mode ainsi lingerie excellent rapport qualité prix côté très branché plus période soldes beaucoup affaires disponibles compris site bravo
                                ## [4]  monoprix narbonne a tres evolué espaces fonctionnels tres organisés tres jolies choses tres pratiquepar contre responsable tres désagréable repond sèchement moque clients certaines salariées aucune discretion dommage peut etre direction pourrait envoyer formation puisse acquérir certaines bases accueil respect clientcela plus ans viens magasin bien plus venais mere ponpon

Finally, the text corpus is sometimes stemmed. Stemming is removing the words endings so that the root only counts for sentiment analysis. In the stemmed text, the words ‘run’, ‘runs’ and ‘running’ will count as the same word three times. It makes sense for machine analysis, but doesn’t please to a human reader. Thus, I skip stemming and move on to building a term document matrix, which is basically a matrix containing the frequency of the terms we encounter in the analyzed text.

Very commonly, it is a sparse matrix, meaning that the object is large, but most of the words are only there once or twice. For the ease of manipulating the data and also because these rare words are not of any use, the sparse terms are removed from the matrix.

So, below are the Top 10 common words that are seen in Monoprix reviews on Trustpilot:

Visually it could be shown in the histogram form:

However, the most common representation is the wordcloud, which is basically just an infographic where the words’ size correspond to their frequency in the text.

While this is a nice representation, undobtedly, it is not very useful on its own. However, it can give the ideas where to explore further. For example, we can see the associated terms with each of the commonly shown terms (the number next to each term is the correlation level).

	service
client	0.55
réponse	0.48
toujours	0.42
nouvelle	0.38
nouveau	0.38
personne	0.37
aucune	0.34
part	0.34
demande	0.32

	jamais
plus	0.3
juste	0.29
manque	0.27
internet	0.25
horaire	0.24
livraison	0.21
bon	0.21
livraisons	0.21

	aucun
nouveau	0.37
respect	0.34

	cher
livraisons	0.33
plus	0.26

The applications of term associations are more interesting. For example, as a consultant hired to improve marketing management, and without any prior knowledge about Monoprix and its problems I can quickly come up with the conclusion that there are underlying problems with missing articles when groceries are delivered, as well as some problems with on-time delivery. It seems that the reviewing customers no longer want to use internet to order groceries. It might also be that the customers perceive the groceries ordered online as more expensive.

Of course, all these findings are preliminary and need somebody looking forward into them. They should not be presented to the board of directors as the cause of the customer dissatisfaction. However, it is a useful basis for data collection and investigation.

Sentiment Analysis

The most ancient and robust technique of sentiment analysis is based on taking the so-called dictionaries (the list of positive and negative words for the given language) and counting the number of instances, or rather, the percentage, for each of the words. For instance, if the first review has 30% negative words and 10% of positive words (because some words are, obviously, neutral), then we might consider that the review will most likely have a negative ranking (one or two stars).

This approach has a number of flaws. One of them is the treatment of negation and another is the question: what do you consider positive and negative? Is ‘milk’ considered positive? It might in some context (ask your toddler!), but if we’re reading supermarket reviews about groceries, it is most likely neutral.

What about the figures of speach? What about building the dictionaries for every language?

Despite these inconveniences, the method is still highly effective, as I’m going to demonstrate. To do so, I will use the FEEL dictionary (FEEL: French Expanded Emotion Lexicon. Special thanks to Amine Abdaoui, Jérôme Azé, Sandra Bringay et Pascal Poncelet).

Below is the list of the top review words that were considered positive, along with the supposedly negative words.

Positive: heure, lendemain, ligne, livraison, client, personnel, qualité, service, site, accueil, bien, magasin, commande, temps, enseigne, bon, message, article, réponse, jour, achat, bref, carte, compte, domicile, expérience, montant, caisse, personne, rayon, aller, mettre, panier, surtout, créneau, horaire, produit, part and suite

Negative: dommage, moins, mal, problème, cher, appel, retard and déplorable

We will see further down that even if some of the positive words don’t have real positive connotation, it doesn’t have much impact on the accuracy of the analysis.

We will combine the words and their (relative, in %) frequencies in the following dataframe, where each row correspond to the review, so there’s 234 rows in total:

Note that the last column is the actual ranking coming from the rating in stars, that was converted for simplicity into zero (negative review) or one (positive review). The reason for this binary conversion is that we have a small amount of data (234 reviews), and the majority of them are very negative. It will be difficult or nearly impossible to teach the algorithm to distinguish between 4- and 5-stars reviews if there aren’t any or if there’s very little of those.

Prediting a positive or negative review

We’re going to split 234 reviews into a training and test set, based on 70%/30% rule. We’ll train the algorithm on training set and evaluate its prediction on the test set, the data the algorithm hasn’t yet seen.