3.1 Data Acquisition

Our process of data acquisition was composed of the following steps:

1. First, we identified the html tags required with the information suitable for the purpose of this analysis and prepared a script (scripts/glassdoor-html) in charge of identifying the desired tags. After that, we combined all of the reviews for a company into a single tibble. We also tried to account for any parsing errors that may happen as the html tags of the website were consistently updated.

2. Next, we used another script (scripts\web-scraper) and created a vector with the URL of the five desired companies. Once defining the minimum number of required reviews, which was 5,010 per company, we looped over each page of the review and combined all the results for the five banks into a single tibble. In order to not have issues while accessing the website a several times (i.e. being blocked by the company), we have defined a rnorm_sleep_generator which introduces random pauses between each visit replicating a human’s browsing behaviors.

3. Finally, we remove any duplicates, parse the review time and finished by writing our CSV file with bank’s processed reviews. A sample of the website and the tags during the scrapping process can be found in data\glass-door-sample.html.

3.2 Preprocessing

4.1 Tokenization

Text Mining results are largely impacted by the tokenization process. After testing different options, we decide to use the lemmatization as it is the most promising option.

We create the function EDA_handler, that deals with every aspect of the tokenization. The tokens are split by word and we also remove the stopwords. We also do not want to take into consideration the case and lowercase of the tokens.

4.3 Semantic Analysis

First we will do a sentiment analysis with the two dictionaries of nrc and afinn. Then, we will introduce valence shifters in the subsequent part.

#> # A tibble: 13,901 x 2
#>    word        sentiment
#>    <chr>       <chr>    
#>  1 abacus      trust    
#>  2 abandon     fear     
#>  3 abandon     negative 
#>  4 abandon     sadness  
#>  5 abandoned   anger    
#>  6 abandoned   fear     
#>  7 abandoned   negative 
#>  8 abandoned   sadness  
#>  9 abandonment anger    
#> 10 abandonment fear     
#> # ... with 13,891 more rows
#> # A tibble: 2,477 x 2
#>    word       value
#>    <chr>      <dbl>
#>  1 abandon       -2
#>  2 abandoned     -2
#>  3 abandons      -2
#>  4 abducted      -2
#>  5 abduction     -2
#>  6 abductions    -2
#>  7 abhor         -3
#>  8 abhorred      -3
#>  9 abhorrent     -3
#> 10 abhors        -3
#> # ... with 2,467 more rows

4.4 Job positions analysis

In order to have the most accurate classifier, we need to have a look at every variable available. During the scraping process, we have also managed to extract the job position. It makes sense to have a look at it to see if the employee role has an impact on the company review score.

We have also explored the job positions to potentially identify patterns. After removing inaccurate job-title, we observe on the top 5 most frequent position per bank that some position seems more specific to certain banks. For instance, UBS seems to hire a larger number of interns for its operations.

To have a more interpretable insight and using LSA techniques, we manage to create a biplot in 2 dimensions with the 50 top job positions. Interestingly, JP Morgan, HSBC and UBS seem to have a similar job structure. On the other, TD offers more position related to representants (customer oriented), whereas DB hire more analysts.

4.5 Compare the reviews in terms of lexical diversity

Lexical diversity can also be an interesting tool for classification. However, it seems unlikely to have a usable result here. The only explanation would be that maybe some specific banks hire more non-native English speakers, whose linguistic abilities in exercising this langauge may be lower.

We use both Yule’s index and TTR’s index to demonstrate that they are not candidate for classification. Both graphs tell very different stories, and we cannot correlate these results with the employee_review.

5.1 Preprocessing

To start our analysis, we tokenized our data and removed the stopwords. To make a relevant analysis, we manually removed some words that, from our point of view, would create a bias in our analysis. Stopwords mainly include uninformative words such as quantifiers or company names which appear in every topic. The stopwords definition was an evolutive process. When we performed the unsupervised and superviseed analysis we changed our stopwords gradually.

We selected the pros and cons reviews with at least five tokens because Glassdoor’s users can not write a review with less than five words.

5.2 Latent dirichlet allocation Analysis

The LDA method displays each topic as a combination of words, and each document as a combination of topics. We created two functions, fors pros and cons, that displays our DFM in an LDA plot.

6.1 Document Term Matrix

The following table shows us that the lowest RMSE is obtained by using 50 topics for the pros and only 5 for the cons. The best result (RMSE= 0.913) is given by a linear model using DTM+LSA with 50 pros and 5 cons reviews.

6.2 TF-IDF

The following table shows us that the lowest RMSE is obtained by using 20 topics for the pros and only 5 for the cons. The best result (RMSE= 0.901) is given by a random forest using TF-IDF+LSA with 20 pros and 5 cons reviews.

6.3 Combining the results

6.4 Best model (TF-IDF + LSA)

Here we added to our best model valence shifter (sentiment function mentioned above) and the length of the reviews. We can see that despite the addition of the new element, it does not improve our RMSE (0.900) and that the results obtained previously are all better. However, we can see that our model does quit well when it has to predict twos and threes, which could be explain by the fact that there a more reviews which have been rated with a two or a three.

6.5 Word Embedding & Glove

To complete our supervised learning, we used the Word Embedding and GloVe embedding to model in a different way. First, we computed the RMSE for each combination of number of topic for pros and cons and then, while creating our data frame, we took the mean words instead of the sum as it worked better.

Finally, we can see that the results are without appeal. GloVe produces results that are worse than the ones obtained through the LSA.