reviews

The reviews data frame is the core component of the shoppingwords package, designed for text analysis, sentiment scoring, or NLP tasks (e.g., customer feedback mining).

It contains a large number of reviews collected from a Turkish e-commerce site.

See the Table for the data dictionary for this data frame.

Table: Data dictionary for the reviews data frame.

glimpse(reviews)

## Rows: 260,308
## Columns: 3
## $ rating  <dbl> 5, 5, 5, 5, 1, 5, 5, 5, 2, 3, 1, 1, 3, 3, 1, 1, 1, 1, 3, 1, 4,…
## $ comment <chr> "Yorum gözüksün diye 5 yıldız verdim 5li paket diye sipariş ve…
## $ id      <dbl> 3573, 3975, 4910, 4950, 5908, 6144, 6192, 6335, 6370, 6371, 63…

English translations of the few comments visible in the output above are given below:

# $ rating <dbl> 5, 5, 5, 5, 5, 1, 5, 5, 5, 5, 2, 3, 1, 1, 3, 3, 1, 1, 1, 1, 1, 3, 1, 4,...
# $ comment <chr> "I gave 5 stars so that the comment would be visible I ordered a 5-pack and ...
# $ id <dbl> 3573, 3975, 4910, 4950, 5908, 6144, 6192, 6335, 6370, 6371,...```

The dataset spans over 3 years (with the earliest review from April 16, 2022 and most recent from June 11, 2025), capturing reviews from users, including 260,308 rows, in customer feedback.

The examples have been translated to English, as the dataset contains reviews in Turkish.

stopwords_tr

stopwords_tr contains commonly found in Turkish shopping-related text and are useful for text preprocessing and analysis.

For example, it contains some words pertaining to clothing sizes.

stopwords_tr |> 
  slice(c(37:39))

## # A tibble: 3 × 1
##   word 
##   <chr>
## 1 xxs  
## 2 xs   
## 3 s

This dataset can supplement stopwords-iso, which is a standardized multilingual stopword collection, including Turkish but potentially with broader coverage.

stopwords::stopwords("tr", source = "stopwords-iso") |> 
  head(n = 5)

## [1] "acaba"      "acep"       "açıkça"     "açıkçası"   "adamakıllı"

English translations of the few comments visible in the output above are given below:

# "I wonder"   "perhaps"    "clearly"    "frankly"    "thoroughly"

Using both datasets reduces the risk of missing critical stopwords that could skew sentiment or topic modeling results.

phrases

The phrases data frame contains phrases that can assist in text processing, linguistic analysis, and NLP applications, making it easier to analyze customer behavior and responses.

phrases |>
  slice(c(7:8))

## # A tibble: 2 × 1
##   word                         
##   <chr>                        
## 1 5 ulduz verdimki yorumum üstə
## 2 5 verdim ki öne çıksın

English translations of the comments visible in the output above are given below:

# "I gave 5 stars to pin my comment to the top"
# "I gave 5 to pin to the top" 

Using match_stopwords() to remove stop words

The match_stopwords() function processes user reviews by removing predefined stopwords while preserving the original rating scores.

It takes a dataframe with “comment” and “rating” columns, cleans each review by filtering out stopwords from both custom stopword lists and stopwords-iso, and returns a dataframe with an additional cleaned_text column for further analysis.

This function helps in text normalization for sentiment analysis and rating-based insights.

To demonstrate a practical analytical use case for the preprocessed review data and how the package’s function work together in a real analysis workflow is given in the examples.

The example calculates if longer reviews correlate with higher/lower ratings.

cleaned_reviews <- match_stopwords(reviews)  # Remove stopwords 
cleaned_reviews |>    
  group_by(rating) |>   
  summarise(avg_text_length = mean(nchar(cleaned_text)))  

## # A tibble: 5 × 2
##   rating avg_text_length
##    <dbl>           <dbl>
## 1      1            65.7
## 2      2            59.8
## 3      3            52.7
## 4      4            48.0
## 5      5            43.1

The function can be applied to another sample dataset as well.

reviews_sample <- tibble(
    comment = c(
      "Bu ürün xs ancak fiyatı yüksek gibi", 
      "Fiyat çok pahalı ama kaliteli iyi"
    ), 
    rating = c(4.5, 3.0)
)

cleaned_sample <- match_stopwords(reviews_sample) 

English translations of the text in comment and cleaned_text columns in the output above are given below:

# comment                                                    
# [1] "This product is xs but seems realy expensive"            
# [2] "The price is not very expensive it's high-quality good" 

# cleaned_text
# [1] "the product price high" 
# [2] "the price expensive high-quality"

reviews_test

The shoppingwords package also includes a reviews_test data for text mining and NLP tasks.

This data contains user-generated reviews labeled as positive (p) or negative (n), being different than reviews data and making it useful for training models in sentiment classification.

See the Table for the data dictionary for this data frame.

Table: Data dictionary for the reviews_test data frame.

Below is a preview of the test dataset, which includes customer ratings, review text, and sentiment labels.

Each row represents user feedback, providing insights into product perception and satisfaction.

reviews_test |>
  slice_head(n = 3)

## # A tibble: 3 × 4
##   rating text                                                      emotion    id
##    <dbl> <chr>                                                     <chr>   <int>
## 1      5 Kesinlikle alın aldırın pahalı markalara o kadar para ve… p           1
## 2      5 10 yıldız olsa veririm super                              p           2
## 3      5 ürün 5 yıldız hakkediyor, ayrıca hem m hem de l beden al… p           3

English translations of the comments in the text column in the output above are given below:

# [1] Definitely buy it! No need to spend so much on expensive brands; the fit looks incredibly good. 
# [2] If I could give 10 stars, I would. Super!  
# [3] The product deserves 5 stars. I bought both M and L sizes, and they fit the same. I really liked it.

Let’s take a look at the emotion distribution in this data frame:

reviews_test |> 
  count(emotion, sort = TRUE)

## # A tibble: 2 × 2
##   emotion     n
##   <chr>   <int>
## 1 p        1184
## 2 n         297

Summarizing the occurrences of rating and emotion in relation to negative and positive reviews is insightful.

Notably, 154 reviews contain negative expressions even though they have been assigned a rating of 5, out of a total of 1481 reviews.

reviews_test |> 
  count(rating, emotion, sort = TRUE) |> 
  arrange(desc(rating))

## # A tibble: 10 × 3
##    rating emotion     n
##     <dbl> <chr>   <int>
##  1      5 p         874
##  2      5 n         154
##  3      4 p         241
##  4      4 n          20
##  5      3 p          49
##  6      3 n          23
##  7      2 n          24
##  8      2 p          13
##  9      1 n          76
## 10      1 p           7

The reviews_test can be used in various ways such as capturing the negative expressions or predicting sentence-level polarity using the reviews.

This discrepancy between user ratings and comments is visualized in the Figure below, which shows the distribution of positive and negative reviews across all rating levels.

reviews_test |> 
    count(rating, emotion) |> 
    ggplot(aes(x = factor(rating), y = n, fill = emotion)) +
    geom_col(position = "dodge") +
    scale_fill_manual(
      values = c("p" = "lightblue", "n" = "darkred"),
      labels = c("p" = "Positive", "n" = "Negative")
    ) +
    labs(
      x = "User Ratings",
      y = "Number of Reviews",
      fill = "Polarity"
      
    ) +
    theme_minimal() +
    theme(legend.position = "right")