3.2 Polarity Scoring
Two packages measure text polarity. The simpler one tidytext: unnest tokens, join to the Bing lexicon, and calculate the net of positive minus negative polarity counts. sentimentr is more sophisticated in that it takes into account valence shifters, surrounding words that change the intensity of a sentiment (e.g., “very”) or switch its direction (e.g., “not”).2
tidytext
The tidytext way to score polarity is to tag words as “positive” and “negative” using the Bing lexicon, then calculate the difference in counts. The qdap and sentimentr packages correct for text length by dividing by \(\sqrt{n}\). It is useful to capture the positive and negative words back in the main data frame for explaining how the polarity score was calculated.
polarity_bing <-
token %>%
left_join(bing, by = "word", relationship = "many-to-one") %>%
summarize(.by = c(review_id, sentiment), n = n(), words = list(word)) %>%
pivot_wider(names_from = sentiment, values_from = c(n, words),
values_fill = list(n = 0)) %>%
select(-c(n_NA, words_NA)) %>%
inner_join(hotel_0 %>% select(review_id, prepped_wrdcnt), by = "review_id") %>%
mutate(
polarity = (n_positive - n_negative) / sqrt(prepped_wrdcnt),
polarity_desc = if_else(polarity >= 0, "Positive", "Negative")
)
polarity_afinn <-
token %>%
inner_join(afinn, by = "word", relationship = "many-to-one") %>%
summarize(.by = review_id, sentiment = sum(value), words = list(word)) %>%
inner_join(hotel_0 %>% select(review_id, prepped_wrdcnt), by = "review_id") %>%
mutate(
polarity = sentiment / sqrt(prepped_wrdcnt),
polarity_desc = if_else(polarity >= 0, "Positive", "Negative")
)
# Attach to main data frame
hotel_1 <-
hotel_0 %>%
left_join(polarity_bing %>%
select(review_id, polarity, words_positive, words_negative) %>%
rename_with(~paste0("bing_", .x)),
by = join_by(review_id == bing_review_id)) %>%
left_join(polarity_afinn %>%
select(review_id, polarity) %>%
rename_with(~paste0("afinn_", .x)),
by = join_by(review_id == afinn_review_id))Let’s see how the polarity scores compare.
hotel_1 %>%
pivot_longer(cols = c(bing_polarity, afinn_polarity),
names_to = "lexicon", values_to = "polarity") %>%
filter(!is.na(polarity)) %>%
ggplot(aes(x = polarity, y = fct_rev(hotel))) +
geom_boxplot() +
facet_wrap(facets = vars(lexicon)) +
labs(title = "Review polarity", x = NULL, y = NULL,
caption = glue::glue("Bing Polarity = (n_pos - n_neg) / sqrt(n_words)\n",
"AFINN Polarity = sentiment / sqrt(n_words)"))
The two lexicons are similar. The data set includes a rating (1-5). I’ll stick with Bing going forward for convenience. The polarity score should correlate with the rating.
hotel_1 %>%
filter(!is.na(bing_polarity)) %>%
ggplot(aes(x = as_factor(rating), y = bing_polarity)) +
geom_jitter(width = 0.2, alpha = 0.3, color = "#5DA5DA", size = 1) +
geom_boxplot(alpha = 0) +
theme_minimal() +
labs(title = "Polarity is associated with overall Likert score",
x = "Overall Likert Rating", y = "Polarity Score")
Sentiment increases with Likert rating, but there are many reviews with a rating of 5 and a polarity score <0. In some cases this is because the reviewer interpreted the scale incorrectly. You can use polarity scores to identify problematic reviews like these.
hotel_1 %>%
mutate(
problematic = case_when(
(rating == "1-2" & bing_polarity > 0.5) ~ "Too Low",
(rating == "5" & bing_polarity < -.5) ~ "Too High",
TRUE ~ "Other"
)
) %>%
filter(problematic %in% c("Too High", "Too Low")) %>%
group_by(problematic) %>%
slice_max(order_by = abs(bing_polarity), n = 1) %>%
select(problematic, rating, bing_polarity, review) %>%
flextable::flextable() %>%
flextable::autofit() %>%
flextable::valign(valign = "top")problematic | rating | bing_polarity | review |
|---|---|---|---|
Too High | 5 | -1.278019 | Some design faults in the bathroom no stool misplaced grab handles and vanity mirror Very disappointing experience in Savoy Grill Good quality ingredients but poorly presented and tasteless Numerous mistakes in service including charging for expensive drinks which we did not have Service charge revoked and booking for following night cancelled |
Too Low | 1-2 | 1.774014 | Having stayed at several Morgans Hotels on both sides of the pond I was really looking forward to checking in to the Mondrian London I have to say I was bitterly disappointed and left questioning whether this hotel was actually a Morgans property Normally the second you step into any Morgans property you instantly know youre in a Morgans Hotel They always have the wow factor amp the dcor has such a quirky distinctive style and the service is always exemplary without being stuffy The same cannot be said about Mondrian London we could have been in any generic highend hotel in the world and some of the staff behaviour we witnessed was highly substandard The room was nice but nice is not a word I would normally use to describe a Morgans room its usually amazing iconic incredible It was stylish but had no soul amp dcor wise we could have been in a room in any hotel in the world However the view of the Thames amp St Pauls was fantastic and was a redemptive feature amp the large marble bathroom which is synonymous with Morgans was amazing amp without doubt the highlight of the room The bed was enormous but was so hard it was like sleeping on breeze blocks possibly one of the worst nights sleep Ive ever had had to get up in the night amp go amp sleep on the couch not really what you expect from a 500 room We had aperitifs in the Dandelyan bar again the dcor is completely unremarkable nothing special when compared to the likes of the groups other bars but was comfortable enough We had really high hopes for this bar as it is apparently run by an award winning mixologist so we couldnt wait to sample to cocktails and wait we did while we pored over the enormous incomprehensible menu It was more like a botanical reference book than a cocktail menu gargantuan in size and too wordy with no comprehensible flow People know what they like to drink either in terms of the spirit base of their cocktail vodka rum gin whiskey or the type of drink they prefer martini flute long etc therefore it is a sensible idea to arrange menus in a way that people can easily access the information they are looking for as most high end places do We got so fed up of trying to look for something suitable that we gave up amp ordered a glass of Champagne which was served warm amp flat Quite ironic that these supposed award winning bartenders werent even able to get a glass of Champagne right But the waitress not sure of her name but she was of an eastern heritage was very welcoming and friendly with a constant smile so made up for the poor drinks We went through to dinner and it was fantastic the food was all executed perfectly amp really tasty amp well presented Service was also very good but we found the waiter to be quite formal and stiff which was quite surprising in a Morgans He was very efficient and slick but wasnt very engaging and definitely needed to smile more The dcor of the restaurant was very elegant and had great views across the Thames There was a clear managerial presence on the floor which always instils confidence in you as a diner The Irish manager checked on our table and the other tables around us a touch that was very much appreciated Breakfast was also excellent there I would wholeheartedly recommend the Sea Containers restaurant as a standalone venue We had drinks up at the Rumpus Room amp were welcomed by a very colourful character who instantly put a smile on our faces amp gave us a very warm welcome not sure if he was the manager or host gentleman with a beard amp earrings Again the bar was nice dcor wise it had a classy feel to it amp the views across the Thames were spectacular This time we opted to sit at the bar so we could speak to the bartenders about the drinks amp avoid another negative experience like we had in Dandelyan We were served by great bartender cant recall his name blue eyed chap who I believe told us he was from Hungary He was friendly amp chatty amp made us feel welcome without being intrusive He made us some bespoke cocktails based on our preferences which were exquisite Faith restored we opted to order the next round from the menu big mistake they pre make their cocktails in batches amp literally just pour it out of a bottle onto ice You may expect this kind of cutting corners from a highvolume lowend venue but we certainly did not expect this from a supposed highend cocktail bar in a luxury hotel When you are paying around 15 for a cocktail plus service charge you expect your drinks to be freshly made amp for there to be a bit more service than simply opening a bottle I couldnt believe what I was seeing The bartender made us some more bespoke cocktails instead which again were great if you visit this bar ask for the bartenders to make you something based on your likes amp avoid their menu We moved to a table amp had several different waitresses check on us amp serve us all of whom were friendly amp attentive The ambience was good in the bar amp it is great for people watching but one thing we did notice which we thought was highly inappropriate was the lady with the short hair amp American accent not sure what her role was but she seemed to work there as went behind the bar amp was talking well shouting across the room actually to all the other staff openly imbibing in the bar In our experience of high end hotels the staff dont normally openly drink to the point of belligerence in front of the guests during hours of service so we were quite shocked by this amp didnt think having a drunk member of staff on the floor was acceptable So this amp the pre made cocktails let this bar down amp clouded our otherwise good experience Overall I would say this hotel amp its facilities are mediocre at best They are ok but when you are paying five star prices you expect a five star experience If you are a Morgans aficionado like me then avoid this place and stay at Sanderson or St Martins Lane instead where you will receive the full Morgans experience But I would certainly recommend the Sea Containers restaurant excellent food amp good service in a beautiful location And if you dont mind a drunken member of staff shouting across the room amp having to sit at the bar to order with the bartenders then I would also recommend Rumpus Room |
The polarity words can help explain why some hotels rated poor or excellent.
token %>%
inner_join(hotel_1 %>% filter(rating %in% c("1-2", "5")), by = join_by(review_id)) %>%
filter(!word %in% c("hotel", "stay", "night")) %>%
filter((rating == "5" & bing_polarity > 0) |
(rating == "1-2" & bing_polarity < 0)) %>%
count(rating, word) %>%
mutate(.by = rating, pct = n / sum(n)) %>%
group_by(rating) %>%
slice_max(order_by = pct, n = 10) %>%
ggplot(aes(x = pct, y = reorder_within(word, by = pct, within = rating))) +
geom_col() +
scale_y_reordered() +
scale_x_continuous(labels = percent_format(1)) +
labs(y = NULL, x = NULL) +
facet_wrap(facets = vars(rating), scales = "free_y")
Word clouds are a nice way to get an overview of the data.
token %>%
inner_join(hotel_1 %>% filter(rating %in% c("1-2", "5")), by = join_by(review_id)) %>%
filter(!word %in% c("hotel", "stay", "night", "london"),
!is.na(bing_polarity)) %>%
mutate(polarity_desc = if_else(bing_polarity > 0, "Positive", "Negative")) %>%
count(word, polarity_desc, wt = prepped_wrdcnt) %>%
pivot_wider(names_from = polarity_desc, values_from = n, values_fill = 0) %>%
data.table::data.table() %>%
as.matrix(rownames = "word") %>%
wordcloud::comparison.cloud(max.words = 30, title.size = 1.5, scale = c(1, 3.5))
sentimentr
sentimentr calculates polarity at the sentence level. It improves on tidytext in that it takes into account the context in which the sentiment words occur by incorporating valence shifters.
- A negator flips the direction of a polarizing word (e.g., “I do not like it.”).
lexicon::hash_valence_shifters[y==1]. - An amplifier intensifies the impact (e.g., “I really like it.”).
lexicon::hash_valence_shifters[y==2]. - A de-amplifier (downtoner) reduces the impact (e.g., “I hardly like it.”).
lexicon::hash_valence_shifters[y==3]. - An adversative conjunction overrules the previous clause containing a polarized word (e.g., “I like it but it’s not worth it.”).
lexicon::hash_valence_shifters[y==4].
sentimentr uses a lexicon package combined from the syuzhet and lexicon packages. Positive words are scored +1 and negative words are scored -1. sentimentr identifies clusters of words within sentences of the text. The 4 words before and 2 words after are candidate valence shifters. Polarized words are weighted by the valence shifter weights: negators = -1; amplifiers and de-amplifiers = 1.8; adversative conjunctions decrease the value of the prior cluster and increase the value of the following cluster. Neutral words hold no value, but do affect the word count.
hotel_sentimentr <-
sentimentr::get_sentences(hotel_1$review) %>%
sentimentr::sentiment() %>%
summarize(.by = element_id, sentimentr_polarity = mean(sentiment))
hotel_2 <-
hotel_1 %>%
mutate(element_id = row_number()) %>%
inner_join(hotel_sentimentr, by = join_by(element_id)) %>%
select(-element_id)Let’s see a few examples where sentimentr differed from tidytext. Looks like bing did a better job on the first one, but sentimentr was better on the next two.
hotel_2 %>%
filter((bing_polarity > 0.2 & sentimentr_polarity < -0.2) |
(bing_polarity < -0.2 & sentimentr_polarity > 0.2)) %>%
select(review, bing_polarity, sentimentr_polarity) %>%
head(3) %>%
flextable::flextable() %>%
flextable::autofit()review | bing_polarity | sentimentr_polarity |
|---|---|---|
Well I am no strange to London's 5star hotels and when a new one comes along I am eager to try So I heard about the corinthia via some friends who work in the industry and said their bar was really something Until the incident I had stay there a total of 3 times but on the 4th time in June I was on my second of 4 nights and was about get some cash from my room as I was going to exchange money for some dollars as I was flying to the US pretty soon When I came to my money it felt light When I counted it it was almost 1000 down Obviously shocked I contacted reception They started an investigation Well a couple of days later I had to checkout and the matter was not resolved I was made to pay my bill in full Upon my return to London I met up with a manager at the hotel Jean Louis He said the matter was still under investigation and he gave me his card to get in touch Nothing was ever resolved This hotel does not only boast to be one of the best hotels in London but one of the best in the world You pay premium for this reason and you expect a premium service This hotel epically failed me and dented my confidence in the hotel industry I didn't even get a formal apology It's a lovely hotel but I would have to say stay at your own risk but for over 500 a night should there be any risk | -0.2279212 | 0.2346432 |
was so looking forward to staying here had high expections what a let down it was we had a connecting room as we took the children rooms are small they are clean but for over 1000 a night u expect a bit more there was five of us altogether and felt like we was crammed in staff are rude I've stayed a cheaper hotels and got treated better | -0.2294157 | 0.3910338 |
The service the Food the room all 100 i simply can't fault The Savoy i have stayed all over the world and paid huge sums The Savoy is in my mind the level that all Hotels should be judged when we stay in the City for work or to just to take in a show we always choose The Savoy | -0.2182179 | 0.2017928 |