Have a Nice Flight! Understanding the Interplay Between Topics and Emotions in Reviews of Luxury Airlines in the Pre- and Post-Covid-19 Periods

Abstract

Few studies assessed the impact of Covid-19 on the aviation industry from the passengers’ perspective. This study examined how airline passengers’ emotions (positive and negative) and sub-emotions (joy, trust, surprise, anticipation, fear, sadness, anger, and disgust) changed in the pre-and post-Covid-19 periods. 21,463 reviews from 2014–2022 of top 10 luxury airlines were extracted from Skytrax. Using the lens of the Appraisal Theory of Emotion, the analysis revealed an increase in negative emotion and related sub-emotions after the pandemic. Using topic modeling seven similar topics (namely food, staff, customer service, board, flight, crew, and luggage) and four dissimilar topics (entertainment and drink for pre-Covid-19 and wait and Covid-19 for post-Covid-19) were identified. Regression analysis showed that the topics food and entertainment generated significant positive emotion whereas wait and customer service generated significant negative emotion. The results would help the luxury airlines to identify offerings to improve during the recovery after Covid-19.

Appendix

1.1 Emotion Analysis

To conduct emotion analysis based on NRCLex, a lexicon emotion score matrix was created as $\it {{E}}=\{{{{e}}}_{{{i}}},{{j}}\}$, where i denoted the $\it {{{i}}}^{{{th}}}$ word in the document and $\it {{j}}$ represented the $\it {{{j}}}^{{{th}}}$ emotion in set $\it {{E}}$. Word frequency analysis was used on the collected reviews using the lexicon corpus to create sparse vectors ${[v}_{{{i}}}]$,where $[v]={frequency\ of\ word} \ i \ {in\ a\ review}\ t$. We calculated the scores for the eight sub-emotions for each review by calculating the dot product of the sparse vector and the emotion matrix.

This yielded a vector of scores $\it {[{{s}}}_{{{t}},{{e}}}]$, for each  and each review $\it {{t}}$ where $E = {set\: of\: sub-emotions}= {anger, anticipation, disgust, fear, joy, sadness, surprise, trust}$. Equations (4) and (5) were used to categorize the sub-emotions into positive emotion (E_POS) and negative emotion (E_NEG) respectively.

$$\lbrack{s}_{ t, e}\rbrack={ V}_{ t}.{ E}_{ T},{ E}_{ T}={Transpose}\;{of}\;{matrix}\; E.$$

(4)

(5)

1.2 Topic Modeling

The LDA model provided two probability distributions, S (topic | review) and S (word | topic) which was determined by $\it {{S}}\left({{z}}|{{d}}\right)$, with each topic described by a set of words following another probability distribution (i.e., $\it {{S}}\left({{t}}|{{z}}\right)$. This is shown in Eq. (6):

$${{S}}\left({{ti}}|{{d}}\right)={\sum }_{{{J}}=1}^{{{Z}}}{{S}}\left({{ti}}|{{zi}}={{j}}\right)={{S}}({{zi}}={{j}}|{{d}})$$

(6)

1.3 Perplexity and Coherence Analysis

The performance of a topic modeling method depended on the number of topics. To determine the optimal number of topics, coherence and perplexity values were calculated. Perplexity was calculated as the normalized log-likelihood of the data as shown in Eq. (7). The lower the perplexity, the better was the choice of number of topics.

$${{perplexity}= 2\times {{{exp}}}^{{{L}}}}$$

(7)

where $\it {\text{L}}$ was the log likelihood, which indicated the probability of observing new data. Sometimes the analysis showed that perplexity was not perfectly related to human judgment, so the coherence score was used to obtain more accurate results. The score was determined by calulating the sum of pairwise similarity scores of words in a topic as shown in Eqs. (8) and (9). The higher the score, the better it was.

$${\it{Coherence}={\sum }_{{\text{i}}<{\text{j}}}{\text{score}}({{\text{w}}}_{{\text{i}}},{{\text{w}}}_{{\text{j}}})}$$

(8)

$${\it{score}\left({{\text{w}}}_{{\text{i}}},{{\text{w}}}_{{\text{j}}}\right)={\text{log}}\frac{{\text{p}}({{\text{w}}}_{{\text{i}}},{{\text{w}}}_{{\text{j}}})}{{\text{p}}{({\text{w}}}_{{\text{i}}}){\text{p}}{({\text{w}}}_{{\text{j}}})}}$$

(9)

where $\it {\text{p}}{({\text{w}}}_{{\text{i}}})$ denoted the probability of occurrence $\it {{\text{w}}}_{{\text{i}}}$ in a random document and $\it {\text{p}}\left({{\text{w}}}_{{\text{i}}},{{\text{w}}}_{{\text{j}}}\right)$ denoted the probability of co-occurring of both and in a random document. Figures 9 and 10 show the perplexity and coherence scores for determining the number of topics in the reviews in the pre- and post-Covid-19 periods based on topic frequency.

Fig. 9

Perplexity and coherence scores for reviews from the pre-Covid-19 period

Fig. 10

Perplexity and coherence scores for reviews from the post-Covid-19 period

Figures 9 and 10 showed that the perplexity score decreased dramatically for both time periods whereas the coherence score oscillated over the number of topics considered. We tested the optimal number of topics, including 8, 9, and 10, and then LDA was computed for the different number of topics. Face-validation indicated that the topics would repeat after 9th topic, so we considered 9 as the optimal number of topics for both periods.

1.4 Regression Results for the Pre- and Post-Covid-19 Time Periods

Figures 11, 12 and 13 show the regression coefficients of emotions and sub-emotions of reviews for different topics over different time periods.

Fig. 11

Regression coefficients for emotions and sub-emotions of reviews from both time periods

Fig. 12

Regression coefficients for emotions and sub-emotions of pre-Covid-19 reviews

Fig. 13

Regression coefficients for emotions and sub-emotions of post-Covid-19 reviews