To fill this gap, we analyzed the emotional content of a large number of online product reviews using Natural Language Processing (NLP) methods. Word of mouth (WOM) has been described as one of the most important means of informal communication among consumers (Sundaram et al., 1998; Dellarocas and Narayan, 2006a; Derbaix and Vanhamme, 2003; Hennig-Thurau et al., 2004 ). This act of catharsis (Alicke et al. 1992) often takes the form of online product reviews.
Shared content such as online WOM is known to play a crucial role in emotional recovery, relief and other aspects of social interaction (Maute and Dube, 1999; Derbaix and Vanhamme, 2003; Hennig-Thurau et al., 2004). We are also able to observe changes in the emotional content of reviews over time for search and experience items. In addition, consumers engage in positive word-of-mouth behavior to express their enjoyment of the experience, increase engagement with the product, and help the company (Hennig-Thurau et al., 2004).
This is followed by another emotion, such as happiness (positive) or anger (negative), leading to the assumption that a positive or negative surprise has been elicited (Anderson, 1998; Hennig-Thurau et al., 2004). Previous research has predicted that stronger emotional responses to a product lead to the posting of customer opinions (Maute and Dube, 1999; Derbaix and Vanhamme, 2003; Hennig-Thurau et al., 2004). In other words, the emotional content of online WOM should be greater in the case of experiential goods compared to search goods.
Indeed, research on online WOM has confirmed that consumers primarily read online reviews to reduce the risks associated with a purchase and to reduce the research costs associated with the decision-making process (Hennig-Thurau et al. 2003).
EMPIRICAL STUDY
In addition, when the change of the average value of a product from the first to the second scale was greater than the average change of all products, the product was also considered an experience product (E). Therefore, based on this classification system, we can place each product in only one of two categories.7. 6In this case, the mean of the second scale must be within the standard deviation of the mean score of all products.
7In both classification evaluations, the description of the product was not present to avoid categorization bias. In addition, adjacent words were also considered in the analysis to correctly identify the words used in the classification. We then aggregated the frequencies of the emotional words used in each of the reviews based on their star ratings.
One of the key aspects of NLP is identifying keywords that carry sentiment semantics. Here we explain several important NLP concepts that we used to identify and categorize emotion words. Once this was done, the document could be represented as a vector of word counts, where the dimension of each vector is the number of unique words in the dataset.
If, on the other hand, a word occurred frequently in the positive class, but rarely in the negative class, then the word would not be considered independent of class, and the chi-square value would thus be high. We used a stop word list via an Application Programming Interface (API)8 in the Onix Text Retrieval Toolkit 5 (Lextek International, U.S.A.). Ranking by word frequency: We have ranked words by frequency, reflecting the general topics of the documents.
Feature selection: We used the feature selection method to compare the chi-squared distribution of the words based on their appearance in positive and negative reviews. Sentiment score: For each word in the top-ranked word lists, we used SentiWordNet9 to retrieve the positive, negative, and objective scores. We counted the frequencies of the words immediately following the highest-ranking words and selected the top five adjacent words in frequency.
This means that the word "memory" is used as "memory card" or "memory cards" in the user reviews of cameras. 9 SentiWordNet is a lexical resource for meaning mining based on quantitative analyzes of the glosses associated with synsets and vectorial term representations for semi-supervised synset classification.
RESULTS AND DISCUSSION
To confirm H1, which stated that more extreme ratings will have a greater proportion of emotional content than less extreme ratings, we conducted an ANOVA of the emotional content associated with each star rating for the products in our data set. ANOVA results show significant differences in the overall emotional content of extreme and non-extreme reviews (see Table 6, “Overall” column). The percentage of emotional content for the positive/negative extreme valences (1-stars and 2-stars vs. 4-stars and 5-stars) is significantly different compared to the middle valence (3-stars).
It was shown that the mean values of percentages of emotional content for the pair of 1-2 stars, the pair of 2-3 stars and the pair of 3-4 stars are statistically different. These findings support our hypothesis (H1) that more extreme ratings will have a higher proportion of emotional content than less extreme ratings. ANOVA for pairs of star ratings for total, positive, negative, emotional content of search and experience products.
Our second hypothesis stated that there is a higher proportion of positive emotional content in positive extreme evaluations compared to the proportion of negative emotional content in negative extreme evaluations. Considering only the two extreme ratings, a 5-star rating has, on average, 57.4% more emotional content than a 1-star rating. A closer look shows that there is no difference in the amount of positive emotional content in the 2- and 3-star ratings, but the 2-star rating contains significantly more negative emotional content (see Table 6, . "Positive" and "Negative" columns).
In fact, 3-star rating does not contain more negative emotional content compared to 4-star and 5-star review, indicating that it is very balanced and rational. For both search and experience products, there were significant differences in the percentage of emotional content between the extreme and intermediate valences. To further investigate these differences in the emotional content between search and experience products for each star rating, we ran multiple T-tests to compare the emotional content of corresponding star ratings.
The results (Table 7, column “All emotional words”) show that there are no significant differences in emotional content between search and experiential products. Once again we segment the emotional content into positive and negative and now we find some interesting results (Table 7, “Positive Emotion” and “Negative Emotion”). We find that experiential products have a small but significantly higher positive emotional content in their negative reviews (1 star and 2 star) compared to search products.
The results show that when only the first three months of reviews are considered, there is a significant difference in the emotional content of 5-star reviews for search vs. By compiling the results of the 3 models, we can conclude that the 3-Star rating is very balanced in terms of its emotional content.
DISCUSSION AND CONCLUSION
The results for Model 1 confirm our finding that more extreme ratings have greater emotional content than less extreme ratings, thus lending further support to H1. The dummy variable for product type is not significant, which confirms H3b that in the long run it is irrelevant whether a product is a good search or experience in terms of emotional content. This finding provides for the first time empirical evidence for the widespread assumption of such a distribution of emotional content.
Second, we found that a higher proportion of positive emotional content exists on the positive side of the rating spectrum compared to negative emotional content on the negative side of the rating spectrum, confirming a positive skew in online product reviews. Our results for H3b broadly validated this hypothesis, as we find no difference between the proportions of emotional content for search vs. These longer length, less emotional words can provide valuable information about the strengths and weaknesses of the products that the consumer has experienced.
Although a certain amount of negative emotional content will always be present, it can be of great benefit to companies located at different points in the value chain to use this information as a feedback mechanism that feeds back into a continuous quality improvement process. That is, although we know that there is more positive emotional content in positive reviews compared to negative content in negative reviews, we do not know whether these emotions have a proportional or disproportionate influence on decision making. Farrell, A Longitudinal Test of the Investment Model: The Impact on Job Satisfaction, Job Commitment, and Turnover of Variations in Rewards, Costs, Alternatives and Investments, Journal of Applied Psychology.
Implications for Revenue Forecasting and Planning, Proceedings of the 24th International Conference on Information Systems, Washington, D.C, (2004). A study of the product-specific antecedents of online movie reviews, Proceedings of the Workshop on Information Systems and Economics, Evanston, USA, (2006b). Pennock, Mining the Peanut Gallery: Opinion Extraction and Semantic Classification of Product Reviews, Proceedings of the 12th international Conference on World Wide Web Budapest, Hungary, (2003).
Empirical findings and analytical modeling of online word-of-mouth communication, Proceedings of the 7th ACM Conference on Electronic Commerce, New York, USA., (2006). However, only a limited number of studies to date have examined the content of reviews for their emotional content. We found that more extreme reviews have a greater proportion of emotional content than less extreme reviews, revealing a bimodal distribution of emotional content, thereby empirically validating a key assumption underlying much of the existing literature on online WOM.
In addition, we found that reviews have a greater proportion of positive emotional content within positive extreme ratings compared to negative emotional content within negative extreme ratings, which is an important factor in online WOM generation. Upon further investigation, we found that there is a difference in the emotional content of reviews between search and experience products in the early stages of product launch.
