Francisco Villarroel Ordenes, Stephan Ludwig, Ko de Ruyter, Dhruv Grewal, Martin Wetzels
«Unveiling What Is Written in the Stars: Analyzing Explicit, Implicit, and Discourse Patterns of Sentiment in Social Media»
Journal of Consumer Research, vol. 43, n.º 6 (2017)
Journal of Consumer Research (@JCRNEWS) | Oxford University
Extracto de apartados en páginas 29-39 de la publicación en PDF. Véanse las referencias en la publicación original del texto.
«DISCUSSION
»Extending Extant Research
»By zooming in on how consumers’ written reviews reflect differential, asymmetric sentiment strength, and how sentence patterns might exert direct negative effects on overall sentiment expressions, we contribute to the literature on consumer sentiments expression and improve predictions of subsequent consumer behavior. By empirically validating the hypothesized relationships and addressing their relevance and generalizability, we extend extant research in three ways.
»First, to decode consumer sentiments and their influence, prior consumer research has relied on simple word frequencies, such as the number of positive or negative emotion words in verbatim customer reviews and posts (Hennig-Thurau, Wiertz and Feldhaus 2014). By accounting for activation level differences, innate to sentiment expressions (Russell and Barret 1999), and the influences of certainty (Pennebaker et al. 2007), we augment such approaches. In particular, disentangling positive or negatively valenced emotions and the degree of certainty with which they are expressed significantly improves estimates of consumer sentiments.
»On the one hand, compared with positive low activation and/or attenuated sentiment expressions, the use of positive high activation and or boosted expressions doubles the probability of a higher star rating designation by a consumer. On the other hand, compared with negative low activation and/or attenuated sentiment expressions, the use of negative high activation and/or boosted expressions did not double the probability of a lower star rating designation by a consumer.
»In particular, we failed to find a significant difference between highly activated and/or boosted and low activated and/or attenuated negative expressions in book reviews. Sentiment expression in book reviews thus depends at least partially on the context, so “sad” might be an appreciated feature for a tragedy genre, and “disgusting” might describe the antagonist character. Our findings which reflecting differences in hotels compared to books is also in line with research on affect suggesting that the use taxonomic structures regarding to emotion, might not work across contexts in the same way (Russell and Barret 1999).
»This is an important finding for research in sentiment analysis, which is highly dependent on word taxonomies associated with sentiment and activation. Overall, in line with Russell and Barret (1999) and Sbisa (2001), we empirically demonstrate the importance of considering the nuanced relationship among sentiment force features (i.e., activation level, certainty, tentative and negations), and overall sentiment in online consumer reviews.
»Second, SAT predicts that assertive, commissive, and directive speech acts implicitly convey the speaker’s sentiment, without using explicit emotion words (Searle 1975). We predict and find that such “emotionless,” implicit acts relate asymmetrically to consumers’ overall sentiment. Implicit sentiment features appeared in about 16% of consumer reviews, and in line with our hypotheses, we found that positive (negative) directive and commissive acts exerted stronger effects on sentiment strength than did assertive acts. The linguistic context suggests that generic assertions hotel reviews (e.g., “We stayed in a superior double room,” “Rooms were clean”) may not really have an effect on the overall sentiment as they are only aligned with general expectations.
»Furthermore, commissive language tended to be used more in hotel but less in book reviews, likely because it is generally less common to commit to read a book again (once in a life product experience), whereas returning to a certain hotel is a likely option. Our findings also contribute to conceptualizations of implicit sentiment expressions (Feldman 2013; Montoyo, Martínez-Barco, and Balahur 2012), in that we introduce and empirically validate a theoretical framework of emotionless speech acts.
»Third, we underscore the necessity of considering the message development (van Dijk 1997) and contribute to conceptualizations of sentiment dynamics (Schweidel and Moe 2014) by exploring how sentence-level developments reflect consumers’ sentiments. A consumer’s overall sentiment is likely negative if the development of the sentiment expressions in a review (explicit and implicit) are incoherent. In line with SAT and discourse literature (van Dijk 1997), as well as the concept of consumer ambivalence (Otnes, Lowrey, and Shrum 1997), we verify that relative incoherence across all review sentences decreases the overall positivity of the sentiment. Our exploratory analysis of positive and negative trends across sentences drove consistent and interesting results.
»On the one hand, we found that positive trends reflect a more negative consumer sentiment overall. Smyth (1998) justifies the association between more negative reviews and positive trends (e.g., decreasing negativity) on the inherent curative process of writing, which provides assimilation and understanding of the negative event. This is also in line with Pennebaker and Seagal (1999), who conceptualizes writing as a process by which people can confront upsetting topics.
»On the other hand, negativity trends are associated with more positive reviews. This finding consistent with empirical research suggesting that positive reviews start with their most activated emotions (e.g., “the hotel was a disaster”) and then dilutes through a constellation of supporting statements (e.g., “I had an issue with the personnel”; De Ascaniis 2013).
»Corroborating Extant Research
»Consumer review phenomena stimulate extensive, insightful research to uncover relations between text-based sentiments and retail performance, yet we still lack a good synthesis of the divergent sentiment analysis approaches (Schweidel and Moe 2014). In this empirical, theorydriven approach, we achieve some corroboration of extant research findings though. For example, in line with Barasch and Berger (2014) and Schweidel and Moe (2014), we confirm that the general presence of positive emotion words relates to more positive consumer sentiment overall.
»However, we find that specific sentiment expressions can also be context dependent in terms of the product/service and the social media platform (Schweidel and Moe 2014). For example, while implicit sentiment expressions through commissive language are very frequent in hotel reviews (e.g., “I will come back”), they are rather an exception the book evaluations (i.e., it is rather uncommon to say “I will read this book again”). In addition the heterogeneity across platforms plays an important role in how consumers express their sentiment. Product evaluations in online reviews are in average 8 sentences long, while in twitter and Facebook 1.6 and 2.7 sentences in average. As such, social media platforms force consumers to be more explicit and brief regarding their overall sentiment strength.
»This is in line with the significant effects of explicit and highly activated and/or boosted sentiment force indicators and the non-significant effect of implicit sentiment expressions on overall sentiment strength. The latest changes in Twitter and Facebook providing consumer more character spaces and new emoticons might be a response to the need of a more complete sentiment expression (Bloomberg 2016; Wired 2016). Our findings that weekly sentiment changes in the verbatim consumer reviews (derived using our algorithm) influence future sales ranks also emphasize the importance of improving the accuracy or precision of sentiment analysis.
»First, we corroborate research by Chevalier and Mayzlin (2006; modeling details provided in appendix 1) by finding that sales on online retail sites are significantly influenced by price fluctuations. Furthermore, in line Ludwig et al. (2013), who suggest that book reviews are processed heuristically, we corroborate that consumers avoid informational overload and resort to heuristic processing, screening for the most easily accessible indicators, which are affect word cues (hence the effects of sentiment and valence).
»The result that particularly negative directives (being the strongest class of speech acts) impact sales is also in line with the findings of this paper, which suggest that more negative will always hurt sales more, meanwhile positivity (especially if it gets too much) gets scrutinized at some point. We corroborate and support the latest marketing research on text mining by suggesting that the focus should extend beyond single words, to include the discourse patterns of sentences and entire paragraphs.
»This suggestion goes in line with moving sentiment analysis research from a “bag of words” to a “bag of sentences” (Buschken and Allenby 2015) and in turn giving researchers and managers a more comprehensive understanding of the individual and aggregated intentions (speech acts) included in product and service evaluations.
»Finally, our findings link to research in psycholinguistics (Pennebaker et al. 2007). In hotel reviews, consumers use first-person pronouns with more positive sentiment, whereas in book reviews, their usage shows the opposite connection. According to Chung and Pennebaker (2007), this finding might reflect the difference in singular versus plural first-person pronouns. First person plural relates more to shared positive experiences whereas singular (e.g., “I” or “myself”) pronouns connect more to negative experiences and depression (Chung and Pennebaker 2007). In fact, we found that hotel reviews showed an almost equal use of first person pronouns in singular and plural (a ratio of 1:1), while book reviews were characterized by the use of mainly first person pronouns singular compared with plural (a ratio of 2:1). This different use of plural versus singular in the two review contexts explains why we find a positive impact in hotels and a negative one on books.
»LIMITATIONS AND FURTHER RESEARCH
»We note the massive potential for further studies on how different patterns of sentiment can drive subsequent consumer behavior (Ludwig et al. 2014). By theorizing about speech acts, this article informs sentiment analysis, resulting in a greater understanding of how consumers express sentiment in product and service reviews. Several limitations of our study also provide worthwhile avenues for continued research.
»First, consumer research often uses direct inverses of the sentiment of a negated valence word (e.g., from positive to negative or vice versa; Ghose, Ipeirotis, and Beibei 2012). Our more granular revision of negations instead showed that for book reviews, negations of negative high expressions (e.g., “not horrible” or “not too bad”) have attenuation effects but do not reverse the meaning completely. Unlike a logical negation, a phrase such as “the service wasn’t horrible” does not translate to its equivalent in positive terms, such as “it was amazing.” Building on this finding, research should zoom in on the differential impacts of negations in customer reviews and social media, which could enhance understanding of the language in user-generated content.
»Second, we propose a new, metric-based approach to improve understanding of sentiment expression and its components, but we do not establish a new class of probability models for sentiment analysis. This important task is beyond the scope of our paper; it also is being addressed by recent developments in computer linguistics and machine learning. In this sense, we view our work as complementary: It provides a theoretical basis for a better elaboration of sentiment analysis and other models derived from language. Regarding our dictionary approach, further research could assess the diverse implications of word taxonomies as the ones suggested by Tausczik and Pennebaker (2010) and Whissell (2009). Further research could also incorporate our findings and assess their implications in other context such as sentiment in voice or videos (Poria et al 2016) and also through other learning algorithms, such as support vector machines and hidden Markov models (Mao and Lebanon 2007; Thelwall et al. 2010).
»Third, despite finding relative differences in how sentiment is expressed in book versus hotel reviews, we did not test specifically whether the different contexts prompted different sentiment expressions. According to SAT, linguistic propositions reflect considerations of the referee or subject (Searle 1969), so a book review likely features a combination of the reader’s experience with the character, story, and plot, whereas sentiment toward a hotel more commonly is conveyed in terms of the customer experience. Additional research could seek to uncover the relation between sentiment and its linguistic context, possibly with nested logit models (Farley, Hayes, and Kopalle 2004).
»Fourth, Luna and Peraccio (2005) note the importance of considering multiple consumer languages in marketing decisions. Although our approach only focuses on English reviews, it would be interesting to study how sentiment is expressed in different languages or different English-speaking countries, to identify implications for decoding consumer sentiments. Further research could apply SAT to assess how different types of speech acts, translated into various languages, exert distinct effects on the overall sentiment expression.
»Fifth, sentiment connotations in customer reviews are not always literal. Ironic or sarcastic connotations use subtleties to communicate meanings opposite those of the actual words (Gopaldas 2014; McGraw, Warren and Kan 2015). Further research might investigate linguistic properties that characterize ironic statements, to help identify the sentiment orientation of user-generated content and enable companies to avoid erroneous sentiment predictions.
»Sixth, we used regular expressions to retrieve commissive, directive, and assertive speech acts, not an exhaustive compilation of non-expressive speech acts that implicitly convey sentiment. This current approach indicated that 16% of the reviews contained at least one of these speech acts. Further text mining studies might improve the retrieval mechanisms for detecting implicit sentiment expressions. Although the automated classification of speech acts is a relatively new area (Zhang, Gao, and Li 2011), developments in the detection of varying speech acts might reveal additional implications of consumers’ reviews. A recent meta-analysis (Purnawirawan et al. 2015) indicates that review valence is key for influencing further consumer recommendations, though a focus on explicit valenced language might mask the effect of commissive, directive, and assertive language.
»Seventh, further research could look into the individual effects of certainty and tentative words (boosters and attenuators) when combined with valenced words (i.e., control condition) and their differential impact on sentiment. Our analysis provided an aggregated overview of positive/negative high vs. low including features of language such as negations, certainty and tentative words. However, we believe that these more granular components and other function words can be studied individually in further research. It would contribute to understand how the interaction of content words together with booster and attenuators has an impact on consumers’ emotional states and behaviors.
»Eight, we encourage researchers to further explore discourse patterns such as trend. Our study provides an exploratory analysis, regarding broad types trend (positive and negative), however there might be more specific types of trends such as from positive to negative, from negative to more negative or from positive to more positive, that are worth studying. Literature in argumentation patterns (e.g., consequential argumentation; Walton 1999), narrative (e.g., genre; Gergen and Gergen 1988) and also psychology literature (e.g., writing as a curative process; Pennebaker and Seagal 1999) could be helpful for researchers interested in this topic.
»A final avenue for further research is to explore curvilinear effects related to extreme positive (negative) reviews or extreme variations or trends. Previous research shows curvilinear valence effects (Ludwig et al. 2014; He and Bond 2015), such that at low levels of activation, reviews drive sales, but at very high levels of activation, they do not (because review readers assumed the review writers were being irrational). It would be interesting to connect the potential curvilinear effects of incoherence with research on ambivalence, though little is known about extreme ambivalence or when consumers use high positive and negative language simultaneously to describe product and service experiences. Further analysis of the non-linear effects of incoherence (ambivalence) in customer reviews would be insightful.
»IMPLICATIONS
»The sheer volume of unstructured, text-based sentiments has led to intensified efforts to gauge their impact and integrate their insights into marketing (Gopaldas 2014). The latest managerial evidence (Magids, Zorfas, and Leemon 2015) suggests that online consumer sentiments represent an enormous opportunity to create new value, so companies should pursue emotional connections as a key strategy. This article illustrates the importance of speech act features for analyzing sentiment, not just to derive the writer’s sentiment but also to predict its value for subsequent sales. Our Study 2 findings—that weekly sentiment changes in verbatim consumer reviews influence readers’ reactions (i.e., changes in sales ranks)—emphasize the importance of moving from sentiment valence to sentiment strength.
»To improve implications, researchers need to discern sentiment appropriately, rather than relying on simple valence. Sentiment is continuous (rather than either positive or negative) and requires consideration of its granular, explicit and implicit conveyance in writing. Researchers then can achieve better results in terms of decoding writers’ willingness to act and readers’ reactions. As we show in Study 3, the findings can be extrapolated to other contexts in which consumers share product and service experiences, without assiging stars. Our Study 2 highlights that improvements in sentiment classification have important applications for sales forcasting.
»Finally, this study provides better understanding of the linguistic markers of sentiment, spanning both word use and message development. Our research offers a theory-based approach to improve understanding of consumer sentiment. This study delineates and validates general cues at each level; the speech act framework provides further guidelines for including additional, context-specific, and independent cues. At the intersection of linguistic and consumer research, these theory-driven improvements are particularly relevant, considering the growing amount of potential research insights that will stem from online, unstructured content.»
