“Love the shape, but hate the weight”: Aspect-Based Sentiment Analysis to Identify Product Innovation

User-generated content, such as product reviews, offer one of the most accessible and abundant
information sources to firms. Coupled with recent technological advances in natural language processing
these review postings provide opportunities for marketing managers to gather market intelligence in the
form of consumer opinions and competitive information. One of the most prevalent analysis methods to
determine consumer sentiment on a given topic is sentiment analysis (Karlgren et al. 2012), which uses a
lexicon-based approach to classify different terms within the review text based on their polarity (i.e. as either
positive, negative, or neutral) or by using a machine learning-based approach in which a model is used to
assess the sentiment of a text, resulting, in both cases, a contextual polarity score for the review text.
However, this approach provides very little insight into the underlying factors influencing the review
sentiment, thus limiting its applicability to future marketing and product innovation decisions. Recently,
researchers have started investigating deeper forms of sentiment analysis aimed at increasing interpretability
by including additional factors to qualify the syntax of the reviews (e.g. readability) or by conducting more
fine-grained sentiment analysis on sub-sentences or aspects within the review text (Thet, Na, and Khoo
2010). Consider for example the following review text: “Phone battery and fingerprint reader are great, the
problem I have with this phone is how slow it is. The phone's touch screen is horrible.”. In this exemplary
review, some product aspects are positively evaluated by the consumer while other aspects are negatively
evaluated. Yet, traditional sentiment analysis would score the overall sentiment for this text as neutral, thus
providing very little insight on how consumers evaluate specific product features.
The current work provides a novel methodological approach to conduct aspect-based sentiment analysis in
order to identify the sentiment linked to different product features mentioned in a text. We leverage
traditional text analysis along with prior knowledge to identify the most representative product features, and
introduce a reusable python script to easily conduct this analysis. To illustrate our approach, we scraped
1,147 consumer reviews on the Samsung Galaxy A10s from a major online review site (amazon.com).
Leveraging LDA topic modelling coupled with previous knowledge about smartphones, we identified ten
relevant aspects of the phone (e.g. software, shape, price, battery life). Next, we extract all nouns from the
review text using part-of-speech tagging and determine which nouns in the review texts relate to each of the
ten aspects. We do this by converting words and aspects into vector representations and then calculating the
cosine similarity between them. To assign a sentiment score to each noun we separate text sentences and
compute the sentiment for each sentence in which one of those relevant nouns is found, taking adjectives as
sentiment indicators and adverbs as multipliers. Finally, we aggregate the sentiment for each noun belonging
to a specific aspect, thus allowing us to calculate a sentiment score for each product aspect. As shown in
Figure 1, such an aspect-based sentiment analysis provides a much deeper insight into consumers’ likes and
dislikes of a specific product.
Unlike traditional sentiment analysis, the proposed, multi-method approach provides far reaching
practical applications for marketing managers and product developers, opening up new opportunities to
leverage unstructured, user-generated text to gain insight into feature-level product performance and identify
consumer unmet needs.