1Automated Product Profiling through NLPJiho Han, Dowon (Ronny) KoAbstract—We design and experiment with an innovativeway to automatically generate product profiles fromAmazon reviews. Using NLP, we extract opinions fromeach review, clusters them by their orientation throughan unsupervised learning (k-means). From these clusteredopinions, we estimate the product profiling kernel θ andthe pricing kernel λ. Finally, we optimize/update the wordpolarity by minimizing the prediction error (supervisedlearning). While the trained model perform only slightlybetter than random guessing, the interim outputs and theestimated parameters seems to provide useful informationwhile showing a possibility for improvement.I. INTRODUCTIONIn this paper, we propose an unconventional approachto summarize Amazon online product reviews. Incontrast to previous researches which attempt to simplypredict the overall product rating, our model aims to:(1) extract and evaluate the sentence-level opinions, (2)aggregate them into the relevant clusters (i.e., evaluationcriteria), and finally, (3) predict the product rating byaveraging the scores from the evaluation criteria.Undoubtedly, this approach has much complicatedinternal structures, which make parameter estimationprocess highly difficult. Nonetheless, there are twoimmediate benefits that can be gained from thisapproach. First, this model generates a product profile,or a scorecard for each evaluation criteria. Second,we can estimate the degree of an average consumer’sconsideration on each evaluation criteria for each marketproduct.This approach, as far as we know, has not beensuggested in the literature. Therefore, our effort lackstheoretical support and the fruitfulness of the researchis never warranted. Even then, we believe this approachsuggests a novel framework for unsupervised – orminimally-supervised – information extraction and isworth the effort to explore.This paper was originally written in joint effort with Hyungoo Kang.Although he decided to pursue a different project, I recognize anddeeply appreciate his contribution for this research.The rest of the paper is organized as follows. Section 2goes over the previous literature and provides the formalmodel specification. Section 3 describes the data whichwe work with. In Section 4 we discuss the estimationprocedures in detail and point out developmental issues.Section 5 shows the model’s outcome and the perfor-mance result. Lastly, Section 6 concludes and suggeststhe directions for future research.II. LITERATURE AND MODEL OVERVIEWA. Literature ReviewThere are several researches conducted on the topicof document-level classification and review-level ratingprediction. Simply because documents have a lotmore words than sentences, most of document-levelclassification models can utilize an N-gram modelcombined with NER to extract keywords while beingagnostic about the syntax structure of each sentence.[1], [3] On the other hand, when evaluating sentences,we do not have such luxury of using a large number ofwords, and thus do not have to deal with semantic-levelof natural language processing!This is where our model can contribute to theliterature. Given a minimal piece of information(i.e., product ratings), we devise an algorithm wheresentence-level information extraction is dependent onsemantic relations as little as possible. In other words,by jointly estimating the polarity and the orientationof opinions (these terms will be defined soon in thenext section) and assuming a certain market structure(what we call the pricing kernel and the orientationkernel, also explained below), we can hopefully bypasssemantic issues to some degree.In our analysis, the previous work that is most closelyrelated to ours is that of Popescu and Etzioni. [2] Infact, their objective and subtask procedures are almostidentical to our model. However, the main differencelies in the use of product ratings and, in general, theassumption of market structures. For example, Popescuand Etzioni extensively use NLP techniques – e.g.,2defining 10 rules to represent semantic relations – toextract relevant product features and the correspondingsentiments while discarding the information from theproduct ratings.In the next section, we present our model specificationand show how we can leverage the market structureassumption.B. Model OverviewThe fundamental building block for our model isthe opinions we extract from each sentence. We defineopinion as a tuple consisting of orientation and polarity,i.e.,opinion = (orientation, polarity)While orientation is an abstract entity, we representit by a m-dimensional vector xifor the ith-opinion. Onthe other hand, polarity of an opinion is simply a scalar.Given a set of opinions, we model the predicted ratingˆr as follows:ˆr =X∀i(λTθxi)pi(II.1)where λ ∈ Rk(pricing kernel)θ ∈ Rk×m(profiling kernel)pi= polarity of opinion ik = the number of evaluation criteriaProfiling kernel θ maps each opinion xiinto k-dimensional space where each dimension representsa criterion with which consumers use to evaluate theproduct (e.g. price or design). Pricing kernel λ is thenthe weights - or the importance - that each evaluationcriterion has on determining the overall product rating.(e.g., How important is the battery life when evaluatinga cell phone?)The overall procedures for training the model aresummarized in Algorithm 1. Note that the key insightof our model is that we update the polarity and theunderlying market structure so that the prediction erroris minimized.Of course, the question is: How exactly can weestimate each entity mentioned above? In Section IV,we will answer that question in detail.Algorithm 1 Overall training processGiven the orientation extraction rule ROandthe polarity extraction rule RP,O ← extractOpinions(R)P0← extractP olarityW ord(R0; RP)O ← extractOrientations(R, P0; RO)Until converge {CO ← clusterOpinions(O)Ki← extractM arketStructure(CO, Pi)Pi+1← optimizeP olarity(CO, K0; RP, RO)Ki+1← updateMarketStructure(CO, Pi)}Return {K = (λ, θ), P ; RO, RP}III. DATAWe used the product reviews from Amazon, which isstored in XML format and is available to public on theweb.1Ideally, we would like to obtain a dataset wherethe ratings for specific evaluation criteria are givenalong the review texts. Unfortunately, we were unableto collect such dataset, and hence the sentence-levelpolarity evaluation can only be performed throughhuman inspection. In other word,