This paper describes a novel approach for summarising usergenerated reviews for the purpose of explaining recommendations. We demonstrate our approach using TripAdvisor reviews.
Product reviews, that are written by real users, are now mainstream online. Sites
like Amazon and TripAdvisor have collected thousands of reviews for all manner
of products, and users are increasingly relying on these reviews make better
choices [
In this paper we explore how to exploit textual reviews to summarise product experiences that can explain recommendations. In particular we describe how we can pro le a user based on the reviews that they have written, and identify product features that matter to them. And we explain how we can represent products in a similar fashion, by extracting opinions and sentiment information from their reviews. We then describe an initial approach to generating personalised product summaries for given user-product pairs on TripAdvisor data. This section describes our approach for generating personalised summaries that are tailored to a user`s preferences, and non-personalised summaries that re ect the general opinion of users about a product. 2.1
Inspired by the methods described in [
To evaluate the sentiment of a given feature Fi in sentence Sj of review Rk,
we identify the closest sentiment word wmin to Fi in Sj ; Fi is labelled as neutral
if no sentiment words are present. In this context, sentiment words are those
contained in the sentiment lexicon [
Our corpus is taken from the hotel review site TripAdvisor.com. It contains 226,110 unique reviews written by 150,352 unique users for 2,500 hotels in 6 di erent cities. We mined over 270,000 unique features mention around 4,129,265 times using the process described in Section 2.1.
We propose two levels of features to represent users and hotels; these levels are determined by their similarity to amenities prede ned on the TripAdvisor website. We obtain a set of base features, which are single-nouns and bi-gram noun phrases from the set of amenities that TripAdvisor uses to described hotels (e.g tness centre and wheelchair access ). We expect these features to be highly meaningful and familiar to users. We hypothesise that there is an extended bag of words that people use when talking about the same thing. For instance, a person talking about `breakfast' may use words like `orange juice' or `bu et'. The key thing to note is that these are related words, but not necessarily synonyms. We apply k-Means to base features and their corresponding sentences to nd other co-occurring, related features. The k most co-occurring features are used to enhance the representation of each base feature as expanded features. 2.3
Users normally talk about the things that matter to them in reviews. Therefore we assume that the preferences of each user consist of the features they mention in reviews and the relative frequency at which they mention them, which may indicate their relevance to the user. Hence we de ne the pro le of a user as the set of all features mentioned by the user in reviews. Each feature in the set is tagged with the relative frequency with which it was mentioned in the user's reviews.
Similarly we de ne a hotel pro le as a set of features mentioned by users about the hotel. Each feature in the set is tagged with its relative frequency and average sentiment score (a value in the range [ 1; +1]).
Generating Personalised and Opinionated Review Summaries
To construct non-personalised summaries for a user-hotel pair, two hotel pro les are built using the base and expanded features respectively. Each feature is assigned a ranking score that is the product of its average sentiment and its normalised frequency. When the non-neutral features in the hotel pro le are ranked by the ranking score, the top-n and bottom-n features form the pros and cons parts of the explanation respectively.
To generate a personalised summary for a user-hotel pair, two pro les are built each for users and hotels using the base and expanded features respectively. Each feature in the hotel pro le is assigned a ranking score that is the product of its average sentiment and its normalised frequency. The ranking score of each feature in the hotel pro le is updated by multiplying its original ranking score with its normalised frequency in the user pro le. When the non-neutral features in the hotel pro le are ranked by the updated ranking score, the top-n and bottom-n features form the pros and cons parts of the explanation respectively. In both explanation types, expanded features are mapped to their corresponding base features that are familiar to the user. 2.5
Earlier in 2.4, we described how we can model a user and hotel pro le using base and expanded features. In Fig. 1 we show a fragment of the user and hotel pro le used to generate the example summary in Fig. 2.
Fig. 2 shows a screenshot of a personalised summary generated for a userhotel pair. The pro features are highlighted in green, and the cons features in red. We always present base features in the summaries regardless of how we choose to model the user and hotel pro le. This is to avoid having too ne a level of granularity which might be unintuitive to users. Therefore `shuttle bus service' is a pro feature of the hotel, ranked by the user's preferences. The tooltips (see (c) in Fig. 2) display expanded features in snippets of sentences from reviews that are associated with the base feature in the summary. Here the reviewers have discussed the `printer' not working, and the closing hours of the `centre' ; these have been summarised to `business centre'
This paper presents a method for constructing personalised summaries of items based on opinions from textual reviews. With TripAdvisor data we show how the pros and cons of hotels can be explained to users using di erent feature representations. Our technique focuses on those features that users write about most frequently in their reviews. This forms the basis for prioritising features that are likely to be of interest to the user compared to non-personalised explanations, which focus on features that are commonplace for a hotel but may not be so relevant for an individual user. This work builds on related work in the area of opinion mining and recommender systems but considers a novel application in the form of explanation generation.