In this paper we describe a novel approach to extract contextual information from user reviews, which can be exploited by context-aware recommender systems. The approach makes use of a generic, large-scale context taxonomy that is composed of semantic entities from DBpedia, the core ontology and knowledge base of the Linked Data initiative. The taxonomy is built in a semi-automatic fashion through a software tool which, on the one hand, automatically explores DBpedia by online querying for related entities and, on the other hand, allows for manual adjustments of the taxonomy. The proposed approach performs a mapping between words in the reviews and elements of the taxonomy. In this case, our tool also allows for the manual validation and correction of extracted context annotations. We describe the taxonomy creation process and the developed tool, and further present some preliminary results regarding the effectiveness of our approach.
In addition to the users’ preferences –i.e., tastes and interests–,
Context-Aware Recommender Systems (CARS) exploit
information about the circumstances under which the users (prefer
to) interact with items, such as the time of the day, the day of the
week, the weather conditions, and the users’ location, mood, and
social companion. Studies have shown that contextual conditions
may have an important, positive effect on the usefulness of
recommended items [
Despite these benefits, CARS are not used extensively. This is
mainly due to the lack of available context data associated with
user preferences, and the difficulty and cost to obtain it.
The simplest method to acquire context data in a recommender
system consists of asking the user to explicitly state the contextual
conditions as she interacts with the system items [
ComplexRec 2017, Como, Italy. 2017. Copyright for the individual papers remains with the authors. Copying permitted for private and academic purposes. This volume is published and copyrighted by its editors. Published on CEUR-WS, Volume 1892.
An alternative technique is to identify and extract contextual
information from freely given user generated contents, such as
product reviews in e-commerce sites, opinion articles in web
blogs, and personal posts in online social networks. Among these
types of user generated contents, text reviews have been the most
investigated for recommendation purposes [
By means of natural language processing techniques and resources, the proposed approach performs a mapping between words in the reviews and the categories and instances in the taxonomy (Section 3). In this case, our tool also allows for the manual validation and correction of extracted context annotations. We present some preliminary results regarding the effectiveness of our approach on a well-known dataset of Amazon reviews for products in three domains, namely books, movies and music (Section 4).
To identify context information in user reviews through the
proposed approach, we first need a definition of context
dimensions (a.k.a. context categories or context variables) and
their respective values. As noted in [
Different to existing approaches, where small and domaindependent context are used, in this paper we advocate for the use of a generic (i.e., domain-independent), large-scale and adjustable context taxonomy. By counting with such taxonomy, it would be possible to extract and exploit context information in different domains, not being limited to recommendation purposes.
Since building from scratch, adapting and keeping updated this
taxonomy manually would be highly costly, we propose to make
use of collaboratively created, consensual and up-to-date
repositories available on the Web. In particular, we propose to use
Semantic Web-based repositories published in Linked Data [
We surveyed previous work regarding context modeling with the aim of identifying the context dimensions considered in the literature. For each of the identified dimensions, we searched for representative DBpedia categories as root categories of the taxonomy. Next, our approach iteratively performs online queries to DBpedia for acquiring subcategories that are included in the taxonomy. To assist the process, we developed a software tool that allows browsing and modifying the taxonomy, as well as establishing the criteria that determine which (sub)categories represent context.
In the literature, there are several context modeling proposals, particularly in the area of pervasive and ubiquitous computing. Revising published work, we found that a major approach for context modeling is the use of taxonomies/ontologies. Table 1 summarizes some important context modeling approaches, briefly describing the context categories they consider. Based on the analyzed models, we decided to include four major context dimensions, namely Location, Time, Environmental, and Social contexts, which will have a variety of context categories. Table 2 shows the DBpedia categories selected as the root taxonomy categories for each of the above contexts.
http://dbpedia.org/resource/Category: A two-level context ontology model, with a generic level and a domain-specific level Basic context descriptors: User, Resource, Location, Service, Activity, Device, Network Instantiating the W4 model components: Who, What, Where, When As contextual components, Where is associated to the location of the fact (What) performed by the subject (Who), and Where refers to the time or time range associated to the fact.
Mapping between the 5W1H model components –Who, What,
[
Status, Location, Time, Goal It is important to note that DBpedia allowed us to retrieve not only related, more specific categories by means of the skos:broader property, but also instances/individuals of the categories by means of the dct:subject property. Thus, the gathered vocabulary in the context taxonomy goes beyond the category names. As illustrative examples, for the case of Location context and its root category dbc:Wheeled_vehicles, our approach identified the subcategories dbc:Automobiles and dbc:Buses, among others. For the subcategory dbc:Automobiles, it also acquired instances such as dbr:Automobile3, dbr:Car, and dbr:Motorcar,
as well as particular vehicles, e.g. dbr:Ferrari_F40 (belonging to subcategories like dbc:1990s_automobiles and dbc:Coupes). Our tool provides a Context Taxonomy Editor, shown in Figure 1 (left), which browses the taxonomy, and allows the user for expanding and removing any of its categories (and corresponding subcategories and instances) by online querying DBpedia. It also allows for establishing criteria that the category names have to be satisfied to be explored or discarded by our approach. Specifically, it lets defining syntactic patterns “starts with”, “contains” and “ends with” to explore/discard the categories whose names respectively start with, contain, and end with certain text; for instance, expanding those categories whose names end with the suffix _transport, like dbc:Road_transport and dbc:Rail_transport (which are subcategories of dbc:Transport_by_mode, the root category of the Location context). Moreover, the tool allows for setting the maximum depth of the taxonomy.
After navigating through the built taxonomy, we observed that
there exist words describing context that were not represented as
either categories or instances of the taxonomy. We realized that,
in many cases, such words were synonyms of already included
categories/instances. For this reason, we decided to enrich the
taxonomy with such synonyms. Specifically, we obtained them
from the well-known WordNet English lexical database [
Table 3. Statistics of the built taxonomy 9591 2560 16729 http://dbpedia.org/resource/ modified by a first-person possessive determiner, my or our, and the sentences that have a noun phrase composed by a preposition for, to or with followed by a first-person personal pronoun, me or us. On each selected sentence, we first extracted candidate (simple or compound) words that may represent contextual conditions. The words used by users for describing context are nouns –e.g., evening, Saturday, restaurant, wife–, and some adjectives –e.g., sunny, cold, happy, nervous. Again, we used CoreNLP, and specifically its Part-Of-Speech tagger, to obtain the above words.
For every candidate word, our approach establishes whether or not
the word matches certain category of the context taxonomy. In
positive cases, as explained below, it generates scores that reflect
the relatedness between the words and the matched categories. The
resultant (word, category, score) tuples will be the context
annotations of the input reviews. The process is done as follows.
For a fast identification of matches between review words and
taxonomy categories, we took advantage of Apache Lucene4, a
well-known information retrieval software library. Hence, before
performing any word-category matching, we used Lucene to create
a search index for the context taxonomy. Specifically, for each
taxonomy category, we created a text document containing the
name of the category, the names of the instances in the category,
and the synonyms of the category. We then added such document
into the index. In this way, given a candidate word as input, a search
in the Lucene index returns a ranked list of documents containing
the word, i.e., a ranked list of categories related to the word.
In order to deal with morphological derivations –e.g. singular and
plural forms of a word–, and user misspellings or deliberated word
modifications –e.g. repeating the last vowel of a word as emphasis–,
the words to be indexed were first lemmatized. For this reason, to
effectively map a word with its corresponding context category, we
further analyzed the document words retrieved by Lucene and to be
compared with the queried candidate words. Specifically, for such
pairs of words, (!, !), we computed the Damerau-Levenshtein
distance [
To preliminary assess the correctness of the context annotations
generated by our approach, and their potential utility for
recommendation, we conducted two experiments, namely a manual
validation of annotations, and an offline evaluation of a
state-of-theart context-aware recommendation model fed with the annotations.
In both experiments, we used part of the Amazon reviews dataset5
published in [
http://lucene.apache.org
http://snap.stanford.edu/data/web-Amazon.html DBpedia categories
574 DBpedia instances Category WordNet synonyms 846 Instance WordNet synonyms Total categories 1420 Redirected DBpedia instances 4578
Total instances
The built taxonomy is used by our approach to identify and annotate in reviews those words that express user contextual conditions, in particular, those words that correspond to (the names of) categories or instances in the taxonomy. For such purpose, we follow two steps: first, selecting a limited number of words that could represent context and second, if possible, mapping such words to taxonomy categories.
In order to avoid wrong context annotations, we did not analyze all
the sentences in the reviews, but only those that express personal
statements and opinions of the review author, e.g., “I watched the
movie with my children at home,” where children and home would
be annotated as social and location contexts, respectively.
To select such sentences, we made use of the Stanford CoreNLP
toolkit [
Before evaluating the utility of the generated context annotations in recommendation, we assessed a subset of them manually. For this, we used a Context Annotation Validator implemented in our software tool. Shown in Figure 1 (right), the tool allows the user to load and view a review together with its metadata (id, title, summary). In a left panel with the review text, the words annotated as context are highlighted. A table on a right panel details such annotations, showing the annotated word, the associated context category, the type of match (category, instance, synonym) and the similarity score for each annotation. In the table, the user is allowed to validate whether an annotated word is context or not and, in positive cases, state whether the assigned context is correct, wrong or can be improved with a child or parent category. Moreover, the tool also allows adding manual annotations, with the possibility of browsing the taxonomy and searching for categories and instances on an interactive panel. From the set of annotated reviews, we carefully read and manually assessed annotations in book, movie and music reviews. Table 4 shows the number of evaluated reviews and annotations, and the percentage of such annotations that were correct/context. We observe that our approach obtained a relative high percentage of correct mapping cases (84.2% on average), and that most of the wrong cases were due to semantic ambiguities, which we will address in the future. The predominant contexts were social contexts in book reviews (e.g., daughter), location contexts in movie reviews (e.g., theater), and emotional contexts (e.g., melancholy) in music reviews.
We also notice that the number of annotations that really referred to context was low (36.7% on average). We believe that other domains, such as restaurant and hotel reviews, may have much more contextual information, and thus should be investigated. As a lesson learnt from the conducted validation, we saw the need of investigating additional syntactic patterns with which selecting the sentences to be analyzed and annotated; for instance, we found out contextualized sentences with a third-person subject, such as the reader/viewers…, and anyone interested in/looking for…, and contextualized sentences with two-person phrases like if you...
In order to get initial insights about the potential benefits of
exploiting our context annotations in CARS, we evaluated their
effect on collaborative filtering by means of the Item Splitting (IS)
context pre-filtering algorithm [
From the 100,000 reviews in each domain, we selected reviews having annotations from the social, location and time context, as they showed best rate of correct context annotations. The final books, movies and music dataset used were respectively composed of 17,543, 15,983 and 14,194 reviews, 16,701, 14,340 and 12,332 users, 307, 306 and 2,540 products. Table 5 shows the achieved MAE and RMSE values for the recommendation methods in each domain using 5-fold cross-validation. Light gray indicate lower error when applying IS over the corresponding baseline. Dark gray indicate the lowest error in the corresponding column. We observe that the extracted context data enable improving recommendation quality in the three domains, particularly in the case of CAMF, and also in the case of IS over MF, preliminary showing the potential of our approach.
For more argued conclusions, we should extend the experiments with other context-aware recommendation methods, and alternative evaluation metrics (e.g., ranking-based) and protocols. We also have to analyze results for each context dimension (location, time, environmental, and social) separately to conclude which of them is more/less useful in each domain.
In this paper we have presented first a method and a software tool to semi-automatically build a context taxonomy with Linked Data. The taxonomy could be exploited by context-aware applications distinct to recommendation. Moreover, the proposed method could be used to build other types of taxonomies. For instance, it may be instantiated to describe features/aspects of items, and thus it may be exploited in feature-based opinion mining and recommendation applications.
Using the built taxonomy, we have developed a method to extract and annotate context in user reviews. We manually assessed some of the annotations, and preliminary evaluated the utility of the annotations in recommendation. We tested them with Item Splitting and Context-Aware Matrix Factorization algorithms. We are interested in evaluating additional context-aware recommendation techniques as well, and we also want to perform an exhaustive experimentation and result analysis comparing alternative approaches to index and match the context taxonomy categories in the annotation process.
This work was supported by the Spanish Ministry of Economy, Industry and Competitiveness (TIN2016-80630-P), and by Dirección de Investigación, Universidad del Bío-Bío (DIUBB151115 4/R and GI 170315/EF).