The field of recommender systems has seen substantial progress in recent years in terms of algorithmic sophistication and quality of recommendations as measured by standard accuracy metrics. Yet, the systems mainly act as black boxes for the user and are limited in their capability to explain why certain items are recommended. This is particularly true when using abstract models which do not easily lend themselves for providing explanations. In many cases, however, recommendation methods are employed in scenarios where users not only rate items, but also provide feedback in the form of tags or written product reviews. Such user-generated content can serve as a useful source for deriving explanatory information that may increase the user's understanding of the underlying criteria and mechanisms that led to the results. In this paper, we describe a set of developments we undertook to couple such textual content with common recommender techniques. These developments have moved from integrating tags into collaborative filtering to employing topics and sentiments expressed in reviews to increase transparency and to give users more control over the recommendation process. Furthermore, we describe our current research goals and a first concept concerning extraction of more complex argumentative explanations from textual reviews and presenting them to users.
Today’s Recommender Systems (RS) have been shown to
generate recommendations for items that match the user’s interest
profile quite accurately. The underlying methods are usually
based either on purchase data or ratings of other users
(collaborative filtering), or on structured data explicitly describing
the items (content-based filtering). Yet, algorithmic maturity
does not necessarily lead to a commensurate level of user
satisfaction [
ExSS ’18, March 11, Tokyo, Japan
When recommendation methods are applied in the real world,
users can often provide textual feedback on items in the form
of short tags or written reviews. Textual feedback from other
customers is known to strongly influence the current user’s
decision-making [
One popular way of increasing transparency of RS is to use
textual explanations [
Building on our prior work, we in the following present a novel concept that relies on extracting and summarizing arguments about products from textual reviews in order to provide users with adapted model-based recommendations, and in particular, explanations that are personalized according to the current user’s preferences and styles of decision-making.
While tagging helps to classify items by attributing specific
properties, the descriptive nature of tags limits their
applicability as an explanatory element. User reviews, on the other hand,
constitute semantically rich information sources that
incorporate sentiment and often an intended sequence of arguments,
i.e. an argumentation flow. Reviews go beyond mere
objective descriptions by (sometimes implicitly) invoking a stance
towards a target. Users reading a review can get an idea of
motives and reasoning behind its author’s words. This verbal
provision of subjective information constitutes a
comprehensible context on which arguments and their interrelations can
be grounded. However, automatically identifying presence,
polarity and quality of argumentation structures has not yet
been considered in RS research, although it is well known
that textual feedback of other users may strongly influence
decision-making [
Item
The hotel is in a central location, with lots of shopping and eateries nearby, yet only few meters from the beach.
Attention-based mechanism central location shopping beach
Linguistic argument mining and stance detection Recommender system
Explained item recommendation We propose a conceptual framework that automates the process of extracting arguments from reviews (see Figure 1) to come up with personalized argumentative item-level explanations for recommendations: We suggest to apply feature-based methods from computational linguistics in order to derive argumentative structures, including argument polarity in the form of stances. Beyond that, detecting personally important arguments requires a deep integration into the process of calculating recommendations. We aim at utilizing deep learning methods that—matching the attention analogy—enable the system to focus on important concepts subject to a user variable. Put together, this leads to the following challenges: Linguistically analyzing review texts via argument mining and stance detection.
Identifying important concepts for a target user via an attention-based mechanism.
Deriving an argumentation flow via multiple applications of the attention-based mechanism.
Unifying the linguistic analyses and the attention-based mechanism.
Linguistic Analyses: Computational linguistic approaches can be distinguished based on granularity of the analysis. Document-level analyses, for instance, aim at determining a sentiment or stance towards the subject of decision, e.g. a hotel. However, for our purpose such an approach is too shallow as a review generally not only consists of utterances regarding the target item as a whole, but usually also includes remarks on sub-aspects. Sentiments or stances towards these sub-aspects may deviate from the review’s overall polarity. For example, a guest may generally like a hotel albeit he or she found the bed uncomfortable. In addition, a review might address sub-aspects not important to the current user. Therefore, we propose a more fine-grained approach relying on aspect-level argument mining that is capable of assigning polarity to a variety of mentioned entities.
Automated argument mining refers to identifying linguistic
structures consisting of at least one explicit claim and optional
supporting structures such as premises [
Another difficulty arises when one wants to decide whether an
extracted argument is in favor or against a particular product.
However, a notion of polarity is necessary to be able to select
adequate arguments supporting or opposing a
recommendation. Therefore, we propose to relate each argument to a stance
expressed by the review author. In this regard, a stance target
is not limited to the subject of decision itself but may
essentially address anything towards which one can have a stance.
It follows that the linguistic analyses need to identify the target
and establish a relation to the subject of decision. Please note
that stance detection involves identifying a subjective
disposition that might often be implicit [
Attention-based deep learning approaches [
Personalizing explanations requires attention to be distributed
with respect to user preferences, i.e. users act as the context
that moderates the (soft-)attention’s output. Thus, it is
necessary to calculate attention weights subject to a vectorial
user representation. We propose to model users analogous
to our previous work [
Argumentative Flow: Argumentation can be considered
more convincing if it consists not only of an aggregation of
potentially important words or phrases. We assume arguments
to become more understandable and effective in case they
follow a coherent argumentative flow, i.e. dependencies between
successively extracted arguments. The analogy in our
proposed concept is the repeated application of the (soft-)attention
mechanism while considering previous output as context.
Informally, this mimics a conversational exchange between user
and attention mechanism where the latter continuously tries to
identify concepts more and more tailored specifically towards
the user’s preferences. The result would be a succession of
relevant word sets that exhibit sequential properties and are
thus tied to each other. As a consequence, the output sequence
would reflect the whole path of how the system came up with
a particular recommendation and thus explain which concepts
played an important role during the process (see Figure 1). It
must be noted that such a procedure would be closely related
to so-called memory networks [
Unification of Linguistic Analyses and Attention: Up to this point, we have covered two independent approaches to process review texts: (1) linguistic analyses aiming at mining arguments and detecting stances, and (2) attention-based extraction of words relevant to the current user. However, both on their own are limited in expressiveness: the linguistic analyses lack the personalization component while the attention mechanism operates on word-level, thus incapable of extracting complete arguments. Consequently, following our superordinate research goal of presenting users more informative, personalized explanations, we plan to exploit the benefits of both approaches by coupling them closely together. A first attempt, for instance, would be to check whether the surrounding context of a word that received a large amount of attention was identified as an argument by the linguistic processor. If this condition is met, the system can interpret the argument as a possible candidate for a personalized explanation. Assume the system detects e.g. the arguments “the food is good” and “beds are uncomfortable”, then identifying food as an important individual concept should lead to the first argument being chosen for explaining the recommendation.
In this paper, we have discussed the state of research regarding usage of product reviews in RS—with a focus on explanations. We set our prior work into context, where we already used tags to increase a recommender’s transparency and analyzed product reviews in order to provide extended interaction possibilities. Based on this, we pointed out a possible way to exploit this rich information source for presenting users with intelligible, personalized explanations by extracting more complex arguments. We outlined several challenges and proposed a concept to address them. For future work, our research goals are to implement this novel concept, use it to additionally personalize recommendations, and to evaluate it in several domains with a focus on the influence on user experience in comparison to other explanation methods.