=Paper=
{{Paper
|id=Vol-1329/paper2
|storemode=property
|title=Sentiment Analysis for Dynamic User Preference Inference in Spoken Dialogue Systems
|pdfUrl=https://ceur-ws.org/Vol-1329/paper_4.pdf
|volume=Vol-1329
|dblpUrl=https://dblp.org/rec/conf/esws/VanrompayCGAL14
}}
==Sentiment Analysis for Dynamic User Preference Inference in Spoken Dialogue Systems==
https://ceur-ws.org/Vol-1329/paper_4.pdf
Sentiment Analysis for Dynamic User Preference
Inference in Spoken Dialogue Systems
Yves Vanrompay1 , Mario Cataldi2 , Marine Le Glouanec1 , Marie-Aude
Aufaure1 , and Myriam Lamolle2
1
MAS Laboratory, Ecole Centrale Paris
Grande Voie des Vignes,
F-92 295 Chatenay-Malabry Cedex, France
2
Université Paris 8, Saint-Denis, France {yves.vanrompay,marine.le-glouanec,
marie-aude.aufaure}@ecp.fr,{m.cataldi,m.lamolle}@iut.univ-paris8.fr
Abstract. Many current spoken dialogue systems for search are domain-
specific and do not take into account the preferences of the user and his
opinion about the proposed items. In order to provide a more personal-
ized answer, tailored to the user needs, in this paper we propose a spoken
dialogue system where user interests are expressed as scores in modular
ontologies and his sentiment about the system propositions is consid-
ered. This approach allows for a dynamic and evolving representation of
user interests. In fact, in order to improve the performance of the detec-
tion mechanism of users preferences, we propose a hybrid model which
also makes use of a sentiment analysis module to detect the opinion of
the user with respect to the proposition of the system. This allows the
system to leverage the degree of user satisfaction and improve the over-
all recommendation mechanism being more precise about the expressed
user interest. An evaluation on a representative set of dialogues is pre-
sented and highlights both the validity and the reliability of the proposed
preference inference mechanism.
1 Introduction
The traditional goal of spoken dialogue system is to approach human perfor-
mance in conversational interaction, specifically in terms of the interactional
skills needed to do so. With this objective, as an attempt to enhance human-
computer interaction, in the PARLANCE project3 , we build a system for in-
teractive, personalized, hyper-local search for open domains such as restaurant
search and tourist information. Current search engines work well only if the user
has a single search goal and does not have multiple trade-o↵s to explore. For
example, standard search works well if you want to know the phone number of a
specific business but poorly if you are looking for a restaurant with several dif-
ferent search criteria of varying importance, e.g. food type versus location versus
price etc. The latter requires the user to collaborate conversationally over several
turns. In order to provide a personalized answer, tailored to the specific user, in
3
https://sites.google.com/site/parlanceprojectofficial/
18
this work we focused on three levels: (1) user modelling in terms of preferences
and interests inferred from past interactions with the system, (2) personalization
of search approach and (3) the opinion of the user himself about the suggested
products.
Personalization in the context of spoken dialogue system has slowly pro-
gressed compared to the field of non-natural language systems. Indeed, current
spoken dialogue systems are mostly domain-specific, using rather static informa-
tion from experts and knowledge bases. In PARLANCE, we choose to represent
the dynamic domain knowledge through modular ontologies, where each ontol-
ogy module represents a domain and can be dynamically loaded at run-time to
meet the current needs of the user. In order to provide a semantic representa-
tion of the user model, the concepts and attributes in these ontology modules
are annotated with scores representing the preferences and interests of the user.
This allows us to learn the specificities of a user, and give responses that fit
the user’s profile. Additionally, user opinions can be detected where system rec-
ommendations are made. The sentiment analysis of the user answers permits
to detect the degree of user satisfaction to systems replies and provide more
tailored recommendations in the future.
This paper is organized as follows: Section 2 presents related work on prefer-
ence systems. Section 3 motivates our research and provides an overview of the
di↵erent components in the system. Section 4 introduces the representation of
user interests while Section 5 describes how to detect and exploit user opinions
about the recommended items. Then, we provide our experimental evaluation.
Finally, section 7 concludes the paper.
2 Related Work
The automatic detection and categorization of sentiment expressed in natural
language constitutes a remarkable research challenge, and has generated a nexus
of techniques and approaches. Recently, [1] and [2] have conducted extensive
surveys of the open challenges in this research area. Most of the works in this
area focus on the categorization of overall sentiment, capturing a form of average
polarity at the document level [3, 4].
A number of researchers have applied natural language processing (NLP)
techniques to detect features in small chunks of text [5, 6]. Some methods make
use of a lexical approach to focus on subjective words of the considered dia-
logue, namely adjectives and adverbs[7]. The study of linguistic complements,
negations, syntactic dependencies, etc., can also be positively used to categorise
the terms in the dialogue [8, 9].
Various API are available for sentiment analysis purposes. Most of them
are built for study of specific environments, like social networks and/or blogs,
and they try to focus on particular subjects, enterprises or events. Among all,
SentiWordnet [10] analyses relationships like antonymy, hyponymy and/or hy-
peronymy to give a triplet value (objective, positive and negative) to each term
of the considered phrase/document.
19
3 PARLANCE Architecture
This section starts with an example dialogue between the user and the system
motivating the need for personalized interaction. Let us suppose that Paul is
visiting San Francisco and he is looking for a good restaurant in the center of
the city. An example dialogue between Paul (U) and the system (S) can be as
follows:
1. U: I’m looking for a moderately priced restaurant near the centre of town.
2. S: I have found 2 Indian and 1 Chinese restaurant. Star India is a moderately
priced restaurant in the centre of town. It has free Wifi.
3. U: Ok, what’s the address?
4. S: The address is 3721 Geary Ave.
5. U: Oh, I don’t want that one, it’s hard to park on that street. Are there any
other Indian restaurants?
In this conversation, not only information on user interests learned from past
dialogues is exploited, but also new information on interests and preferences is
learned. In item 2, the system has learned from previous conversations that the
user is quite fond of Indian and Chinese restaurants. Therefore, it starts with
proposing restaurants with these food types. The system also knows that the
availability of Wifi is important to the user, so it proactively gives this informa-
tion. In item 5, as the user asks for an alternative restaurant, user preferences
are updated: the user does not like Geary Avenue.
Considering this scenario, the starting point for inferring user interests are
abstract representations of the history of dialogues (dialogue act units) between
the user and the system. A user model manager analyses the dialogue history
of the user and derives interest scores associated to concepts, attribute types
and attribute values in a weighted ontology module corresponding to a specific
domain.
4 Evolving User Preferences as Weighted Modular
Ontologies
User preferences regarding concepts and attributes are inferred from the user’s
dialogue history. All user and system utterances from past dialogues are saved
in a so called dialogue act unit (DAU), which is the abstract representation of
utterances in PARLANCE. For example, when the user asks for the price range
of a restaurant, this is represented as the DAU request(price). Interest scores are
derived from logged traces of DAUs. We keep track of the positive and negative
occurrences of attribute values and concrete instances. These frequencies allow
us to rank the di↵erent elements. Based on this ranking it is decided which
system response is best suited with regards to the user interests. If the user often
queries for pricing information with given attribute value ”cheap” in searching
for a restaurant, the value of the price attribute will have a high frequency, and
the system will proactively inform the user about the price in its answers, and
lead the system to recommend restaurants from a cheaper price range.
20
Our mechanism for expressing user interests is integrated in the approach
which represents information as (hierarchical) modular ontologies. Ontology
modularization is defined as a way to structure ontologies, so that large domain
ontologies will be the aggregation of self-contained, independent and reusable
knowledge components (considered as Ontology module (OM)). An OM can be
seen as an ontology fragment that has a meaning from the viewpoint of applica-
tions or users. Each ontology module corresponds to a particular domain and its
size should be small for easy maintenance. Each OM is characterized by a basic
concept, called the pivotal concept. A tourism ontology can contain several on-
tology modules like lodging, transportation and restaurant information. Taking
the restaurant ontology module, this contains (amongst others) the restaurant
concept with attribute types name, food type, dress code and location. To per-
sonalize the responses given to the user, our user model incorporates the interests
and preferences of the user by assigning scores to elements in the appropriate
ontology modules. These scores are updated according to what is being learned
from the history of past dialogues, ensuring that the interests evolve as user
preferences may change through time. The weights are useful in two di↵erent
aspects. First, the scores are used to rank and recommend concrete instances
that are of interest to users. Second, attribute value scores are used to generate
a system response tailored to the user needs. For example, based on the interest
scores, the system can decide to proactively inform the user on the food type of
the restaurant, but not on the dress code.
4.1 Evolving update of the User Model
Calculation of scores happens o✏ine based on all available dialogues in the di-
alogue history component. This means that the user model contains both the
representation of user preferences (as weights in modular ontologies), as the
mechanism to calculate the scores. To update scores based on recent dialogues
of the user with the system, the user model manager aims to the recalculation
of scores on a regular basis. The scores are thus dynamic, within the modular
ontology structure which itself is relatively static. The scores of the attribute
values are relative and sum up to 1. If an attribute type has m possible val-
1
ues, the initial score wi for each value will be m . The score wi is updated by
counting how often attribute value attvali was selected and dividing this number
by the total number of times a value for the corresponding attribute type was
specified by the user. This means however that the past is as important as the
present. In our context, user interests will typically evolve over time. So, if the
dialogue history includes for instance the dialogues of the user during the last
six months, it is reasonable to have more recent dialogues having relatively more
influence on the user interest model than older ones. To this end, the scores for
attribute values should be updated in such a way that recent dialogues have
more relevance than older ones, which we do using an exponential smoothing
method as follows: wi = ↵ ⇥ xj + (1 ↵)wi0 where wi0 represents the old score
and xj 2 {0, 1} is the value for the choice taken at moment j in the dialogue
history. If xj = 1 then attvali was specified by the user, if xj = 0 it was not.
21
Using this method, the sum of all scores of the attribute values belonging to
an attribute type remains 1, and the scores represent the relative importance of
each attribute value. The learning rate ↵ 2 [0, 1] is a real number that controls
how important recent observations are compared to older ones.
5 Mining User’s Opinions wrt System Recommendations
This component is responsible for analyzing the positive, neutral, or negative
opinions produced from user with respect to the propositions of the system. In
order to tackle this issue, we make use of a novel feature-based polarity analysis
technique[11], which combines statistical techniques with natural language pro-
cessing. As in literature, we define a polarity as a real number that quantifies
the user’s positive, neutral, or negative opinion about a feature.
With this goal, for each dialog (intended as the complete set of system-user
interaction), we model the user’s sentiment with respect to the proposition of
the user by estimating the degree of positivity/negativity with respect to the
considered features. To extract such fine-grained sentiment information from raw
text, we model each review as a set of sentences. A sentence is then formalized
as a syntactic dependency graph, used to analyze the semantic and syntactic
dependencies between its terms, and identify the terms referring to features.
More formally, a sentence S can be formalize as an ordered vector of terms
S = {w0 , w1 ...wm }, where the order represents the original position of each term
within the sentence. The sentence s can be represented as a dependency graph
G. The dependency graph is a labeled directed graph G = (V, E, l), where V is
the set of nodes representing the lexical elements wi and E the set of edges (i.e.
dependency relations) among the nodes.
The graph is obtained through a preliminary POS tagging phase, achieved by
training a tagging model on the annotated corpus proposed by [12] and therefore
by calculating the probability p(tj |wi ) of assigning a tag tj to the term wi using
a maximum-likelihood estimation as in [13].
Subsequently, the dependency graphs are then utilized to detect the terms
referring to a feature, which expresses some non-neutral opinion, including com-
pound expressions, e.g. “the restaurant serves a very good pizza.” In this phase, a
SentiWordNet-like approach[10], which attributes polarity values to each Word-
Net synset, is used as a source of polarity values. In detail, using the synset graph
proposed by WordNet, we calculate the polarities of each term by using a two-
step algorithm. A first step is a semi-supervised learning step in which polarity
values are assigned to two sets of seed nodes. This set consists of two subsets;
one subset of “paradigmatically positive” synsets and another one consisting
of “paradigmatically negative” synsets [14]. The polarities are then propagated
automatically to other synsets of the WordNet graph by traversing selected se-
mantic relations. This propagation is performed within the minimal radius that
guarantees no conflicts among the relations, that is, until a node labeled as pos-
itive points to a node already linked to some negative seed, or vice-versa. In
other words, we only propagate the polarities to the nodes that are univocally
connected to a positive or a negative seed. Second, a random-walk step is exe-
cuted on the whole WordNet graph starting from the seed nodes, and iteratively
22
propagates the positive and negative polarity to all of the synsets. This approach
preserves or inverts the polarity of each node based on the number of positive
and negative relations that connect it to the seeds. The process ends when a
convergence condition is reached. This condition is satisfied when all the nodes
have maintained the same polarity sign (positive or negative) after two consecu-
tive steps. Finally, the polarities of terms are aggregated into single values, each
one referring to a specific feature.
6 Evaluation
The evaluation of our approach consists in two di↵erent analysis; from one side,
capturing the evolution of the user preference scores in time, i.e. with respect to
the size of the dialogue history or the number of interactions of the user with
the system, and from the other side, study the user feedback expressed in the
dialog with the system.
6.1 Analysing the evolution of the user preference scores
By using the fading factor ↵ in our interest score update formula, we make sure
that more recent dialogues provide more information on interests than the older
ones, leading to an evolution of interests. To obtain real spoken dialogues, we
used Amazon Mechanical Turk, which is a tool for crowd-sourcing where users
can call a toll-free number, solve tasks assigned to them, and earn money. First,
a number of well defined tasks, expressed in natural language, were constructed.
As an example, we ask a user to find an Italian restaurant in the center of
town. As a check for task success, the user has to give in the phone number of
the restaurant he has found. This means the test users have every incentive to
succeed, since they are only paid in case of task success. We varied the content
of the tasks as to reflect changing user interests over time. The basic task was
in each case to find a restaurant with variations in the attribute types food, area
and price. The experiments are based on 60 dialogues, and we set ↵ to be 0.1.
In analyzing the evolution of the interest scores, we plot the maximum score
P across attribute values for each attribute type as a function of the number
of dialogues considered, as can be seen in Figure 1. Indeed, the score P serves
as a basic metric for showing how outspoken the user interests are at a certain
moment in time. A low P signifies that the user changes his interest rather quick,
while a high P means that in recent dialogues he has shown a rather consistent
and stable choice behavior.
Figure 1 shows the evolution of the values of P over the dialogues for di↵erent
attribute types. It can be noticed that there is a correlation between the ”peaks”
in the graph for food and area, reflecting that when the user is in a stable period
with respect to his interests, this holds for some attribute types. The evolution
of the graph allows us to identify periods in which the user very dynamically
changes his interests, and periods in which his choices remain merely stable.
Also, by looking at the relative values of P for the di↵erent attribute types at a
given moment, it is possible to make a ranking of those types where the user does
not change his behavior a lot, compared to the ones that are more fluctuating.
23
1
0,9
M 0,8
a 0,7
x 0,6
0,5 Area
P
0,4 Food
r
0,3 Price
e
f 0,2
0,1
0
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55
Iterations
Fig. 1. Evolution of user interests in function of dialogue history
6.2 Dissecting user opinion with respect to System
recommendations
In order to assess the proposed approach to sentiment analysis, we analysed the
detected sentiments in the considered corpus of dialogs. The aim of this analysis
is to study, from one side, how the users interact with the system and, on the
second side, analyse the retrieved opinion with respect to the recommended
items. For this, we considered three features: food type, area and price. Figure
2 shows the obtained polarities.
Fig. 2. Statistics about the performed feature-based polarity on user dialogs.
These results highlight, for all the considered features, that the user express
more likely positive opinions rather than negative ones. In a sense, the users
accept, in great majority of the cases (up to 97% of the cases regarding the
food type) the recommendations of the system, explicitly proving the goodness
of the whole recommender system and the user preference model. In fact, the
previously collected preferences lead the system to recommend an item that
perfectly match the real user preferences.
Notice that this feedback value can be positively use to dynamically tune the
learning rate ↵ (explained in Section 4.1). In fact, a negative opinion can suggest
a change of the user preference wrt the attended one. Thus, we can dynamically
set the learning rate, which express how important recent observations can be
compared to older ones, proportionally to the negative sentiment expressed by
the user in order to reflect a sort of user preference change. Following this idea,
the more negative the sentiment expressed by the user, the higher the learning
rate which will reflect the necessity of the system to quickly update the user
preferences.
24
7 Conclusion
In this paper we described an approach for modeling the dynamics of user pref-
erences in a spoken dialogue system for searching items of interest. A fading
method allows for the interests to keep track of the evolution of user behavior.
We then leverage the opinion of the user about the suggested items to improve
the recommender system and tune the learning rate of the system. We evaluated
our approach on a set of real dialogues and showed it can provide useful insights
into changing interests. The ontology-based representation of interests lets us
tailor recommendation of items to the recent preferences the user has exhibited,
for each search domain involved.
References
1. Pang, B., Lee, L.: Opinion mining and sentiment analysis. Foundations and Trends
in Information Retrieval 2(1-2) (2008) 1–135
2. Liu, B.: Sentiment analysis and opinion mining. Synthesis Lectures on Human
Language Technologies 5(1) (2012) 1–167
3. Beineke, P., Hastie, T., Manning, C., Vaithyanathan, S.: Exploring Sentiment
Summarization. In: Proceedings of the AAAI Spring Symposium on Exploring
Attitude and A↵ect in Text: Theories and Applications. (2004) 1–4
4. Hiroshi, K., Tetsuya, N., Hideo, W.: Deeper Sentiment Analysis Using Machine
Translation Technology. In: Proceedings of COLING ’04. (2004) 1–7
5. Holz, F., Teresniak, S.: Towards Automatic Detection and Tracking of Topic
Change. In: Computational Linguistics and Intelligent Text Processing. Volume
6008. (2010) 327–339
6. Missen, M., Boughanem, M., Cabanac, G.: Opinion Mining: Reviewed from Word
to Document Level. (2012) 1–19
7. Turney, P.D.: Thumbs up or thumbs down?: Semantic orientation applied to un-
supervised classification of reviews. In: Proceedings of the ACL ’02. ACL ’02,
Stroudsburg, PA, USA, Association for Computational Linguistics (2002) 417–424
8. Benamara, F., Chardon, B., Mathieu, Y.Y., Popescu, V.: Towards context-based
subjectivity analysis. In: IJCNLP. (2011) 1180–1188
9. Wilson, T., Wiebe, J., Hwa, R.: Just how mad are you? finding strong and weak
opinion clauses. In: Proceedings of the AAAI’04, AAAI Press (2004) 761–767
10. Baccianella, S., Esuli, A., Sebastiani, F.: SentiWordNet 3.0: An Enhanced Lexical
Resource for Sentiment Analysis and Opinion Mining. In: Proceedings of LREC’10.
(2010) 2200–2204
11. Cataldi, M., Ballatore, A., Tiddi, I., Aufaure, M.A.: Good location, terrible food:
detecting feature sentiment in user-generated reviews. Social Netw. Analys. Mining
3(4) (2013) 1149–1163
12. Marcus, M.P., Marcinkiewicz, M.A., Santorini, B.: Building a Large Annotated
Corpus of English: The Penn Treebank. Computational Linguistics 19(2) (1993)
313–330
13. Klein, D., Manning, C.D.: Accurate unlexicalized parsing. In: Proceedings of
the 41st Annual Meeting on Association for Computational Linguistics. ACL ’03
(2003) 423–430
14. Turney, P.D., Littman, M.L.: Measuring Praise and Criticism: Inference of Seman-
tic Orientation from Association. ACM Trans. Inf. Syst. 21(4) (October 2003)
315–346