=Paper=
{{Paper
|id=Vol-2220/03_CONFWS18_paper_24
|storemode=property
|title=Insights for Configuration in Natural Language (short paper)
|pdfUrl=https://ceur-ws.org/Vol-2220/03_CONFWS18_paper_24.pdf
|volume=Vol-2220
|authors=Andrés Felipe Barco,Élise Vareilles,César Iván Osorio
|dblpUrl=https://dblp.org/rec/conf/confws/BarcoVO18
}}
==Insights for Configuration in Natural Language (short paper)==
https://ceur-ws.org/Vol-2220/03_CONFWS18_paper_24.pdf
Insights for Configuration in Natural Language
Andrés F. Barco1 and Élise Vareilles2 and César I. Osorio1
Abstract. Usually, in configuration processes, customers interact 2. answer a predefined series of questions, always in the same order
with a decision support system, also named configurator, by explic- whatever his/her knowledge and needs about the product,
itly selecting components or required functionalities through a writ- 3. express his/her needs and preferences in such a way they fit the
ten series of questions, until the complete configuration is done and predefined set of choices and options proposed online by configu-
the desired product is defined. The interactions during a configura- rators, and
tion process may vary vastly depending on the customers’ knowledge 4. click to select the relevant functions or components meeting the
about the product and his/her understanding of its potential function- best of his/her needs.
alities. However, configurators are not conceived for making a differ-
ence between expert and inexpert customers as interfaces and input All these facts gathered make the configuration of products more
information are all expected to be the same for everyone. This pa- and more a counter-intuitive process. Also, current interaction makes
per discusses how natural language can enhance configuration pro- no difference between expert and non-expert user as most of the input
cess by making possible for customers to express their desires, needs mechanism (such of as graphical windows, drawings, text fields) are
and preferences in natural language, and for configurators to interpret all expected to be the same for everyone [2].
their words and better help them to find relevant solutions. This kind This paper discusses how the mature techniques from AI may be
of configuration process could have as foundations an expert systems used to allow a more natural interaction between customers and con-
that that maps speech into constraints and objectives. We present the figuration systems. In essence, we describe the future of configura-
artificial intelligence trends motivating our research, an initial archi- tion in which natural language interactions could replace the tradi-
tectural design and potential applications of the research. tional one based on writing, explicit selection of items and rigid se-
ries of questions. More precisely, we argue that current AI advances
like natural language processing could help customers to express
1 Introduction their needs implicitly in speech format (e.g. “I need a cheap lap-
For several decades now, customers want to bring a personal touch to top with good video card”) while an expert system could infer the
their products to make them special and unique. To meet this grow- requirements and set the goals of the configuration (e.g. video card
ing demand of personalization, companies nowadays no longer of- ≥ good). We call this kind of configuration Configuration in Natural
fer standard products, but more and more personalizable ones [1]. Language. We also present an idea of architecture of such an expert
Thanks to the Web technologies and dedicated decision support sys- system.
tems, named configurators, this personalization is done directly and The paper is divided as follows. In Section 2, the traditional and
interactively online [1]. Customers can play with the wide range of future customer-system interactions in configuration are discussed.
choices and options offered by companies: they can assemble, cut, In section 3, a software architecture based on current advances in AI
color, choose, ..., and visualize the result of their desires and ulti- is introduced. In Section 4, some of the mathematical frameworks
mately order it, all in few clicks and minutes. from which the configuration in natural language can take advantage
This concept of personalization or configuration of products con- are discussed. Finally, in Section 5, some applications of the research
sists in assembling modules or predefined components, to produce a and conclusions are presented.
unique and specific product [10]. For businesses, this is a way to of-
fer personalized products to stand out from the competition and build 2 Configuration Interactions: Past and Future
customers’ loyalty through more accurately reflecting their tastes and
needs. We present in this section the interaction typically made in configu-
Interactions between potential customers and configurators, be- ration systems and contrast it with the idea of configuration in natural
come now one of the key aspects of configuration problems [22]. language proposed in this paper.
Nevertheless, often the configuration relies in a long data capturing
process. Normally, to configure the product object of desires, any
potential customer has to:
2.1 Traditional Configuration Interactions
In most of the cases, configuration systems follow a series of itera-
1. Face the increasing range of choices and options without being tive steps that guide the customer and help him/her to progressively
completely able to focus on his/her essential items, configure his/her own product. These steps, that we call the standard
1 Universidad de San Buenaventura Cali. Santiago de Cali, Colombia. email: way of configuration, are as follows:
{anfelbar,ciosoriod}@usbcali.edu.co
2 Université de Toulouse, Mines Albi. Albi, France. email: 1. The system shows some sets of components and functionalities to
elise.vareilles@mines-albi.fr the customer in a predefined way,
2. The customer either selects the most relevant component or func- functionalities, and goals; critical to reach an appropriate configura-
tionality which meets his/her needs and desires, or specifies a tion solution. For instance, in the first of the two previous examples,
value for the most important criteria (such as the price he/she is both the processor speed and the computer cost may be seen as con-
willing to pay for the product), figuration goals. For tackling this challenge, we propose an expert
3. The system removes or assigns the set of remaining components system architecture built upon AI trends.
and functionalities according to the set of constraints in the prod-
uct,
4. The system computes the price of each possible product and valu- 3 AI Trends-based Architecture
ates their criteria. The field of AI is generating broad interest. Technological advances
using AI techniques, such as the DeepMind GO system developed
In step 2, the selection is usually done manually by clicking on by GoogleTM [19] and the IBM WatsonTM analytic system [11],
the relevant item via a mouse-click or a touch screen, or by inputting draw the attention of the academic and non-academic world on the
its specific value directly from a keyboard. At the end of the standard innovative role of mathematical models from computer science and
way of configuration, the customer selects his/her unique product and philosophy. Current trends show that the use of AI and other related
orders it. fields are being widely used sectors such as economy, health, trans-
port industry, aviation and games, among others [20].
2.2 Future Configuration Interactions The configuration in natural language is motivated by the recent
advances in AI and by industrial trends, in particular the growing in-
The configuration in natural language changes the interactions be- terest in the construction of machines that understand human emo-
tween customers and configurators. Customers will be able, what- tions and act according to the interaction with human [16]. It is
ever their requirements and knowledge about the product, to express sought that the machines assist decision-making processes whereas
better in a more natural way their preferences, desires and needs and the understanding of human emotions helps to improve the expert
configure faster their own products. The significant difference be- systems behavior and interaction [8]. This idea, known as Human
tween traditional and forthcoming configuration interactions lies in Aware AI, is not new in configuration as it has been used as a goal
the way of expressing and capturing customers’ needs and goals. The in different configuration systems (see for instance [3]). Further, the
steps for the configuration in natural language are as follows: idea of understating or inferring user needs has been widely study
in the plan recognition problem [6]. Nonetheless, to the best of our
1. The system welcomes the customer. knowledge, current advances in natural language use remain to be
2. The customer writes or says what he/she wants or needs by using adopted in configurators implementations.
his/her own words. Within the human aware AI field, the Natural Language Understat-
3. The system infers the set of components and functionality the user ing (NLU) [4] and Natural Language Processing (NLP) [12] draw at-
wants or needs, and the goal of the configuration. tention for its capabilities of human computer interaction. In essence,
4. The system removes or assigns the set of remaining components NLU and NLP systems allow the user to ask questions in everyday
and functionalities according to the set of constraints in the prod- language and try to understand these questions in order to return
uct. appropriated answers. Typically, these systems makes some hypoth-
5. The system computes the price of each possible product and valu- esis according to the question and a knowledge base, such as Inter-
ates their criteria. net, and then process an output. This is akin to the problem of plan
recognition, i.e., knowing the user’s plans and goals [6]. Further, if
To exemplify this kind of configuration, limit us to written inter- these systems are improved with natural language generation (NLG)
actions in natural language via a keyboard or similar device. Three in order to produce responses, the system then becomes a question
examples of such interactions when configuring a computer, and the answering system (QAS) [13]. These systems were conceived to re-
respective responses of the inferring engine, are: ceived provide argued answers to user queries. From these systems,
WolframAlphaTM [7] and IBM WatsonTM [11] present the more
• Customer express: “I want a really fast computer but not too ex- innovative results for configuration as these systems are able to rec-
pensive” ognize some information in form of requirements within informal
System infers: Component(processor, speed, high) speech in text format.
Goal(computer, cost, low) To illustrate these capabilities, Figure 1 shows the result when
• Customer express: “I just want to play video games, preferably querying “I want a computer of less than 1000 dollars.” in the
not heavy so I can carry it easily” WolframAlphaTM system. To construct a response, the system
System infers: Component(video card, processing, high) maps the input into a more elaborated query, encoding the query thus
Goal(computer, weight, low) making a syntactic and semantic analysis. Nevertheless, as the sys-
• Customer express: “I need to write my texts” tem does not focus on inferring mathematical notions, it is easily con-
System infers: Component(keyboard, comfort, high) Goal(none, fused by adding words that add relevant constraints. For instance, no
none, none) result shown when querying the same computer but DellTM man-
ufactured; “I want a dell computer of less than 1000 dollars”. We
As expected, inferring components or functionalities and configu- consider that this is a major drawback in the system when addressing
ration objectives is a major challenge. On the first hand, the universe configuration problems.
of words used by humans to express the same thing may be vast. The IBM WatsonTM system works similarly to the
Second, the way to build expressions may vary largely depending on WolframAlphaTM . It encodes a query by applying automated
academic background, experience, state of mind and mood, to name a reasoning, machine learning and several other techniques to analyze
few. Finally, it is difficult to set a difference between components and the speech. One of the more innovative applications of the IBM
Figure 1. Output example of WolframAlphaTM system.
WatsonTM system is the personality characterization from written
speech. To do this, in addition to question answering, different
techniques of sentiment analysis [21] are applied. This kind of
analysis may be useful to infer expertise level of the user and then
behave accordingly.
In spite of the recent advances of NLU, NLP and QAS, these
are not well-suited for addressing configuration problems given that
they do not focus on constructing the mathematical notions (con-
straints and objectives) needed in most configuration problems. Be-
sides, question answering systems are too powerful in the sense they
are of general purpose having different question types and extensive
knowledge bases. At the other end of the spectrum, specific applica-
tions using question answering systems technology do not necessar-
ily deal with extensive vocabularies, hypothesis and so on, as these
applications are domain-dependent. Ergo, its underlying mechanism
may be simpler although more robust and may count with reduced
knowledge, question types and small knowledge bases. In conse-
quence, we have devise an architecture, presented in Figure 2, for
implementing configurators that exploits the aforementioned natural
language elements. The mandatory module is that of NLU whereas
NLP and NLG are optional (used if formatted answers are desired). Figure 2. Architectural view of expert system for configuration NL.
External services may be attached in order to fulfill specific tasks.
The differentiating element in the architecture is the inference en-
gine. This key element is in charge of discriminating among the set
that sum, subtract, division and multiplication are allowed as well
of words those referring to components of the product, functional-
as inequality symbols. This helps to make the mapping unambigu-
ities and/or configuration objectives. This is in fact a challenge as
ous. Given that our main innovative application is the configuration
different conclusions may be reached from a given sentence.
in natural language, mathematical frameworks used to tackle config-
uration problems comes naturally. Here, we briefly describe three of
4 Mathematical Frameworks for Configuration in these models.
Natural Language
• The first framework, Constraint Programming (CP), has been
Unlike NLP and QAS, configuration processes are mapped, gen- identified as a key paradigm in the expansion of applied computer
erally, into mathematical (optimization) models that unify the cus- science [17]. CP is part and good representative, of the declarative
tomers requirements and problem domain limitations into a single programming frameworks. This is one of the most used framework
framework. To infer constraints and objectives from informal speech, to address configuration problems as it suits their constrained na-
it is needed an underlying mathematical framework in which such ture [15]. First, the knowledge (constraints) that restricts possible
notions are built. In other words, to construct a mathematical formula configuration of elements (variables) is easily modeled under the
it is necessary the set of values and operands allowed in the formula. declarative framework of constraint satisfaction problems. And
As an example, if constructing inequalities, the expert system knows second, constraint-based configurators are able to present different
solutions to users, often optimal, even when they do not provide REFERENCES
all configuration parameters.
[1] L. Ardissono, A. Felfernig, G. Friedrich, A. Goy, D. Jannach,
• Among the same lines we have integer programming (IP). It is a M. Meyer, G. Petrone, R. Schafer, W. Schutz, and M. Zanker, Person-
mathematical optimization in which some or all of the variables alising on-line configuration of products and services, 225–229, Thiel,
are restricted to be integers [18]. Likewise the constraint satisfac- S., Pohl, K., Lyon France, 2002.
tion model, it is built from constraints and limitations over vari- [2] Liliana Ardissono, Anna Goy, Matt Holland, Giovanna Petrone, and
Ralph Schäfer, Customising the Interaction with Configuration Sys-
ables, although variables do not have a given domain, and objec- tems, 283–287, Springer Berlin Heidelberg, 2003.
tive functions such as minimization and maximization. This model [3] Virginia E. Barker, Dennis E. O’Connor, Judith Bachant, and Elliot
or extensions of it (Mixed IP or Linear IP) have been widely used Soloway, ‘Expert systems for configuration at digital: Xcon and be-
to reason and solve configuration problems (see for instance [9]). yond’, Commun. ACM, 32(3), 298–318, (March 1989).
[4] M Bates, ‘Models of natural language understanding’, Proceedings of
• Finally, another framework that can be used in configuration in
the National Academy of Sciences, 92(22), 9977–9982, (1995).
natural language is SAT; determining if a given propositional for- [5] A. Biere, A. Biere, M. Heule, H. van Maaren, and T. Walsh, Handbook
mula is satisfiable by an interpretation [5]. The inclusion or ex- of Satisfiability: Volume 185 Frontiers in Artificial Intelligence and Ap-
clusion of a given feature in a product may be seen as boolean plications, IOS Press, Amsterdam, The Netherlands, The Netherlands,
assignments. Thus, it would be possible to deduce if a solution ex- 2009.
[6] Sandra Carberry, ‘Techniques for plan recognition’, User Modeling and
ists by constructing a boolean formula using the requirements of User-Adapted Interaction, 11(1), 31–48, (Mar 2001).
the customers and the configuration knowledge (e.g. compatibil- [7] John B. Cassel, Wolfram—Alpha: A Computational Knowledge
ity among components). The modeling of a configuration problem “Search” Engine, 267–299, Springer New York, New York, NY, 2016.
under SAT is not new (see for instance [14]). [8] Celso M. de Melo, Stacy Marsella, and Jonathan Gratch, ‘”do as i say,
not as i do”: Challenges in delegating decisions to automated agents’,
in Proceedings of the 2016 International Conference on Autonomous
Agents #38 Multiagent Systems, AAMAS ’16, pp. 949–956, Richland,
5 Concluding Remarks SC, (2016). International Foundation for Autonomous Agents and Mul-
tiagent Systems.
In this paper, we have presented insights on how to enhance the inter- [9] Ingo Feinerer, ‘Efficient large-scale configuration via integer linear pro-
action between customers and configuration systems into something gramming’, Artif. Intell. Eng. Des. Anal. Manuf., 27(1), 37–49, (January
called configuration in natural language. In summary, we have dis- 2013).
cussed the possibility of using AI techniques, namely, human-aware [10] Alexander Felfernig, Lothar Hotz, Claire Bagley, and Juha Tiihonen,
AI to implement an expert system that infers constraints and objec- Knowledge-based Configuration: From Research to Business Cases,
Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 1 edn.,
tives from a speech input. 2014.
We have built out the idea around the concepts of NLU, NLP and [11] D. A. Ferrucci, ‘Introduction to “This is Watson”’, IBM Journal of Re-
QAS. Then, we argue that although current systems have reached search and Development, 56(3.4), 1:1–1:15, (May 2012).
a stable development, there is still research to be done when infer- [12] Chowdhury Gobinda G., ‘Natural language processing’, Annual Review
of Information Science and Technology, 37(1), 51–89.
ring mathematical notions from informal sentences, i.e., everyday [13] L. Hirschman and R. Gaizauskas, ‘Natural language question answer-
language. In addition, we have argued that advanced question an- ing: The view from here’, Nat. Lang. Eng., 7(4), 275–300, (December
swering systems such as IBM WatsonTM are not yet well suited 2001).
for configuration problems as they are not intended to map implicit [14] M. Janota, Do SAT Solvers Make Good Configurators?, 191–195, Thiel,
requirements into mathematical models. S., Pohl, K., Limerick, Ireland, 2008.
[15] Ulrich Junker, Configuration., Chapter 24 of Handbook of Constraint
To give a view of the mathematical models that can prove use- Programming (Foundations of Artificial Intelligence). Elsevier Science
ful to construct the proposed expert system, we have briefly de- Inc., New York, NY, USA, 2006.
scribed three techniques from AI and operational research. These, [16] Stacy Marsella and Jonathan Gratch, ‘Computationally modeling hu-
techniques, namely, constraints programming, integer programming man emotion’, Commun. ACM, 57(12), 56–67, (November 2014).
[17] Francesca Rossi, Peter van Beek, and Toby Walsh, Handbook of Con-
and boolean satisfiability, have been used to solve different configu- straint Programming (Foundations of Artificial Intelligence), Elsevier
ration and optimization problems and are in our point of view inter- Science Inc., New York, NY, USA, 2006.
esting models to construct a configurator in natural language. [18] G. Sierksma, Linear and Integer Programming: Theory and Practice,
The main application of inferring constraints and goals from Second Edition, Advances in Applied Mathematics, Taylor & Francis,
speech is linked to the mathematical modeling of real-world prob- 2001.
[19] David Silver, Aja Huang, Christopher J. Maddison, Arthur Guez, Lau-
lems. Simply stated, the inference system may be used to build rent Sifre, George van den Driessche, Julian Schrittwieser, Ioannis
mathematical models to solve linear problems, discrete optimiza- Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander Dieleman,
tion problems and probabilistic problems, among others. Further, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Tim-
specific properties from each mathematical framework may be ex- othy Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel,
and Demis Hassabis, ‘Mastering the game of go with deep neural net-
ploited, such as the expressivity of declarative approaches like logic works and tree search’, Nature, 529, 484–503, (2016).
and constraint programming. Intuitively, many more applications ex- [20] Peter Stone, Rodney Brooks, Erik Brynjolfsson, Ryan Calo, Oren Et-
ists; those in which user requirements, preferences, limitations or ob- zioni, Greg Hager, Julia Hirschberg, Shivaram Kalyanakrishnan, Ece
jectives are needed. For instance package managers in unix-based Kamar, Sarit Kraus, Kevin Leyton-Brown, David Parkes, William
operating systems. Typically, a package manager from a Unix-based Press, AnnaLee Saxenian, Julie Shah, Milind Tambe, and Astro Teller,
‘Artificial intelligence and life in 2030’, Technical report, Stanford Uni-
system must be asked to search or install a specific package. Using versity, Stanford, CA, (September 2016). One Hundred Year Study on
a expert systems that maps speech into constraints and objectives, a Artificial Intelligence: Report of the 2015-2016 Study Panel, Stanford
manager would accept inputs like “I need a powerful UML diagram University.
editor” to present some potential editors to be installed. Further, ar- [21] Maite Taboada, ‘Sentiment analysis: An overview from linguistics’, An-
nual Review of Linguistics, 2(1), 325–347, (2016).
guments used by such managers can be replaced by everyday words [22] Mitchell M. Tseng and S. Jack Hu, Mass Customization, 836–843,
thus preventing the non-expert user to learn the specifics of the pro- Springer Berlin Heidelberg, Berlin, Heidelberg, 2014.
grams.