Process-oriented intents: a cornerstone

for superimposition of natural language conversations

over composite services
Sara Bouguelia, Auday Berro, Boualem Benatallah, Marcos Báez, Hayet
Brabra, Shayan Zamanirad, Hamamache Kheddouci

To cite this version:

Sara Bouguelia, Auday Berro, Boualem Benatallah, Marcos Báez, Hayet Brabra, et al.. Process-
oriented intents: a cornerstone for superimposition of natural language conversations over composite
services. The 20th International Conference on Service-Oriented Computing (ICSOC), Nov 2022,
Seville, Spain. pp.575-583, �10.1007/978-3-031-20984-0_41�. �hal-03990939�

HAL Id: hal-03990939

https://hal.science/hal-03990939
Submitted on 15 Feb 2023

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 Process-oriented Intents: A Cornerstone for
Superimposition of Natural Language
Conversations over Composite Services

Sara Bouguelia1 0000-0003-4538-0927, Auday Berro1 0000-0003-2411-5761,

Boualem Benatallah2 0000-0002-8805-1130, Marcos Báez3
0000-0003-1666-2474, Hayet Brabra4 0000-0001-7484-2268, Shayan Zamanirad5
0000-0002-3371-9631, and Hamamache Kheddouci1 0000-0002-5561-6203
1
LIRIS – University of Claude Bernard Lyon 1, Villeurbanne, France
{sara.bouguelia, auday.berro, hamamache.kheddouci}@univ-lyon1.fr
2
Dublin City University, Ireland [email protected]
3
Bielefeld University of Applied Sciences, Germany [email protected]
4
SAMOVAR – Telecom SudParis, Institut Polytechnique de Paris, France
[email protected]
5
University of New South Wales (UNSW), Sydney Australia
[email protected]

Abstract. Task-oriented conversational assistants are in very high de-

mand these days. They employ third-party APIs to serve end-users via
natural language interactions and improve their productivity. Recently,
the augmentation of process-enabled automation with conversational as-
sistants emerged as a promising technology to make process automation
closer to users. This paper focuses on the superimposition of task-oriented
assistants over composite services. We propose a Human-bot-Process in-
teraction acts that are relevant to represent natural language conversa-
tions between the user and multi-step processes. In doing so, we enable
human users to perform tasks by naturally interacting with processes.

Keywords: Task-oriented Conversational Bots · REST APIs · Software-

enabled Services · Composite Services · Process-oriented intents

1 Introduction
Task-oriented conversational services (or simply chatbots) emerged as engines
for transforming online service-enabled digital assistance and powering natural
interactions between humans, services, and things [15]. Recently, organizations
leveraged chatbots in a variety of assistance tasks. For instance, the augmen-
tation of process-enabled automation with task-oriented chatbots emerged as
a promising technology to make process automation even closer to users [1,5].
This evolution promises to increase the benefits of automation by simplifying
access and reuse of concomitant capabilities across potentially large number of
evolving and heterogeneous data sources, applications and things [5,9]. While
today’s chatbots may automate some tasks, bot developers have recently started
2 Authors Suppressed Due to Excessive Length

investigating the incorporation of robotic process automation (RPA) to increase

automation [14]. For instance, Devy chatbot was proposed to provide automated
support in DevOps processes [6]. Authors in [13] developed a chatbot for agile
software development teams which analyzes teams’ project data to provide in-
sights into their performance. In [12], the authors proposed an approach that
automatically builds a chatbot from a process model to query process structure.
Another work proposed a chatbot to query event data allowing users to get
insights into specific process executions [11]. All these works are either about
domain-specific chatbots or about querying the process execution or structure
but do not focus on performing process tasks. Other works propose approaches
to interact with business processes and perform process tasks through chatbots.
For instance, Google proposes the use of a chatbot in so-called communication-
enabled business process applications [8]. However, no specific details about the
internals of the chatbot infrastructure are provided. The closest work to ours
is [10], which proposed a methodology that takes a business process model as
input and generates a chatbot to help the users interact with the process. How-
ever, the work does not focus on the recognition of process-oriented intents. In
our previous work, we proposed various techniques for the superimposition of
task-oriented chatbots on top of APIs [17,16,4,3].
In this paper, we focus on the superimposition of task-oriented chatbots over
composite services. In doing so, we enable users to perform tasks by naturally
interacting with service orchestrations involving multiple actions. Orchestrating
human–machine conversations over composite services requires rich abstractions
and knowledge to: (i) interact with a multi-step processes using natural language
utterances, (ii) automatically recognise nuanced, context sensitive and possibly
ambiguous process-aware user intents including starting a new task, inquiring
about task progress, switching from one task to another and exceptional behavior
such as canceling. Specifically, we identify fine-grained Human-bot-Process (HP)
interaction acts that are relevant to represent natural language conversations
between user and multi-step processes. In a nutshell, interaction acts are dialogue
acts that characterise process-oriented intents in user utterances.

2 Preliminaries and Architecture

In this section, we first introduce some process-related concepts and assumptions.

Second, we present a scenario illustrating interactions between a user and multi-
step processes. Finally, we present the architecture.
Preliminaries. A business process is a collection of coordinated tasks to achieve
a concrete goal [7]. The schema of a process can be represented in a variety of
forms, such as Petri nets and Event-Driven Process Chains [7]. For simplicity,
we represent the process schema as a directed acyclic graph. Figure 1 shows an
example of a Travel Booking Process graph. The process graph nodes represent
activities and decision points. A process is associated with a set of exception han-
dling policies. Exception handling policies are directives that model exceptional
situations together with a set of actions that are used to handle exceptions (e.g.,
 Title Suppressed Due to Excessive Length 3

UpdateFlight
CancelFlight
CancelHotel
UpdateHotel

CancelTravel
UpdateTravel

[Method]
[Method]
[Method]
[Method]

[Method]
[Method]

is-realized-by is-realized-by is-realized-by is-realized-by

is-realized-by is-realized-by

Cancel Travel Update Travel

Update Flight Cancel Flight Cancel Hotel Update Hotel
Booking
Booking

Booking
Booking
Booking
Booking

[Process]
[Process]

[Process]
[Process]
[Process]
[Process]

has-cancellation-policy
has-cancellation-policy has-cancellation-policy has-update-policy
has-update-policy
has-update-policy

Travel Booking
Flight Booking
Hotel Booking

[Process]
[Activity]
[Activity]

has-process-schema
is-realized-by
YES YES

BookTravel
Do you want to Do you want to Send Booking
[Method]
book a flight?
NO book a hotel room?
NO details

[Decision Point]
[Decision Point]
[Activity]

Travel Booking Process graph

Fig. 1. Example of a Travel Booking Process Model

cancel a travel booking) [2]. In this paper, we consider that a process is realized
by a composite service. Furthermore, a composite service is accessible through
an API that includes: a main method to invoke the normal process behavior
(e.g., BookTravel method), and exception handling methods to handle excep-
tional behaviors of a process (e.g., CancelTravel method) or an activity (e.g.,
CancelHotel method). We also consider that a process has a set of correlation
attributes that uniquely identify an instance [7].
Scenario. Figure 2 shows an example of a user-chatbot conversation in which
the user is interacting with the Travel Booking process illustrated in Figure 1.
There are interactions that are triggered by the user and others by the chatbot.
During normal process execution (e.g., booking travel): (i) the chatbot can ask
for more information to fulfill a task (message 10) or provide information about
a performed task (message 6), (ii) the user can provide information (message 11)
or inquire about task progress (message 23). As mentioned before, a process is
associated with possible run-time exceptions. In message 27, the user wants to
change the check-out date. This interaction (changing date) is triggered by the
user and involves the update of a previously performed task. However, exception
interactions can also be triggered by the chatbot. For example, assume that the
airline company canceled the user’s flight. The chatbot can trigger an interaction
that involves notifying the user about the cancellation and proposing alternatives
such as changing the travel date. In this paper, we focus on interactions from
the user side; chatbot-initiated interactions are outside the paper scope.
Architecture. To support natural language conversations with processes, the
chatbot needs a set of services to initiate, monitor, and control task-related
conversations. The Natural Language Understanding (NLU) service aims to ex-
tract HP interaction acts and slot-value pairs from the utterance. The Process
Embedding Service (PES) aims to identify the process that corresponds to the
utterance. The Dialogue Manager (DM) service aims to infer the dialogue state
in terms of user intent and its slot-value pairs. This DM relies on the Context
Knowledge Service (CKS) to recognize user intent and infer missing information.
Once the DM recognizes the intent and collects all required information, it per-
4 Authors Suppressed Due to Excessive Length

[START NEW PROCESS INSTANCE] 1.   USER: I would like to plan a one-way trip from Paris to Montreal

[ASK FOR INFORMATION] 2.   BOT: What is the departure date?

[PROVIDE INFORMATION] 3.   USER: 10 July      

[ASK FOR MORE INFORMATION]  4.   BOT: Do you want to book a flight?

[PROVIDE INFORMATION] 5.   USER: yes

[PROVIDE INFORMATION] 6.   BOT: Here is the list of flights I found for the given date: [...].

[ASK FOR MORE INFORMATION] Which one do you prefer?

[PROVIDE INFORMATION] 7.   USER: book the first flight on the list, the one with air Canada
1-14

[PROVIDE INFORMATION] 8.   BOT: Done! I have booked flight tickets.

Book Travel
[ASK FOR MORE INFORMATION] Do you want to book a hotel room?

[PROVIDE INFORMATION] 9.   USER: Yes, please

[ASK FOR MORE INFORMATION] 10. BOT: What is the check-out date from the hotel?

[PROVIDE INFORMATION] 11. USER: 5 august

...

[PROVIDE INFORMATION] 14. BOT: Done! I have booked the hotel room. I sent you details about your travel booking by email.

[START NEW PROCESS INSTANCE] 15. USER: Can you help me to plan another one-way trip from Montreal to New York on 3 August. 

...
15-22

[PROVIDE INFORMATION] 22. BOT: I sent you details about your travel booking by email.
Book Travel
[FOLLOW-UP ON PROCESS STATUS]  23. USER: What is the status of my travel?

[ASK FOR INFORMATION]  24. BOT: What are the departure city and date of this travel?
23-26

[PROVIDE INFORMATION]  25. USER: Paris, 10 July

Inquire about

[PROVIDE INFORMATION]  26. BOT: The status of your travel from Paris to Montreal on 10 July is completed. You have booked
 travel status
 a one-way flight on 10 July and a hotel from 10 July to 5 August.
27-28

[UPDATE AN INSTANCE] 27. USER: Please update my hotel booking for this travel. Change the check-out date to 3 August.
Update a hotel

[PROVIDE INFORMATION] 28. BOT: Done! I updated your hotel booking in Montreal from 10 July to 3 August.
booking

Fig. 2. Example of a user-chatbot conversation. Interaction acts in blue are triggered

by the user and those in green are triggered by the chatbot.

forms the corresponding action and sends the results to the Natural Language
Generator (NLG). The NLG generates then human-like responses to the user.
In what follows, we describe PES and CKS.
The PES aims to identify processes from natural language utterances. It has
(i) a process knowledge model and (ii) a set of services to leverage this knowledge.
The process knowledge model is denoted as a process knowledge graph (P-KG)
with specific types of nodes and relationships. In particular, nodes can describe
Processes, Paths, and Activities. Part of the information in the P-KG is the graph
representation of what we find in the process model definition. Such information
typically includes the process name, the process description, activity name, and
activity description. Furthermore, a Process node has relationships such as is-
realized-by that denotes that a process is performed by an API method (refer to
Figure 1). The PES features the following three services:
– Vector generation Service is used to construct vector embeddings for process
elements. It takes as input a process element e and generates its vector
embedding. It generates: (i) an activity vector by aggregating the information
from activity name and its description; (ii) a path vector by aggregating
vectors of activities in this path; and (iii) a process vector by averaging the
information from process name, process description and all path vectors.
– Process Identification Service aims to identify the corresponding process of a
given utterance u. First, it generates the embedding vector of the utterance.
Second, it calculates the cosine similarity between this utterance vector and
the vector of each process in the P-KG. Then, according to a predefined
threshold, processes with similarities greater than this threshold are kept and
ordered. Finally, the service returns the process scoring the highest similarity.
– Method Identification Service aims to identify a process API method that
corresponds to a given process intent. It takes as input a process p (e.g.,
Travel Booking), a process intent i (e.g., canceling task) and an activity a
(e.g., Hotel Booking) and returns the corresponding API method.
In [3], we proposed the CKS that enables to capture contextual knowledge
from different sources. We extend the CKS with two additional services:
 Title Suppressed Due to Excessive Length 5

– Process Instance Identification Service returns the list of instances for a given
process. It takes as input a process p and returns the list of instances of p.
– Correlation attribute Value Retrieval Service provides values of correlation
attributes for a given process instance id. It takes as input an instance id
and returns an array of attribute-value pairs.

3 Process-aware User Intents

Conversations regarding a given process-aware task may involve several turns
(e.g., starting a travel booking, later inquiring about booking status, modify-
ing travel dates, or canceling the booking). We propose HP interaction acts
to characterize a set of elementary user intents in conversations between users
and multi-step processes. Specifically, we derive four types of process-oriented
intents: start new process instance intent, follow-up on process status intent,
canceling task intent, and task update intent.
We propose five general steps to recognize these process-oriented intents from
an utterance u: (i) detect the HP interaction act class expressed in u; (ii) in-
voke the Process Identification PES service to get the corresponding process p;
(iii) invoke the Process Instance Identification CKS service to retrieve the set
of instances set_i of p; (iv) extract the values of correlation attributes of the
instance that the user is referring to; (v) compare the extracted values of corre-
lation attributes with those of set_i to check if the identified instance already
exists or not. In what follows, we describe the process-oriented intents and define
rules that combine detection of HP interaction acts with additional context and
process knowledge to recognize and realize each of these intents.
Start New Process Instance. This intent allows to identify whether the user
utterance expresses a task that requires the creation of a new process instance.
In general, when users ask for a new task, they provide general information de-
scribing this task, and sometimes they provide more detailed information about
this task. The chatbot needs this information to identify the process and to
check if the utterance concerns the creation of a new process instance. Figure 3
shows the specification of the rule related to start-new-process-instance intent.
This rule consists of trigger and action clauses. The trigger clause defines three
boolean conditions. (1) The condition IS_START() checks if the utterance u ex-
presses a start new instance HP interaction act. (2) The condition IS_SIM()
checks if the process p corresponds to the utterance u. (3) The last condition
EXIST_INSTANCE() compares values of the correlation attributes with those of
the existing instances to check if the identified instance does not exist. If the
conditions are satisfied, the chatbot (1) invokes the Method Identification PES
service to get the method m and (2) triggers m to start the process execution.
Follow-up on Process Status. This intent allows inquiring about process
instance status (e.g., pending, in-progress or completed). For example, in Figure
2 utterance 23, the user is inquiring about travel status. Figure 3 shows the
specification of the rule related to this intent. The conditions are the same as
those defined in the previous rule, except that this rule needs to detect a follow-
up HP interaction act and the process instance should exists. If all conditions
6 Authors Suppressed Due to Excessive Length

(1) (1)

Trigger (2) (2)

(3) (3)

(1) (1)
Action
(2) (2)

(1) (1)

(2) (2)

(3) (3)

(1) (1)
(2) (2)

(3) (3)
(4) (4)

Fig. 3. Rules to recognize and realize the identified process-oriented intents.

are satisfied, the chatbot (1) retrieves the corresponding instance and (2) lists
the status of this instance.
Task Update. This intent allows to identify whether the user wants to update
an existing process instance (e.g., update the hotel check-out date). A user can
request to update an information in the whole process, or a specific activity
in the process. We model an activity as an input parameter of the intent task
update, thus the chatbot can extract from the user utterance the activity that
the user wants to update. The conditions are the same as those defined in the
first rule, except that this rule needs to detect a task update HP interaction
act and the process instance should exists. If all conditions are satisfied, the
chatbot (1) retrieves the corresponding instance, (2) extract the value of the
activity parameter, (3) invokes the Method Identification PES service to get the
the corresponding method m and (4) triggers m to update the corresponding task.
Canceling a Task. This intent allows to identify whether the user utterance
expresses a task cancellation of an existing process instance. The user can request
to cancel the whole process (e.g., canceling travel bookings), or a specific task
in the process (e.g., canceling hotel booking). The steps to recognize and realize
canceling task intent are the same as those in the task update intent (Figure 3).

4 Implementation and Experiments

The first objective of the study was to explore the effectiveness of the proposed
approach, i.e., its capability of recognizing correctly the process-oriented intents
presented in Section 3 and reducing unnecessary interactions. The second objec-
tive was to assess the impact of enabling interaction with a process as opposed
to leaving users to orchestrate services themselves to fulfill their goals.
Experimental design.6 Participants were recruited via email from the ex-
tended network of contacts of the authors. The call for volunteers resulted in a
total of 17 participants. The evaluation scenario required participants to perform
tasks associated with three underlying processes (Travel Booking, Shopping and
6
Experimental materials: https://tinyurl.com/ICSOC22StudyMaterials
 Title Suppressed Due to Excessive Length 7

Table 1. Performance of experimental conditions for each task according to the rele-
vant metrics. Values in bold denote best performance.
Baseline-bot Process-bot
M1 M2 M4 M1 M2 M3 M4
Task TURNS PROMPTS INTENT TURNS PROMPTS PROCESS INTENT

T1 (new) 54,14% 45,59% 51,75% 95,73% 91,67% 92,86% 88,68%

T2 (update) 31,58% 26,67% 25,00% 85,71% 80,00% 91,67% 83,33%
T3 (follow up) 47,15% 51,28% 27,86% 82,61% 88,89% 85,71% 85,71%
T4 (cancel) 27,27% 18,75% 33,33% 80,00% 75,00% 91,67% 83,33%
Mean 40,04% 35,57% 34,49% 86,01% 83,89% 90,48% 85,27%

Scheduling an appointment). Participants were asked to complete 4 tasks in this

scenario (T1: starting new process instances, T2: updating information of process
instances, T3: following up on process statuses, T4: canceling process instances).
We followed a within-subjects design tasking participants to complete the above
tasks by interacting with two chatbots representing the following conditions:
(i) Baseline-bot that implements a standard conversational management; (ii)
Process-bot that support PES and CKS services as well as the defined rules.
Procedure. The study was conducted online. After reading the informed con-
sent and agreeing to participate, participants were introduced to the evaluation
scenario and tasks (T1-T4). They were asked to perform those tasks with the
two chatbots, in a counter-balanced design. After interacting with each chatbot,
participants were asked: to select their preferred chatbot, to specify why, and
provide quantitative feedback on their experience along three dimensions: nat-
uralness (ability to fulfill user tasks in natural language), repetitiveness (ability
to avoid redundant questions) and understanding (ability to interpret requests).
Data analysis. We performed an analysis of conversation logs so as to as-
sess the effectiveness of our approach in recognizing the process-oriented in-
tents. These are computed in relation to optimal conversation scenarios (i.e.,
scenarios assuming ideal accuracy of process-oriented intent recognition) that
we designed based on participants conversations. The effectiveness is calculated
by considering the following metrics: number of (M1) conversation turns, (M2)
prompts asking for missing information, (M3) process correctly identified and
(M4) process-oriented intents correctly recognized.
Results. Table 1 shows the relative performance by task of both baseline-bot
and process-bot in relation to the optimal reference scenario. For the four tasks,
we can see that process-bot experienced a boost in performance M1 and M2,
approaching the efficiency of the reference scenario in terms of number of turns
(M1) and prompts (M2). This level of performance is possible thanks to the PES
and CKS services and the defined rules that allow to perform a mean relative per-
formance across tasks for process identification (M3) and intent recognition (M4)
of 90,48% and 85,27% respectively. In contrast, not supporting these rules leads
the baseline-bot to perform poorly in comparison, with the best performance be-
ing at around 36,70% for the considered metrics. Regarding the user experience,
all but two participants (15/17 participants) expressed a preference towards
the process-bot as opposed to the baseline-bot. The feedback to the specific
user experience questions, highlighted the reasons behind the preference. The
majority of participants reported that process-bot interactions described natu-
8 Authors Suppressed Due to Excessive Length

ralness (11/17), less repetitiveness (11/17) and understanding (12/17), whereas

the baseline-bot was poorly rated on these fronts (2/17).

5 Conclusions and Future Work

We proposed process-oriented intents that are relevant to represent natural lan-
guage conversations between the user and multi-step processes. We devised an
approach that combines recognition of these intents from user utterances with
additional context and process knowledge to enable users to perform tasks by
naturally interacting with service orchestrations. Future work includes identify-
ing a new pattern that allows selecting a service based on subjective attributes.

References
1. Barukh, M.C., et al.: Cognitive augmentation in processes. In: Next-Gen Digital
Services. A Retrospective and Roadmap for Service Computing of the Future, pp.
123–137. Springer (2021)
2. Benatallah, B., et al.: The self-serv environment for web services composition.
IEEE internet computing 7(1), 40–48 (2003)
3. Bouguelia, S., et al.: Context knowledge-aware recognition of composite intents in
task-oriented human-bot conversations. Proc. CAiSE 2022
4. Bouguelia, S., et al.: Reusable abstractions and patterns for recognising composi-
tional conversational flows. Proc. CAiSE 2021
5. Brabra, H., et al.: Dialogue management in conversational systems: a review of
approaches, challenges, and opportunities. IEEE Transactions on Cognitive and
Developmental Systems (2021)
6. Bradley, N., et al.: Context-aware conversational developer assistants. In: 2018
IEEE/ACM 40th ICSE. pp. 993–1003. IEEE (2018)
7. Dumas, M., et al.: Fundamentals of business process management, vol. 1. Springer
(2013)
8. Gaulke, D., et al.: Interactive user interface to communication-enabled business
process platforms method and apparatus (May 26 2015), uS Patent 9,043,407
9. Hofmann, P., et al.: Robotic process automation 30(1), 99–106 (2020)
10. Kalia, A.K., et al.: Quark: a methodology to transform people-driven processes to
chatbot services. In: Proc. ICSOC. pp. 53–61. Springer (2017)
11. Kobeissi, M., et al.: An intent-based natural language interface for querying process
execution data. In: Proc. ICPM. pp. 152–159. IEEE (2021)
12. López, A., et al.: From process models to chatbots. In: International Conference
on Advanced Information Systems Engineering. pp. 383–398. Springer (2019)
13. Matthies, C., et al.: An additional set of (automated) eyes: chatbots for agile
retrospectives. In: Proc. BotSE. pp. 34–37. IEEE (2019)
14. Rizk, Y., et al.: A unified conversational assistant framework for business process
automation. arXiv preprint arXiv:2001.03543 (2020)
15. Sheth, A., et al.: Cognitive services and intelligent chatbots: current perspectives
and special issue introduction. IEEE internet computing 23(2), 6–12 (2019)
16. Zamanirad, S., et al.: Hierarchical state machine based conversation model and
services. Proc. CAiSE 2020
17. Zamanirad, S., et al.: Programming bots by synthesizing natural language expres-
sions into api invocations. Proc. ASE 2017