Proposed approaches for modeling knowledge-intensive processes include declarative, constraint-based solutions, which meet halfway between support and exibility. A noteworthy example is the Declare framework, which is equipped with a graphical declarative language whose semantics can be expressed with multiple logic-based formalisms. So far, the practical use of Declare constraints has been mostly hampered by the di culty of modeling them: the formal notation of Declare represents a steep learning curve for users lacking an understanding of temporal logic, while the graphical notation has proven to be unintuitive. This work presents Declo, a chatbot that allows users to easily de ne declarative constraints using natural language statements provided in the form of vocal or textual input. The supported constraints cover the entire Multi-Perspective extension of Declare (MP-Declare), complementing control- ow constraints with data and temporal perspectives.
Knowledge-intensive processes are known to be complex and unpredictable [
However, Declare is known to be di cult to use in practice since its formal
notation represents a steep learning curve for users lacking an understanding
of temporal logic, while its graphical notation has proven to be unintuitive [
With Declo, the user can build declarative process models using natural
language. A natural language statement can be provided using a chat, or a vocal
input. In the latter case, the vocal input is recorded and interpreted using speech
recognition. In both cases, to interpret the text, the tool rst presented in [
The extracted verbs, objects, and subjects are further enhanced with speci ers, i.e., modal verbs, negations and prepositions. Modal verbs are auxiliary verbs indicating feasibility (mainly \can/could", \may/might", \must", \will/would", \shall/should") and are used to identify certain constraints, e.g., response (expressing the concept of obligation) and precedence (expressing the concept of permission). Negations are identi ed through the negative dependency relation and are used to identify negative constraints. Prepositions express relationships between nouns and other parts of a sentence; for example, the term \immediately" speci es the immediacy of the action (e.g., when specifying constraints of type chain).
Activity extraction. Activities can be either verb-based or noun-based. Verbbased activities are constructed from the semantic components extracted during the previous processing step, where the verb stands for the action, its subject for the actor and its object for the activity's business object. For example, the sentence \a manager processes a claim" translates to the activity process claim, with manager as the actor. Noun-based activities appear in a statement as noun phrases which correspond to process activities, e.g., \processing of a claim". In order to extract these activities, the tool focuses on identifying verbs which describe the process ow, i.e., temporal verbs such as \to start", \to precede", \to end", \to follow", \to happen". The object and/or the subject of the verb are used to generate the activity.
Data and time conditions. Declo supports adding data and time perspectives to declarative constraints, e ectively turning Declare constraints into MPDeclare constraints. Three types of conditions are supported: activation, correlation and time conditions. Data and time conditions express more textual fragments rather than full sentences, unlike constraints; in addition, as they 1 The tool can be found at https://github.com/hanvanderaa/declareextraction are short statements, their variance is much lower than that of declarative constraints. As such, the tool uses pattern matching to extract conditions instead of grammatical parsing used during constraint generation.
Activation conditions de ne requirements that must be met for a constraint to be activated, e.g., the creation of a claim must be followed by a check only if the amount of the claim created is greater than a set value. Correlation conditions must be adhered to when the target of a constraint is being executed, expressing relations that must exist between attributes of the activation and the target. Declo supports two types of correlation conditions, i.e., either that the attribute values should be the same (e.g., \the amount is the same") or di erent (e.g., \the resource is di erent") for the activation and the target of a constraint. Complex conditions are supported by concatenating multiple logical statements: an input string is rst split into substrings by detecting coordinating conjunctions (and and or ) and splitting the string accordingly. Each substring is expected to match an expression. Expressions are then joined with logical operators. Pattern matching is supported on both numerical and categorical attributes.
Time conditions specify the time elapsed between the activation and the target of a constraint, e.g., that the target should occur between 2 to 6 days after the activation. Time conditions speci ed using seconds, minutes, hours and days are supported. 3
Declo is integrated into the Rule Mining toolkit RuM2 and is implemented using Java 11 and the speech recognition API provided by Google Cloud Speech.3
To interact with the chatbot, the user needs to access the MP-Declare editor of RuM, create a new MP-Declare model and press the \Start listening" button or the \Let's write" button in the chat. In the rst case, Declo will listen for an input sentence from the microphone. Once the tool has recognized a sentence, it displays the sentence and adds the detected activities and constraints to the graphical model, to the list of activities and to the list of constraints. Then, Declo asks the user if the sentence was recognized correctly. If this is not the case, it is possible to edit the recognized sentence or discard the recognized constraints entirely and record again. If the user presses the \Let's write" button the sentence can be directly typed into the chat. Data and time conditions can also be added via speech recognition or the chat using the corresponding buttons: \Activation" for activation conditions, \Correlation" for correlation conditions, and \Time" for time conditions. Also these conditions can be modi ed or discarded. The user can repeat the steps above to create additional constraints.
Fig. 1 displays a screenshot of the MP-Declare editor in RuM. The top-left
area of the screen shows the chat. The graphical notation of the constraints is
shown in the top-right area of the screen, with the modi able lists of activities
and constraints in the bottom part of the screen. A screencast demonstrating
the tool is available at https://www.youtube.com/watch?v=M kBu9Bqso.
2 www.rulemining.org
3 https://cloud.google.com/speech-to-text/
A qualitative user study, presented in [
The study was conducted via Skype by a team consisting of a facilitator
and an observer, with the facilitator guiding the participant and the observer
serving in a supporting role. The participant shared her/his screen, whereas the
facilitator started the video recording and began guiding the participant through
a prepared test scenario, which consisted of constructing a declarative process
model using speech recognition. At the end of the study, a follow-up interview
was conducted, focusing on resolving questions about the issues the participants
had encountered. The participants were asked what they liked about the tool,
wished was di erent, under which circumstances they would use it. The study
concluded with a post-survey questionnaire consisting of multi-point Likert scales
including the System Usability Scale (SUS) [
tier 1 tier 2 tier 3 tier 4
(P6, P7) (P3, P8) (P2, P4) (P1, P5)
71.25
3.67
3.67
2.33
2.63
81.25
3.67
3.50
2.00
3.13
[
Compared to the commonly reported average SUS score of 68 [