=Paper=
{{Paper
|id=Vol-2515/short1
|storemode=property
|title=A Planning Strategy For Dialogue Management in Healthcare
|pdfUrl=https://ceur-ws.org/Vol-2515/short1.pdf
|volume=Vol-2515
|authors=Milene Santos Teixeira,Mauro Dragoni,Claudio Eccher
|dblpUrl=https://dblp.org/rec/conf/semweb/TeixeiraDE19
}}
==A Planning Strategy For Dialogue Management in Healthcare==
https://ceur-ws.org/Vol-2515/short1.pdf
A Planning Strategy For Dialogue Management in
Healthcare
Milene Santos Teixeira1,2 , Mauro Dragoni2 , and Claudio Eccher2
1
University of Trento, Italy
m.santosteixeira@unitn.it
2
Fondazione Bruno Kessler, Italy
{dragoni,cleccher}@fbk.eu
Abstract. Planning strategies aim to support an agent in achieving a specific
goal. Similarly, dialogue-based strategies are thought for (i) equipping intelligent
systems with data-acquisition capabilities, and (ii) supporting users by provid-
ing, for example, task orientation. In the digital health domain, the interactions
between physicians and patients have the objective of classifying the patients cur-
rent condition to, subsequently, give him/her new instructions. This work aims to
develop a non-deterministic planning based approach for human-machine dia-
logue management in the health domain. The approach is divided in two parts:
the first part focuses on slot-filling dialogues for acquiring information about the
patient in order to classify him/her with respect to clusters. The second part con-
cerns planning for the management of a task-oriented dialogue to give advice to
patients with the intention of giving some guidance on the patient’s treatment.
Both parts are supported by a knowledge-based back-end performing reasoning
every time new data are provided by users. For demonstrating the complexity of
this task, we provide a sample scenario based on the asthma domain showing how,
even with few variables, the management of a whole conversation is challenging.
Keywords: Healthcare · Semantic Technologies · Conversational Agents · Plan-
ning
1 Introduction
Planning is an important component of rational behaviour, it is part of the human in-
telligence and reasoning. It commonly precedes acting; we plan whenever we face a
situation that is complex or that may offer some high risk or cost that can be aggravated
by acting without a plan. In Artificial Intelligence (AI), planning [6] is a process that,
given an initial state, searches a sequence of actions that leads to the goal. To achieve
this goal, it explores all possible paths by anticipating the outputs of the available ac-
tions. In other words, by comprehending the consequence of executing an action, the
Copyright 2019 for this paper by its authors. Use permitted under Creative Commons License
Attribution 4.0 International (CC BY 4.0).
2 M.S. Teixeira et al.
planner selects and structures these actions in some rational order aiming at achieving
a final pre-stated objective.
Dialogue managers (DM) also deal with the problem of action selection. Differ-
ent approaches have been applied to DM [4] and planning is a promising alternative
since, like in a conversation, it aims to achieve a goal. In planning for human-machine
dialogue, each utterance will be treated as an action and, by anticipating the outcome
of each of these actions, a planner is capable of identifying the states that could be
achieved and of choosing the best path that leads to the goal of the dialogue. For exam-
ple, a health assistant agent that conducts a dialogue with a patient facing an emergency
could benefit from planning to optimize the information retrieval and prioritize the use
of questions that will identify the type of situation as fast as possible.
The development of methods that implement automated planning to structure and
manipulate effective human-machine dialogue is still in its early stages, but it has gained
substantial attention in recent years [1, 8, 5]. Several works [9, 2, 7] have identified that
planning for dialogue may become challenging due to the huge number of paths that
a dialogue can take. In addition, dialogues can be complex since they may have non-
deterministic outcomes, the goal may change during the dialogue and user’ answers
may be partial or even misunderstood by the natural language processor.
In this paper, we propose a novel approach combining reasoning and planning for
supporting the effective and efficient management of dialogues. The synergy of these
two techniques allows to dynamically update the behavior of conversational agents
based on the data provided by users. The reasoner is responsible for inferring the most
suitable status of a user (or patient). This activity is performed by exploiting not only
the user data and the integrated conceptual model, but also the proper resources of the
Linked Open Data (LOD) cloud. On the other hand, the planner is in charge of gen-
erating the interactions for supporting a multi-turn conversation with users in order to
acquire the missing information enabling the classification of the users’ status. This
aspect advances the state of the art of conversational agents in healthcare from two
perspectives. First, the planner allows to dynamically update the conversation, hence,
the system is able to acquire information in a most efficient way. Second, the reasoner
enables the access to external knowledge resources that, combined with information
provided by users, allow to improve the inference of the users’ status.
Considering that domain-independence has always been one of the main concerns
in planning [12, 3, 3], for both parts our approach is domain configurable: it expects to
receive domain knowledge for the topic of the dialogue, but a plan can be automatically
generated by any domain-independent planning engine. The human effort to build a
plan is reasonable, since it requires the knowledge of an expert to identify the variables
and constraints related to health, but does not require the developer to understand the
planning generation process.
The remainder of the paper starts with a brief presentation of the background of our
research (Section 2). Then we present our proposed approach (Section 3) followed by
the description about how the reasoning process supports the planning activities (Sec-
tion 4). As an example, we provide a scenario (Section 5) showing the complexity of
performing an effective management of conversations. Finally, we discuss our inten-
tions for future work (Section 6).
A Planning Strategy For Dialogue Management in Healthcare 3
2 Background
We introduce below some basic notions concerning the background inspired the devel-
opment of the proposed work. Due to space limit, we are not able to discuss and cite in
detail all contribution. Hence, the reader is invited to read the main references provided
below for having more details about the mentioned topics.
Automated Planning In AI, automated planning [6] explores the process of planning
in order to comprehend and computationally model it. Automated planning is intended
for systems that aim to reproduce the human reasoning, presenting an autonomous be-
haviour. Whenever verified that planning is required for the problem, in order to mini-
mize risks or costs, it should be made before starting to execute any action.
In planning we consider that the world has different states. A state can be changed
to a new one by the so called actions, which have preconditions that must be met to be
applicable to that state; and present effects that describe what changes in the world after
its application. Planning can be viewed as a search problem that describes the sequence
of actions that must be applied in the world in order to achieve a goal. As result, a plan
is a path that takes an initial state to a goal state.
Dialogue Managers A Dialogue Manager (DM) [10] is responsible for the flow of the
conversation in a dialogue system. The DM communicates with the other parts of the
dialogue system to (i) receive the input from the user and (ii) to provide the output,
which corresponds to the response given by the system in the dialogue. Depending on
the chosen design for the DM, the natural language for both input and output can either
be processed inside the DM or through an external natural language processing (NLP)
module. This is because the focus of the DM is not the NLP, but the strategy that will
be used to manage the conversation and decide what to say next.
3 Proposed approach
In this work we propose a plan based approach exploiting knowledge for dialogue man-
agement of health systems used to monitor and support patients. Our approach is do-
main configurable and can be applied to any topic related to health. It consists of two
parts: (i) information acquiring and classification of the patient and (ii) orientation.
3.1 Information Acquiring and Classification of the Patient
The first part of our work aims to generate a non-deterministic plan for managing a
slot-filling dialogue to retrieve the information that influences the classification of the
current health condition (status) of the patient. The generated plan must not only be
satisfactory (achieve the goal), it must be as short as possible (optimal) with the aim
of detecting critical situations as fast as possible. We provide an architecture for the
specification of a planning domain and, from that, an instance of a planning problem
can be easily generated. The components necessary to specify the planning domain are
represented in Figure 1 and are explained in detail below.
4 M.S. Teixeira et al.
Fig. 1. Proposed architecture to specify the planning domain.
The first step is the identification of the set of slots to be filled through the dialogue
with the patient. A slot corresponds to an attribute (e.g. a symptom) whose value is
required to classify the patient’s health condition. This value is initially fixed as “un-
known” and the further acceptable values should be identified with the support of expert
knowledge. After the specification of this set, each slot is converted to an action that is
understandable by a classical planner. This action is identified by a type (DIALOGUE)
and, basically, corresponds to a question that will be made with the aim of obtaining
the value for the slot. We expect open answers for the dialogue, since they reproduce
closely a human-human dialogue and do not limit or influence the user’s answer or opin-
ion. Consequently, when asking a question (executing the action), it is not possible to
know what will be the patient’s answer and, therefore, the actions of type DIALOGUE
hold non-deterministic effects. These effects correspond to the set of acceptable values
previously defined for the slot and that are mapped to our model by an external natural
language understanding (NLU) component or inferred through the use of a reasoner
since the slot to be filled cannot be directly mapped to the content of users’ answer but
it is the result of a logic analysis of data packages. Details about the interaction between
the reasoner and the planner can be found in Section 4. Instead, concerning the use of
an NLU component, it is anyway mandatory for extracting from natural language text
the intents that the patient wants to communicate to the system as such intents can be
directly mapped to slots or provided as input to the reasoner. The investigation of an
effective NLU component is out scope of the paper. In our work, we rely on existing
NLU strategies.
The actions of type DIALOGUE presented in the plan do not contain the actual
content of the dialogue. This content could either be generated through natural language
generation (NLG) or it can be retrieved from a database. In our work, we are currently
using a database containing questions that should lead to the answers that fill out each
slot. Considering that a non satisfactory answer can be obtained for a given slot and that
repeating an already asked question could lead to a non-natural dialogue, we encourage
the storage of several questions associated with a single slot. Consequently, to support
the decision on which question to ask, each question is associated with a heuristic value
that indicates its relevance or priority. Therefore, during the plan execution, whenever
an action of type DIALOGUE is called, a method identifies the type of slot and retrieves
from the database the still-non-asked question with the highest heuristic for that slot.
The slot filling dialogue may require further actions that do not correspond to ques-
tions. In our approach, these actions are identified by the type SERVICE and can be
A Planning Strategy For Dialogue Management in Healthcare 5
generated to manage the conversation (e.g. setting some constraint). Both types of ac-
tions DIALOGUE and SERVICE were borrowed from [3].
The classification of the patient’s status must also be defined in the planning domain
according to a set of mutually exclusive zones. A zone can be seen as a label that iden-
tifies a status that a patient may be found in and it is composed by a set of slots-values
that must hold for this zone to be recognized. Similarly to the process that converted
slots to actions, each zone will become an action in the planning domain. An action of
type CLASSIFY classifies the patient in a zone and, since this classification is our goal,
it has as effect a predicate called GOAL ACHIEVED. Consequently, this will be the last
action executed in the plan. Finally, the planning problem generation becomes trivial:
– Initial state: all values of slots are set to “unknown”;
– Goal state: GOAL ACHIEVED
3.2 Orientation
Based on the classification made by the synergistic use of reasoning and planning, the
next step consists in the generation of a further plan for managing a task-oriented di-
alogue able to provide directions to the patient, or to communicate with an external
system, in order to act and change the current condition that the patient was found to be
in.
This aspect is part of the future work and it presents interesting research challenges
to be addressed like the generation of natural language text for explaining the results of
the reasoning process and the management of the interaction with two further actors:
the healthcare organization that is in charge of managing possible emergencies of its
patients and the physician that is responsible for the patient’s treatments.
4 The Role of Reasoning
Reasoning plays an important role in inferring the values of attributes exploited for
classifying patient’s status. If we consider the management of conversations and data
acquisition sessions within the healthcare domain, values of attributes can be the result
of the analysis of more complex and structured information and not extracted by single
question/answer interaction. Let us consider, as example, the monitoring of parameters
associated with chronic diseases like the glycemic index for people affected by diabetes
or the number and type of cough episodes within a specific time-span for people affected
by asthma.
The use of a knowledge base designed for specific domains allows to define classes
modeling risky and undesired situations that can be inferred from the analysis of users’
data. The use of reasoners enables the detection of such situations and triggers specific
actions for the planner and, in turn, for the conversational agent. This kind of approach
significantly advances the state of the art of dialogue management, especially within
the healthcare domain where aspects like user profiling, tracking, environmental data
acquisition, etc. play a crucial role in supporting the interaction with patients.
Our solution goes into this direction by providing the theoretical background neces-
sary for developing a new generation of systems able to manage complex conversations
6 M.S. Teixeira et al.
within the healthcare domain for supporting patients better in an effective and efficient
way.
Figure 2 shows how the reasoner can support a classic NLU component during the
slot-filling activity. The figure refers to a single interaction during a dialog.
Fig. 2. Representation of the reasoning support to the slot-fill process.
When a user sends an answer to the system, the NLU component processes the
natural language text and provides a list of intents that are detected. The intents that
the NLU component is able to detect are part of the knowledge provided by the domain
experts. Indeed, for every intent the system needs to know what to do, i.e. to fill a slot
or to invoke the reasoner.
The reasoner is composed by two modules: the Analyzer and the Inference Engine.
The Analyzer is invoked when specific intents are detected within a user’s answer. Here,
the detected intent is merged with three other kinds of information:
– User Profiles: they contain personal information of the user, e.g. the clinical history
stored into his/her Personal Health Record.
– Past User Interactions: the register of all interactions that the user had with the sys-
tem. Such information is relevant for stream reasoning tasks, recidivity detection,
and inference over historical data.
– External Sources: all external knowledge bases, like the LOD Cloud, connected
with the system that can be exploited for both inference and slot-filling purposes.
A Planning Strategy For Dialogue Management in Healthcare 7
The output of the Analyzer is used for filling the slots of the attributes used by the
planner. Once the slots have been filled, the status of planner’s variables is sent back
to the Inference Engine of the reasoner that tries to infer if the user can be classified
within a precise status or not. If the classification is positive, the orientation component
is invoked in order to conclude the dialogue with the user. Otherwise, the planner is
invoked for updating the dialogue workflow based on the slots that have been filled and
for running the subsequent interaction.
5 Sample Scenario
In this Section, we present a simple scenario showing the complexity of managing ef-
fective dialogues supporting the self-management of a chronic disease. We chose the
asthma domain for two reasons. First, this domain needed the management of conver-
sations that are not limited to single sessions but that require a multi-turn conversation
due to the necessity of acquiring information about the different type of symptoms held
by the patient as well as possible recent treatments the patient did. Second, the size of
domain is limited enough for performing “debugging” operations for both the reasoning
and planning activities in order to verify the effectiveness of our solution.
We introduce the example of a chatbot that dialogues with an asthma patient to un-
derstand his current condition and give him some orientation. A plan is built to manage
this dialogue and retrieve which symptoms the patient hold. To build a very simple
scenario, we distinguish only 2 symptoms (slots) that are common for patients with
asthma: chest constriction and fatigue. Each slot may hold true, false or unknown (de-
fault) values and will be converted to an action of type DIALOGUE (a1 d and a2 d,
respectively), whose actual content is mapped to a database of questions (e.g. Have you
had episodes of coughing recently?). Every DIALOGUE action has a non-deterministic
effect, since we cannot know in advance which will be the answer given by the patient.
The symptoms can be classified by the actions of type CLASSIFY in 2 different
clusters/zones (orange a3 c and red a4 c), for which different orientations are given to
the patient. However, to keep this example simple, we consider that the goal is achieved
when the classification of a zone is made.
We have defined that the dialogue is system-initiated and the question for the initial
state (when all values are unknown) is a0 d, which corresponds to the question “How
are you feeling?”. This question also has a non-deterministic outcome since the patient
may inform none, a single symptom or several symptoms (up to 2 in this scenario). This
implies that this single question, may result in several different states, more precisely,
23 .
After generating the planning domain and problem instance, we used PRP plan-
ner [11] to generate a strong cyclic plan. The generated plan is illustrated in Figure 3.
We can appreciate the complexity of the generated dialogue by considering only two
symptoms. The real asthma scenario on which we are working on for our pilot study
contains 10 symptoms. Hence, the size of the plan graph dramatically increases as well
as its computation complexity.
8 M.S. Teixeira et al.
Fig. 3. Generated dialogue plan for the asthma scenario. Suffix d corresponds to DIALOGUE
actions; suffix c corresponds to CLASSIFY actions
6 Discussion and future work
Our proposed approach facilitates how to reason, plan, and converse in the healthcare
domain. It provides a high abstraction level to generate dialogues for different topics,
requiring only the knowledge in the topic.
In dialogue systems for the healthcare domain, making the right question at the right
moment is a relevant challenge. This is because the patient may be in an emergence state
that should be detected and confirmed by the system as fast as possible to execute the
right action. In our approach, whenever the system pre-detects a given zone (state), it
must try to identify if the patient is in an even more critical zone. Proven that this is not
the case (if the state does not hold), the pre-detected zone that identifies that the patient
may be in the most critical condition, has priority to be confirmed (or disregarded). If
no other slot that influences this zone can be confirmed, the system tries to confirm
the previous one. For example, when a symptom that is known to influence a cluster is
acknowledged, instead of keep verifying random symptoms, the dialogue should focus
on confirming the other symptoms that are crucial to determine that the patient is in this
cluster or in a more critical one. So far, this was implemented by associating constraints
to each slot defined to the dialogue. But we are exploring new strategies to improve the
confirmation of a pre-recognized zone in order to avoid false positives.
The adoption of planning techniques for managing this context is not simple: it
has a human cost for modeling the domain and it is a time-consuming process. A rel-
evant challenge that we identified in planning for dialogue is the unreasonable num-
ber of states that are generated due to the many possible outcomes (non-determinism)
of the dialogue. We intend to find a strategy to minimize this issue and optimize the
plan-generation performance, possibly, by discarding the automatically generated alter-
natives that are computationally true, but unlikely to happen in real life. We also aim
to explore the use of open answers by allowing our model to manipulate answers with
A Planning Strategy For Dialogue Management in Healthcare 9
full, partial or no information at all. These types of answers will be used to improve
the heuristic value in the relation slot-question in the database (Section 3.1) and also to
relate a slot to the other (e.g. when a question leads to an answer that fills more than
one slot), facilitating the decision on which question should be made next.
Finally, with the intention of contributing to planning for dialogue as a whole, we
intend to investigate how the characteristics in the health domain can contribute to im-
prove algorithms or methodologies currently used in planning for dialogue in general.
From this, we expect that our approach can be applied in dialogue managers for any
other domain that requires some expert knowledge for giving orientation, rather than
the health domain.
References
1. Bercher, P., Biundo, S., Geier, T., Hoernle, T., Nothdurft, F., Richter, F., Schattenberg, B.:
Plan, repair, execute, explainhow planning helps to assemble your home theater. In: Twenty-
Fourth International Conference on Automated Planning and Scheduling (2014)
2. Black, E., Coles, A., Bernardini, S.: Automated planning of simple persuasion dialogues.
In: International Workshop on Computational Logic and Multi-Agent Systems. pp. 87–104.
Springer (2014)
3. Botea, A., Muise, C., Agarwal, S., Alkan, O., Bajgar, O., Daly, E., Kishimoto, A., Lastras, L.,
Marinescu, R., Ondrej, J., et al.: Generating dialogue agents via automated planning. arXiv
preprint arXiv:1902.00771 (2019)
4. Burgan, D.: Dialogue systems and dialogue management. Tech. rep., DST Group Edinburgh
Edinburgh SA Australia (2016)
5. Foster, M.E., Petrick, R.P.: Separating representation, reasoning, and implementation for in-
teraction management: Lessons from automated planning. In: Dialogues with Social Robots,
pp. 93–107. Springer (2017)
6. Ghallab, M., Nau, D., Traverso, P.: Automated Planning: theory and practice. Elsevier (2004)
7. Kominis, F., Geffner, H.: Multiagent online planning with nested beliefs and dialogue. In:
Twenty-Seventh International Conference on Automated Planning and Scheduling (2017)
8. Lee, K., Lee, Y.S., Nam, Y.: A model of fsm-based planner and dialogue supporting system
for emergency call services. The Journal of Supercomputing 74(9), 4603–4612 (2018)
9. Lison, P.: Towards online planning for dialogue management with rich domain knowledge.
In: Natural Interaction with Robots, Knowbots and Smartphones, pp. 111–123. Springer
(2014)
10. LuperFoy, S., Loehr, D., Duff, D., Miller, K., Reeder, F., Harper, L.: An architecture for
dialogue management, context tracking, and pragmatic adaptation in spoken dialogue sys-
tems. In: Proceedings of the 36th Annual Meeting of the Association for Computational
Linguistics and 17th International Conference on Computational Linguistics-Volume 2. pp.
794–801. Association for Computational Linguistics (1998)
11. Muise, C.J., McIlraith, S.A., Beck, C.: Improved non-deterministic planning by exploiting
state relevance. In: Twenty-Second International Conference on Automated Planning and
Scheduling (2012)
12. Petrick, R.P., Foster, M.E.: Using general-purpose planning for action selection in human-
robot interaction. In: 2016 AAAI Fall Symposium Series (2016)