Proceedings of Recent Advances in Natural Language Processing, pages 295–303,Varna, Bulgaria, Sep 2–4, 2019.

https://doi.org/10.26615/978-954-452-056-4_034

295

OlloBot - Towards A Text-Based Arabic Health Conversational Agent:Evaluation and Results

Ahmed FadhilDepartment of Computer ScienceUniversita Degli Studi di Trento

Trento, Italyahmed.fadhil@unitn.it

Ahmed AbuRa’edInformation & Communication Technologies

Universitat Pompeu FabraBarcelona, Spain

ahmed.aburaed@upf.edu

Abstract

We introduce OlloBot, an Arabic conver-sational agent that assists physicians andsupports patients with the care process.It doesn’t replace the physicians, insteadprovides health tracking and support andassists physicians with the care deliverythrough a conversation medium. The cur-rent model comprises healthy diet, phys-ical activity, mental health, in addition tofood logging. Not only OlloBot tracksuser daily food, it also offers useful tips forhealthier living. We will discuss the de-sign, development and testing of OlloBot,and highlight the findings and limitationsarose from the testing.

1 Introduction

According to World Health Organisation (WHO),poor diet and physical inactivity have tremendousimplications on individuals health (Michie et al.,2009; Hard et al., 2012). Healthy diet helps pro-tect against malnutrition in all its forms, as well asnoncommunicable diseases (NCDs) (Michie et al.,2009). Middle eastern countries are among themostly affected nations by poor diet and physi-cal inactivity and its effect on cardiovascular dis-eases (Organization et al., 2010, 2012). Accordingto WHO report, total deaths caused by NCDs areamong the highest in countries such as, Saudi Ara-bia, Iraq, Qatar, and Bahrain (See Table-1). This isan evident for a global risk and there is a need fornationwide approach to help mitigate or preventthis escalation.

By 2025, AI systems could be involved in ev-erything from population health management, tovirtual assistants capable of answering specific pa-tient queries. Overall, AI has the potential to im-prove outcomes by 30 to 40 percent while cutting

The NCDs and Middle East (WHO)Countries % of total deaths by NCDs

Saudi Arabia Cardiovascular diseases = 46%Qatar Cardiovascular diseases = 24%Iraq Cardiovascular diseases = 33%Bahrain Cardiovascular diseases = 26%

Table 1: Middle East Countries and NCD Burden.

treatment costs by as much as 50 percent (Kohet al., 2011). Conversational agents, an example ofAI powered systems can communicate with usersthrough an intelligent conversation using a naturallanguage, they can aid doctors in enhancing pro-ductivity and enabling them to respond to patientsquickly. For example, doctors could interact witha diet chatbot to recommend the appropriate foodto their patients. Patients could also engage withthis chatbot to get instant information about theirdietary choices.Currently, some ways to use healthcare chatbotsinclude: scheduling doctor appointments basedon the severity of the symptoms, monitoring thehealth status and notifying a human nurse imme-diately if the parameters are out of control, help-ing home-care assistants stay informed about pa-tients evolution. However, current health conver-sational agents are mature for English languagebut still in their infancy for specific demograph-ics and languages. For example, Hebrew and Ara-bic are far more complex languages for conversa-tional agents. To ensure timely health informa-tion delivery to Arabic users, the patient shouldbe able to tell the bot what symptoms they are ex-periencing and receive medical advice. However,this is not possible with current approaches, sincethey merely focus on generic approaches. Hence,in this research we highlight the role of chatbotsto provide health services to individuals with lan-

296

guage barriers. The approach targets demograph-ics who are not bilingual and can speak only Ara-bic language. This is important, since from Table-1 we see the burden of NCDs in Arabic speakingcountries. We propose OlloBot, an Arabic conver-sational agent able to converse with users and han-dle daily tasks about diet, physical activity, men-tal wellness and coping. The bot can also trackusers daily food and keep a record of their dailydietary habits. To our knowledge, few studies ex-ists on conversational agents that supports Ara-bic language, able to flexibly converse and han-dle dialogues. However, no study exists that haveconsidered the health benefit achievable with lan-guage specific conversational agents. Few plat-forms provide support to Arabic language technol-ogy to build interactive and intelligent conversa-tional agents. This might be due to the complexityof the language and the resource scarcity availableto support it. Our approach relies on IBM Wat-son Conversation API (IBM bluemix)1 to handlethe dialogue structure and Telegram Bot Platformto build the chatbot. We tested OlloBot with 43Arabic speaking users and presented the findingsin this paper.

2 Background Research

Although cognitive behavioral therapeutic (CBT)apps have demonstrated efficacy, still they arecharacterized by low adherence rate (Fitzpatricket al., 2017). Conversational agents, on the otherhand, may offer a convenient and engaging alter-native of giving support at any time. A work byGraf et al., (Graf et al., 2015) built a chatbot thatassimilates into daily routines. The bot communi-cates with user and gathers nutrition data. Keep-ing interaction with the bot final is a good designpractice, however building a great dialogue flowis also essential to ensure smooth user interaction.This includes the way the bot handles several userinteractions and requests. Zaghouani et al., (Za-ghouani et al., 2015) presented a correction an-notation guidelines to create a manually correctednonnative (L2) Arabic corpus. The work extendsa large scale Arabic corpus and its manual cor-rections to include manually corrected non-nativeArabic learner essays. The approach uses anno-tated corpus to develop components for automaticdetection and correction of language error to helpstandard Arabic learners and improve the quality

1https://www.ibm.com/cloud-computing/bluemix/it

of Arabic text produced. Chatbots could be usedas language learning tools, to access information,to visualise the context of a corpus and to give an-swers to specific questions (Abu Shawar, 2005). Awork by Shawar et al., (Abu Shawar, 2005) built ageneral language chatbot that supported differentlanguage among which is Arabic and English. Thestudy also used different corpora structure, such asdialogue, monologue and structured text to buildthe dialogue model.A work by Ali et al., (Ali and Habash, 2016)presented BOTTA, an Arabic dialect chatbot thatcommunicates with users using the Egyptian Ara-bic dialect. Another work by Shawar et al.,(Shawar and Atwell, 2009) described a techniqueto access Arabic information using chatbot withnatural language processing (NLP) tools. The botprovides responses to capture the logical ontologyof a given domain using a set of pattern-templatematching rules. The literature shows no studiesconsidering conversational agent as assistive toolfor Arabic speakers. Which could provide instantaccess to health information and data, and assistphysicians in providing a follow up to the patients.Conversational agents are great in handling repet-itive tasks which consumes most of the health-care providers time. Sundermeyer et al., (Sunder-meyer et al., 2014) build a two translation recur-rent neural models. The first one is a word-basedapproach using word alignments, while the secondpresents phrase-based translation models that aremore consistent with phrase-based decoding. Themodels are capable of improving strong baselinesin BOLT task for Arabic to English.

3 OlloBot Architecture

To provide the required level of Natural LanguageUnderstanding (NLU), we built the Arabic dia-logue states on IBM Watson Conversation2, whichdefines the conversation flow and dialogue states.The platform provides Artificial Intelligence sup-port to catch different user intents and entities.The OlloBot starts with the user sending a mes-sage to the bot (OlloBot) running on TelegramBot Platform3. The application provides the userwith the topics they can chat about, then basedon user selection the bot forwards the request toIBM Watson Conversation cloud. The conversa-tion slots takes user input and provides them with

2https://www.ibm.com/watson/services/conversation/3https://telegram.org/blog/bot-revolution

297

relevant answer after checking their intent, entityand condition of the conversion. The logical partsof the dialogue are handled by a Node.js wrap-per to handle unmatched dialogues by the dialogueflow ( Figure-1 illustrates the high-level architec-tural view of OlloBot).

Figure 1: OlloBot High-level Architecture.

The bot detects patterns in user phrasing andcorrelates them with the intents defined. For ex-ample, conversing about users daily meal and ask-ing users to enter their breakfast, they can eitherprovide a list of food items or send an emojis as alist. The bot can extract the intend from each itemor emoji and build a list of food items. In prin-ciple, this induces the system to recognize fooditems inserted by users. The need for conversa-tional agents rises due to the on-demand 24/7 ac-cess to care. However, with millions of peopleglobally who struggle with poor diet, poor exer-cise, anxiety and depression, there isn’t enoughhealthcare provider or psychologists to provide thenecessary care. Moreover, in some parts, thereis limited technological services supporting na-tive languages and sometimes services are prac-tically nonexistent. Due to the associated cost forcare, providing conversational agents as a noveltechnology that is cost effective, efficient and tar-gets the right demographics is essential. OlloBotbridges these gaps and provides Arabic users witha quick and accessible health services. The botconverses about diet, exercise, emotion and pro-vide coping skills. It can also ask the user tolog their daily food and gather data about theirpreferred style of diet and exercise day, exercise,sleep. The fact that OlloBot is supported by Wat-son and easily accessible by many users makesit able to support many users at the same time.

The chatbot gathers users interaction data, and per-formance indicators from their interaction. Thesedata are saved in a database in the form of acces-sible reports by the caregivers. OlloBot acts as ahealth assistants, meaning that it offers help andsupport rather than treatment.

3.1 User Intents

The intents refers to what the users want fromthe bot or what’s the objective behind a user’s in-put. For instance, the user intents hi, hello, hiOlloBot are translated to the intent “Greetings”,whereas the intents food logging, daily food, log-ging data...etc are all translated into “Food Log-ging”. We carefully sketched the dialogue scopethe bot covers and those that it does not (see Table-2 for a list of user intents). The dialog flow struc-ture was designed in collaboration with a health-care clinic. After defining the flows for key in-

User IntentsGreetings Coping Skills

Diet and nutrition Daily meals LoggingMood Jokes

Physical activity ConditionsAnything else Farewells

Table 2: The User Intents

tents, we defined the bot responses or follow up,once the task is performed. When the user startsa conversation, the bot can either track user dia-logue or leave it at resolution and reset. The botcan switch between intends, based on user input.Since the interaction medium is conversational,users can switch intents on the chatbot. For ex-ample, while the bot waits for the user to providerelevant information about their physical activity,the user can ask the bot to insert their breakfast.OlloBot switches between the topics given in theabove table, by detecting user intent from the con-versation. To accurately handle this switching, wedesigned additional flows represented by condi-tions. Although this adds flexibility to the inter-action, however, it can also create additional cog-nitive load.The next bot response depends on what the userwill say or choose from the buttons. For example,after the user provides their daily food logging, thebot can move on and ask them about their suitabletime to be notified again. Users can come back tothe chatbot at a later time, with no recollection of

298

the task they were trying to accomplish, thereforetracking the dialogue flow is essential to offer amore flexible conversation to the users. Finally, tohandle fallbacks where the bot has no clue aboutthe respond, we designed the “anything else” in-tent, to handle unhandled intents.

3.2 Health Report

The chatbot application provides a health reportgenerated by users interaction with the chatbot.This report contains user activity data, their over-all interaction with the application and indicationsabout their overall health. The health measure ismainly about their diet, exercise and stress pat-tern. All these data are structurally generated bythe chatbot and saved into the database. These datacan then be accessed by the caregiver to obtain rel-evant user specific data.

3.3 Entities

Entities are the pieces of valuable information hid-den in users input. They’re important keywordsextracted from a sentence. For example, in the ut-terance “I want to talk about physical activity”, theword “physical activity” is detected as an entity,and hence the bot switches to the physical activ-ity topic. Entities focus on defining the topic theuser is talking about. This is important to providethe right respond to user questions. We defined abig range of entities to build our dialogue model.These entities represented the four main topics andthe daily food logging. In Table-3 we provide theentities listed to structure the dialogue flow.

Topic EntitiesTopics Meal times

Daily times Weekly timesQuantity FoodNumbers Sport

Table 3: The Topic Entities

The above table lists the “Topics” entity, namelydiet, physical activity, mood, and coping that thebot converses about. The topics entity also coversusers daily food logging. Other entities includethe “Meal times” which defines the meal periodsper day that includes: breakfast, lunch, snacksand dinner. The “Daily times” refers to the peri-ods of the day, namely morning, noon, afternoon,evening and night. Whereas, “Weekly times” en-tity includes the period of the day, namely yes-

terday, today and tomorrow. The “Quantity” en-tity refers to the quantity measurements the usermight mention in the conversation (i.e., kg, g, teaspoon, bread loaf), whereas the “Numbers” en-tity refers to the countable number the user mightmention. We have also defined a comprehensivelist of common Arabic food items and includedthem in the “Food” entity. Finally, we defineda list of any kind of sports in the “Sport” entity.The entity list was accompanied by a list of syn-onym to add flexibility for the entity detection.This is important to detect user intents and sen-timents from their conversation. While there hasbeen a lot of research on sentiment analysis in En-glish, the amount of research and datasets for Ara-bic language is still limited. A work by Alayba etal., (Alayba et al., 2017) built a sentiment analy-sis dataset in Arabic from Twitter data. The studyapplied machine learning algorithms (i.e., NaiveBayes, Support Vector Machine and Logistic Re-gression) for the analysis. Another work by Ismailet al., (Ismail and Ahmad, 2004) proposed a newtype of recurrent neural network architecture forspeech recognition and Backpropagation ThroughTime (BPTT) learning algorithm to observe differ-ences in alphabet “alif” until “ya”.

3.4 Dialogue Engine

The dialogue structure was designed for each ofhealthy diet, physical activity, and mental wellnesstopics by referencing the Cognitive-behaviouraltherapy (CBT) (Rothbaum et al., 2000). We basedthe chatbot dialogue on techniques suggested bythe CBT.The dialogue was then developed on IBM WatsonConversation, where we listed each of the intentsand entities and built the dialogue structure andflow. The tasks were intentionally built to be sim-ple to interact with, so to decrease the time andamount of physical and mental efforts needed, andincrease user’s ability. For example, when askedto log their food, users can either write the list orsend an emoji of the food item. Moreover, but-ton replies were provided to further simplify thedata logging. Simplifying food tracking processwill increase users ability and decrease the learn-ing curve associated to health tracking. This is be-cause interaction with the bot shouldn’t be onlyconversational, since some interactions are betterwith Graphical UI and others with ConversationalUI.

299

3.4.1 Healthy Diet

Around 2 billion people are overweight, but manyare ready to change (Van Itallie, 1985). However,according to studies (Mann et al., 2007) tempo-rary fixes to old habits makes people to regain theirweight. OlloBot acts as an interactive AI-powereddiet tracking bot that converses with individuals ina friendly way directly through the Telegram mes-saging application. The goal is to give instant ad-vices with each meal eaten and help improve theeating habits on the go. Once the conversation isexecuted, OlloBot starts conversing about user’sdiet and asks them questions about their eatinghabits and highlights the values associated withhealthy diet. For example, OlloBot stresses thefact that following a diet rich in vegetables andfruits helps decrease escalation into overweight,obesity or even chronic conditions, and the detri-mental effects associated otherwise.

3.4.2 Physical Activity

Studies (Cooney et al., 2014; Mead et al., 2009;Artal et al., 1998; Byrne and Byrne, 1993) haveshown that exercise can treat mild to moderatedepression as effectively as antidepressant medi-cation and with no associated side-effects. Ourconversational agent chats with users about theirphysical activity and daily energy level, and canprovide personalized plans and keep track ofworkout progress by storing relevant user inputs.Whether the user wants to stay fit, loose weight,or get toned, OlloBot can later provide an effi-cient and consistent workout plan while keepingtrack of the progress. Although the bot is notconsidered as a tool to loose weight through ex-ercise, but rather a supportive tool to help individ-uals track their exercise and improve in on the go,making users conscious about their health habitsalso makes them more likely to set aside time forphysical activities in the future.

3.4.3 Mental Wellness

Mental health refers to our overall psychologicalwell-being. This includes the way we feel aboutourselves, the quality of our relationships and theability to manage our feeling. OlloBot chats withusers about mental health and asks them questionsabout their stress, sleep and other measures rele-vant to their mental wellbeing. There exists sev-eral mental health chatbots that provide supportat different stages of mental illness. For exam-

ple, X2AI4 created a set of chatbots for mentalhealth applications. Their flagship AI, Tess, helpspatients in tandem with their doctors by providingresources on cognitive-behavioral therapy, medi-cation side effects, and questionnaire automation.Woebot5 is a mood tracking chatbot with person-ality. Backed by scientific research, Woebot canhelp reduce depression, share CBT resources, andlearn from conversations over time. Finally, Joy6

uses a chatbot approach to mental health. She of-fers options for both individuals and therapists.

3.4.4 Coping SkillsBeing mentally or emotionally healthy is morethan being free of depression, anxiety, or otherpsychological issues. Rather than the absence ofmental illness, mental health refers to the presenceof positive characteristics. With OlloBot, we han-dle this by providing coping skills and mindful-ness support through clever motivational quotes,relevant workout suggestions mixed with somehealth facts to help users understand the benefitof health and wellness. The bot provides the useswith tips about coping with stress, diet, exercise,sleep, and mindfulness. The quotes and sugges-tions are all based on best recommendations pro-vided by the World Health Organisation (WHO)(Michie et al., 2009). Coping skills are important,since finding a moment to take a few deep breathsand quiet your mind is a great way to relieve stressand improve your overall health.

3.4.5 Daily Food LoggingWhile many food-logging tools are already on themarket, most are either too complicated or bor-ing for the average individual. OlloBot keeps afood diary, tracks calories, and provides basic nu-tritional tips based on user’s eating habits. Wewanted to make the food logging process sim-pler, more engaging, and more informative for theusers. The bot tracks user meals, by asking themto insert each meal items. This helps to know moreabout users diet. With the NLP capabilities of-fered by IBM Watson Conversation, the bot sup-ports any food combinations. The user can evenlog their food by sending emojis of the food items.Once the food logging is done, the bot asks theuser about their preferred time to recheck withthem. OlloBot is in its early stages, due to the time

4https://x2.ai/5https://woebot.io/6http://www.hellojoy.ai/

300

Figure 2: Daily Food Logging Conversation with OlloBot.

required to build and integrate new features, thenvalidate it through testing with real users. We un-derstand that existing food logging apps are nowfar ahead of us in terms of calorie tracking, pre-cision, and various interface features. However,we aim to close this gap and simplify the foodtracking/logging and feedback providing processin the future (see Figure-2 for the daily food log-ging with OlloBot).

4 OlloBot Platform

The dialogue flow in OlloBot were handled withIBM Watson Conversation, whereas the logicalpart of the dialogue were handled with a Node.jsimplementation. Our choice of IBM Watson wasdue to its language support, to our knowledge itwas the only conversational agent dialogue build-ing platform that supports Arabic. We deployedthe bot on Telegram Bot Platform and built thecomponents to handle user requests and responses.We used various UI elements made available byTelegram and integrated them into our dialoguemodel.

5 OlloBot - USE Questionnaire

After building and integrating OlloBot with Tele-gram Bot Platform, we conducted a user exper-iment with 43 Arabic speaking users. The dia-logue and various questions the bot asks were allbased on a WHO questionnaire we extracted abouthealth and wellbeing (Kessler and Ustun, 2004;Parslow and Jorm, 2000; , WHO et al.(2017; Or-ganization et al., 2006). After testing the chatbot,we performed a survey analysis to test various as-

pects of the bot. The questionnaire is based on astandard framework, namely USE Questionnaire(Lund, 2001) to measure the usability of the chat-bot. The framework tests four items of usabilitywithin a product. It consists of 30 questions totest the usefulness, easy of use, ease of learningand satisfaction with the application. The ques-tionnaires are organised in a scale of 1 - 5, where1 is “strongly disagree” and 5 is “strongly agree”.We describe the experiment settings and results inthe following sections.

5.1 Participants DemographicsThe user demographics consisted of both male(n=26) and female (n=17) participants with an agerange of 20 - 65 years old for both gender. Wechecked the participants familiarity with trackingdevices (e.g., any diet, sleep, physical activity, ormood tracking application) and chatbot applica-tions (see Table-4 for the results). There was nosignificant differences in both cases with eithertracking devices (e.g., wearable, sensor, or mobileapplications) or chatbot applications. The major-ity of participants have shown no familiarity withtracking devices (n=26) and chatbot applications(n=27). We provided additional questions at theend of the survey to evaluate their overall experi-ence with the chatbot. A reimbursement of 5ewasgiven to all participants in the form of Amazoncoupon.

5.2 ExperimentWe distributed the chatbot to all the Arabic speak-ing participants recruited from Iraq. After carryingout the experiment for 1 week, we collected data

301

Gender Female 17Male 26

Age Mean 29.8Std. Dev 9.28

Tracking devices familiar No 26Yes 17

Chatbot familiar No 27Yes 16

Table 4: Participants Demographics.

through questionnaire to test the four main pointsrelevant to the usability of OlloBot.

5.2.1 UsefulnessThis item helps to test how useful the participantsperceived the application. The usefulness includesmeasuring whether the bot helps the user to bemore effective and productive and enhance theircontrol over their daily life activities. For exam-ple, “I believe is effective to track my health” eval-uates whether the user thinks the bot is effective intracking their health.

Dimensions Mean Std.Dev.

t value(df=42)

p value

Usefulness 3.381 0.373 6.695 p <.01Ease Of Use 3.626 0.3944 10.405 p <.01Learning 3.942 0.5341 11.564 p <.01Satisfaction 3.505 0.3667 9.031 p <.01

Table 5: The Results from the USE Questionnaire.

5.2.2 Ease of UseThis point helps understand the ease of use aspect.It measures easiness, simplicity and user friendli-ness of the application (chatbot). This point alsomeasures the steps and effort required to achievethe goal set and whether its easy to recover frommistakes. For example, to measure the effort re-quired for each step, we asked users to providetheir scale for “It requires the fewest steps possibleto accomplish what I want to do with it”. We per-formed a descriptive statistics about the four itemschecked in the overall experience (see Table-6).Figure-3 below lists each of the overall bot relia-bility (Figure-3a), and the participants overall ex-perience (Figure-3b, Figure-3c and Figure-3d).

5.2.3 Ease of LearningThis point measures the learnability of the tool.We measure whether its quick to learn and easy

to remember each time. For example, the question“I quickly became skilful with it” checks for howquickly a skill is learned using the chatbot.

5.2.4 Satisfaction

This step checks for user’s overall satisfactionwith the chatbot. It checked whether the users aresatisfied and would recommend the tool to othersand how entertaining they perceived it. For exam-ple, the question “I would recommend it to a friendor family member” checks whether the user wouldrecommend the bot to their relatives or friends.We performed descriptive statistics on the resultsand reported them in Table-5. One Sample t testshowed that averages for all four scales were sta-tistically significantly different from the middlevalue (value = 3, see Table-5). Further analy-ses were performed considering gender, familiar-ity with tracking technology, and familiarity withchatbot as between factors. No differences amongthe four usability dimensions were observed be-tween male and female participants, and no dif-ferences emerged between participants who fre-quently use or not use tracking devices. A signifi-cant difference for the “ease of learning” scale wasobserved when comparing participants who werefamiliar in interacting with chatbot application andparticipants who were not. The latter group re-ported lower scores for learnability compared tothe former group (t(41)= -2.46, p <.01).

5.2.5 Overall Experience

This part evaluated the overall experience with Ol-loBot and is not part of the USE questionnaire. Wetested user satisfaction with the reliability of Ol-loBot, their overall experience with the bot fromdifferent perspectives (see Table-7 for the expe-rience comparison). We have checked whetherusers like chatbots or rather use a mobile appli-cation. Finally, we considered the list of mostpositive and negative aspects the users mentionedduring the survey. In addition, we asked the usersabout the features they would like to see/use in fea-ture versions of OlloBot.

6 Discussion

This work was the first to evaluate the applicationof NLP powered health tools into language andcontext scarce domains. The work involved ini-tial design phase, which involved researchers andhealth experts in the context of healthy lifestyle

302

(a) Overall Reliability Satisfac-tion.

(b) Participants Overall experi-ence.

(c) Participants Overall experi-ence.

(d) Participants Overall experi-ence.

Figure 3: The Overall Experience with OlloBot.

App vs. Chatbot Not at all satisfied vs. Extremely satisfied Terrible vs. Wonderful Frustrating vs. Satisfying Dull vs. StimulatingValid 43 43 43 43 43Mean 3.860 3.163 3.767 3.581 3.512

Std. Dev. 0.9900 0.8710 0.7508 0.6980 0.70028Min 1.000 1.000 2.000 2.000 2.000Max 5.000 5.000 5.000 5.000 5.000

Table 6: The Descriptive Statistics for Chatbot Experience.

Overall ExperienceQuestions Scale = 1 Scale = 5

How satisfied are you with the reliability of this chatbot? Not at all satisfied Extremely satisfiedCan you rate your experience with the chatbot: Terrible WonderfulCan you rate your experience with the chatbot: Frustrating SatisfyingCan you rate your experience with the chatbot: Dull StimulatingWould you rather use the chatbot or prefer a mobile app totrack your everyday health ?

Use an app Use a chatbot

Table 7: Users Overall Experience With OlloBot.

promotion. Involving the health expert was nec-essary to guarantee the real context applicabilityof the application. The dialog structure was builtin collaboration with a dietitian who described thesteps necessary when building a dialog with a pa-tient. We then applied this design paradigm intoour dialog engine. The existing work has severallimitations worth mentioning. Since most of theresearch work was carried out in Italy in collabora-tion with Spain, the dietitian involved in designingthe dialog was a native Italian. Yet, the guidanceobtained from the expert was translated into Ara-bic language and integrated into the Watson dia-log engine. We also acknowledge that the size ofthe participants and the period of experiments, al-though provide good indications, are not enoughto conclude any long-term effect. Future work willconsider building a standalone dialog model andincluding a larger user samples and over an ex-tended period.

7 Conclusion

Chatbots can be a trustworthy assistant, like a car-ing nurse, which provides registration services andpatient follow up. Medical practices can rely onchatbots to capture leads and provide 24/7 supportto existing patients, answering their simple, repet-itive questions using a pre-designed answers. Itwill not offer a diagnosis, but it can remind pa-tients to take their drugs or help them check foran unusual side effect. Most users showed interesttowards social intelligence of the bot. Hence, toneand empathy matters, people can be turned off ifthe experience is too robot-like or casual, so weshould opted for a polite tone. In summary, chat-bots offer a great user experience to patients byjust chatting with the bot to get relevant answers totheir queries. Future work will integrate OlloBotinto a health coaching system and make it specificto support Arabic speaking, diabetic patients.

303

ReferencesBayan Aref Abu Shawar. 2005. A corpus based ap-

proach to generalise a chatbot system. Ph.D. thesis,University of Leeds.

Abdulaziz M Alayba, Vasile Palade, Matthew Eng-land, and Rahat Iqbal. 2017. Arabic language sen-timent analysis on health services. arXiv preprintarXiv:1702.03197 .

Dana Abu Ali and Nizar Habash. 2016. Botta: Anarabic dialect chatbot. In COLING (Demos). pages208–212.

Michal Artal, Carl Sherman, and Nicholas A DiNubile.1998. Exercise against depression. The Physicianand sportsmedicine 26(10):55–70.

A Byrne and DG Byrne. 1993. The effect ofexercise on depression, anxiety and other moodstates: a review. Journal of psychosomatic research37(6):565–574.

Gary Cooney, Kerry Dwan, and Gillian Mead. 2014.Exercise for depression. Jama 311(23):2432–2433.

Kathleen Kara Fitzpatrick, Alison Darcy, and MollyVierhile. 2017. Delivering cognitive behavior ther-apy to young adults with symptoms of depres-sion and anxiety using a fully automated conver-sational agent (woebot): A randomized controlledtrial. JMIR Mental Health 4(2):e19.

Bettina Graf, Maike Kruger, Felix Muller, AlexanderRuhland, and Andrea Zech. 2015. Nombot: sim-plify food tracking. In Proceedings of the 14th In-ternational Conference on Mobile and UbiquitousMultimedia. ACM, pages 360–363.

Eat Smart Play Hard, Healthy Kansas Schools, In-ClassPhysical Activity Breaks, and Blast Off Game. 2012.Physical activity. Fitness and health: Internatio .

Saliza Ismail and A Ahmad. 2004. Recurrent neuralnetwork with back propagation through time algo-rithm for arabic recognition. Proceedings of the18th ESM Magdeburg, Germany pages 13–16.

Ronald C Kessler and T Bedirhan Ustun. 2004. Theworld mental health (wmh) survey initiative versionof the world health organization (who) composite in-ternational diagnostic interview (cidi). Internationaljournal of methods in psychiatric research 13(2):93–121.

Hian Chye Koh, Gerald Tan, et al. 2011. Data miningapplications in healthcare. Journal of healthcare in-formation management 19(2):65.

Arnold M Lund. 2001. Measuring usability with theuse questionnaire12. Usability interface 8(2):3–6.

Traci Mann, A Janet Tomiyama, Erika Westling, Ann-Marie Lew, Barbra Samuels, and Jason Chatman.2007. Medicare’s search for effective obesity treat-ments: diets are not the answer. American Psychol-ogist 62(3):220.

Gillian E Mead, Wendy Morley, Paul Campbell, Car-olyn A Greig, Marion McMurdo, Debbie A Lawlor,et al. 2009. Exercise for depression. CochraneDatabase Syst Rev 3.

Susan Michie, Charles Abraham, Craig Whittington,John McAteer, and Sunjai Gupta. 2009. Effectivetechniques in healthy eating and physical activity in-terventions: a meta-regression. Health Psychology28(6):690.

World Health Organization et al. 2006. Global strategyon diet, physical activity and health: a framework tomonitor and evaluate implementation .

World Health Organization et al. 2010. WORLDHEALH REPORT (The): Health Systems Financ-ing: the path to universal Coverage (Arabic). WorldHealth Organization.

World Health Organization et al. 2012. World healthday 2012: ageing and health: toolkit for event orga-nizers .

Ruth A Parslow and Anthony F Jorm. 2000. Who usesmental health services in australia? an analysis ofdata from the national survey of mental health andwellbeing. Australian and New Zealand Journal ofPsychiatry 34(6):997–1008.

Barbara Olasov Rothbaum, Elizabeth A Meadows, Pa-tricia Resick, and David W Foy. 2000. Cognitive-behavioral therapy. .

B Abu Shawar and Eric Atwell. 2009. Arabic question-answering via instance based learning from an faqcorpus. In Proceedings of the CL 2009 InternationalConference on Corpus Linguistics. UCREL. volume386.

Martin Sundermeyer, Tamer Alkhouli, Joern Wuebker,and Hermann Ney. 2014. Translation modeling withbidirectional recurrent neural networks. In EMNLP.pages 14–25.

Theodore B Van Itallie. 1985. Health implications ofoverweight and obesity in the united states. Annalsof internal medicine 103(6 Part 2):983–988.

World Health Organization (WHO et al. 2017. Obesityand overweight factsheet from the who. Health .

Wajdi Zaghouani, Nizar Habash, Houda Bouamor, AllaRozovskaya, Behrang Mohit, Abeer Heider, and Ke-mal Oflazer. 2015. Correction annotation for non-native arabic texts: Guidelines and corpus. InLAW@ NAACL-HLT . pages 129–139.