1. Introduction

2. System Architecture

Figure 1. ExhibitXplorer architecture.

Figure 1. ExhibitXplorer architecture.

2.4. CMS Microservice

The CMS microservice is implemented in Python. It uses the FastAPI framework to create the Web user interfaces. FastAPI was preferred over Flask or Django for the CMS microservice due to its exceptional performance and scalability. Its asynchronous capabilities enable efficient handling of exhibit requests, ensuring a responsive web interface for users. This microservice allows creating and editing information in the database for each exhibit.

The functioning of the CMS microservice is shown in Figure 2.

Figure 2. CMS microservice.

Figure 2. CMS microservice.

CMS communicates with Exhibits and Chatbot microservices via a queue named cms. The microservice provides a REST interface with multiple endpoints through which information about a selected exhibit can be created, edited, or deleted. The textual information about the exhibit (name, description, author, year of creation, material used, dimensions, location in the museum, etc.) is recorded in a MongoDB database using the Exhibits microservice.

An HTML form is used to describe the exhibits. The only required fields are the name of the exhibit and its detailed description. Filling in the remaining fields of the form can be implemented automatically via the ChatGPT. For this purpose, appropriate prompts are generated through which ChatGPT generates the missing information. For this purpose, it is necessary to pass the complete description of the exhibit as the value of the assistant field of the GPT query. All requests are sent to the Chatbot microservice, which implements the requests to ChatGPT.

Multimedia files, including images, videos, and audio, can be associated with each exhibit. The multimedia files are written to the IPFS distributed file system. IPFS is ideal for this purpose as it provides a decentralized and distributed file system, ensuring content availability and reliability. The MongoDB database records only file CID and MIME type of each multimedia file, facilitating their retrieval.

2.5. Chatbot Microservice

The Chatbot microservice, implemented in Python, enables query submission to the ChatGPT using the GPT API. This microservice retrieves personalized information based on visitor segmentation, ensuring customized responses to inquiries and questions.

Museum visitors’ queries to the chatbot microservice is implemented through a message broker, as shown in Figure 3.

Figure 3. Chatbot microservice.

Figure 3. Chatbot microservice.

A microservice Chatbot works as a server that implements visitor requests. Using a mobile application, visitors send their requests. They are transmitted to a chatgpt queue. The requests are JSON objects that contain the following information:

• visitor_id - visitor identification string.
• timestamp – specific request time in Unix format.
• command - name of the command that is passed to the microservice. Two command names are recognized: 1) prompt (user request) and 2) reset - clears the context associated with the given visitor.
• system, assistant, and user - content of the request.

The "user" field contains the visitor request itself. In the context of the GPT API, the "system" and "assistant" fields are used to provide additional instructions and context to the language model, allowing you to customize and guide the behavior of the model. The "system" field is used to specify the initial behavior or role of the assistant. It helps to establish the context of the conversation. The "assistant" field is used to track the context of the conversation. It records both the questions and the chatbot’s responses. When communicating between visitors and a Chatbot microservice, the "system" and "assistant" fields are left blank. This means that the server side must initialize them. In this case, the role being set is museum curator, and the assistant field is initially initialized with the full text information for the specific exhibit. Through the effective use of the "system" and "assistant" fields, dynamic and interactive conversations are created with the GPT language model. To generate content specific to each visitor, the Chatbot microservice adds information to a "user" field so that the visitor’s profile is accounted for. Thus, depending on the segments to which it belongs, content unique to each visitor is generated. The following is a sample response that is generated when a visitor who is visually impaired requests a description of the Mona Lisa painting:

“The Mona Lisa painting is a portrait of a woman with an enigmatic smile. She is looking directly at the viewer with her eyes, and her hands are crossed in her lap. She is wearing a draped garment and a scarf around her neck. Her facial features and expression are painted with delicate brushstrokes and subtle shading, giving her a mysterious and enigmatic look. The Mona Lisa painting is a portrait of a woman with a mysterious expression on her face. She is seated against a background of rivers and rolling hills. Although the portrait is painted in earth tones, there are hints of yellow and ochre in her face and clothing. Her hands are crossed delicately in her lap, and she stares out at us with her small lips and captivating eyes. The painting is thought to have been painted by the Italian artist Leonardo da Vinci in the 1500s. The size of the painting itself is 77cm x 53cm. The painting has become world famous and is admired for its enigmatic expression”.

The description of the picture for a visitor who is a student is as follows:

“The Mona Lisa is a famous painting created by the artist Leonardo da Vinci. It shows a woman with a mysterious smile on her face. She is wearing a dark dress and a beautiful headscarf. The painting is very special because it is one of the most famous and recognizable works of art ever made. People come from all around the world to see it!”

To complete the user requests, each request is processed in a separate program thread. Similarly, in separate program threads, the received responses are processed. The microservice chatbot sends the response to the corresponding visitor using the message broker. The response is sent to a queue whose name matches the visitor ID (see Figure 3).

The microservice keeps the context of the conversation with each visitor through a "Context cache" block. This is a synchronized object that can only be accessed by one program thread at a time. Deleting the context is possible via a “reset” command. This command is generated at the start of each new conversation or after a timeout.

Proactive content generation

The developed service is proactive in generating personalized content. The response time of the GPT API depends on the pattern set at the request but is no shorter than 5-10 seconds. For this reason, this content starts to be generated earlier in time - when the visitor approaches an exhibit at a set distance. This distance is specific to each exhibit and is set in the Exhibits database. When the visitor gets close to an exhibit, he receives a push notification informing him about which exhibit it is. At this point he receives only short, non-personalized content. The visitor can confirm or reject the delivery of the personalized content. If visitor is hesitant, he/she has the option to use the chatbot to get further information. It is likely that the content to be generated has already been generated for another visitor with a similar profile. For this reason, before a request is sent to ChatGPT, it is checked that similar content has not already been generated. To this end, the Visitors database is searched for visitors who had a similar profile to the visitor for whom custom content is to be generated. This is easily implemented in a MongoDB database as it provides operators through which the Jaccard algorithm is implemented. The Jaccard similarity measures the similarity between two sets of data to see which members are shared and distinct. The Jaccard similarity is calculated by dividing the number of observations in both sets by the number of observations in either set.

The algorithm for generating personalized content is as follows:

• Get the segments to which the visitor belongs - array targetSegments. If targetSegments is empty go to 2, otherwise go to 3.
• There is still no profile for the visitor. Retrieve information from the Exhibits microservice using the getInformation command. The exhibit identifier, E_id, is passed to the command. A JSON object info is built that contains all the information about the exhibit that is available in the database. Go to 5.
• The visitor is profiled, and the visitor information should be dynamically generated by the microservice chatbot depending on the value of targetSegments. To speed up the dynamic content generation process, a calcSimilarity command is sent to the Visitors microservice. The goal is to check if there are other visitor(s) that have similar profiling. For this purpose, the database capabilities of MongoDB are used. Using operators $setIntersection, $setUnion and $devide, the probability that each of the other currently active visitors has similar segmentation is calculated. Jaccard’s algorithm is used for this purpose. For the similarity profiling of visitor n and m we can write:

  $${Sim}_{n,m}=\left(Sn\cap Sm\right)/\left(Sn\cup Sm\right),$$

  where Si is the set containing the segments a visitor with identifier i belongs to, and Sim〗_(n,m) is the probability that visitor n and visitor m have similar profiling. The similarity coefficient is a number in the interval [0, 1]. If there are no visitor(s) with a similarity coefficient greater than a pre-set threshold value SimTh goes to 4. For each visitor with Sim ≥ SimTh, it is checked whether there is any generated content for an exhibit with identifier E_id. If such content is missing, go to 4. Otherwise, this content is retrieved and written to the info object. Next, go to 5.
• Since there are no visitors with similar profiling or no dynamic content has been generated for exhibit E_id, this content needs to be generated by chatbot microservice. For this purpose, a getInformation (E_id, Sn) command is executed to obtain dynamic content (object info) for the specific exhibit and visitor. The text part of the response is obtained using the GPT API. For this purpose, the full description of the exhibit is passed as a prompt, and depending on the segmentation, filters are set for this content. The remaining information (images, video, audio, and links to Internet sources) is filtered according to the segments to which the visitor belongs.
• Return info object.

3. Results

3.2. Databases

For the application to deliver content it is necessary to create the databases that describe the geofences and exhibits.

3.2.1. Geofences database

A Web application has been developed that allows a quick description of the contours of the geofences as a polygon or circle with a given center and radius. The possibility of converting all geofences described with a circle to a polygon is provided. Each geofence is assigned a unique name that is used as the geofence identifier in the database. At any time, geofences can be exported to a file or saved to a MongoDB database in GeoJSON format. The application also allows the description of geofences using geohashes to reduce the execution time of geospatial queries. When working with MongoDB and indexing the database with the 2dsphere index, it is not necessary to use a geohash description, as MongoDB automatically implements the conversion of polygons to geohashes.

Figure 4 shows the interface of the application that creates the geofences database.

Every visitor to the museum falls into one or more geofences. To keep track of which visitors are in the museum, the outline of the museum itself is a geofence. Several geofences can be combined into a common geofence, e.g., the workshops of the craftsmen trade street. A Web app has been developed that can be used to check if the geofences database has been successfully built. The application displays on a GPS map all geofences as well as the position of all visitors if there is already information in the visitor database (see Figure 5).

Figure 4. Application for creating geofencing database: a) Web app; b) Mobile app.

Figure 4. Application for creating geofencing database: a) Web app; b) Mobile app.

Figure 5. Web application for testing geofences database.

Figure 5. Web application for testing geofences database.

3.2.2. Exhibits database

In open-air museums, each geofence describes a building (e.g., workshop) or an open-air exhibit. In this example, the exhibits are the workshops and shops to which visitors have access. A database of exhibits is created through a CMS microservice. It provides a Web interface through which museum staff can describe exhibits, edit their properties, and delete information associated with an exhibit. For example, when entering a new exhibit, a form is displayed for the staff member to fill out (see Figure 6).

Figure 6. Exhibit Description Form.

Figure 6. Exhibit Description Form.

Fields in black must be initialized by the user, and those in gray can be initialized automatically. In this example, the information to be entered manually is the exhibit name, the exhibit type, the media files to be associated with the exhibit. For each selected media file, tags must be entered that are associated with the file. These tags are needed when filtering the multimedia content to return personalized content. The rest of the information can be obtained automatically using the Chatbot microservice. For museums with a very large number of exhibits, the possibility to train their own language model is provided. This is also necessary in cases where the museum is small and ChatGPT has not indexed information about it. The training of the model is implemented using the LlamaIndex data framework. LlamaIndex is a data framework designed for creating Large Language Model (LLM) applications. The available API allows LlamaIndex to be used for data ingestion and search. The framework offers connectors for training data using different file formats, for example PDF. It provides different ways to index the data so that this data can be easily used with LLM and an advanced interface to query the indexed data. We use gpt-3.5-turbo-0613 as the LLM. Input data should be prepared by museum curators in PDF format. The queries are implemented using the ChatGPT model when the resulting index file is set as the source. For museums with few exhibits, another strategy is possible. In this strategy, the detailed description of the exhibit is entered manually, and the rest of the form fields are automatically populated using appropriate ChatGPT prompts based on the detailed exhibit information. In this way, the data for all exhibits does not need to be indexed, but the functionality of the service is lower since ChatGPT cannot return responses that use information for multiple exhibits.

3.3. Mobile App

To access the service ExhibitXplorer, the visitor must install a mobile application for Android OS on their mobile device. The installation of the application is realized automatically by scanning a QR code that is positioned at the museum entrances. After the app is launched for first time, it registers to receive push notifications via Google FCM. Upon successful registration, an FCM access ID is returned. This unique string is also used as the visitor identifier. The ExhibitXplorer service does not require registration by names or e-mail address to ensure anonymity of visitor data that is collected.

The mobile app registers listeners for events from NFC, BLE beacons and GPS. These listeners allow activation of a program code when certain events occur, for example: the mobile device is near an NFC tag; the mobile device is in proximity to a BLE beacon; or new GPS coordinates are received. In terms of battery consumption of the mobile device, the critical hardware part is the GPS receiver. For this reason, the GPS receiver is only accessed when detects that the visitor is on the move. For this purpose, a human activity detector (HAR) is used to classify the visitor’s activity into one of the following categories: still, walking, running, and transportation. Data from BLE beacons and GPS are only analyzed if the HAR returns a walking category. The implementation of HAR is based on the analysis of accelerometer data.

To use the ExhibitXplorer service, the app does not need to be in focus as event listeners from BLE beacons and GPS are running in the background. When a visitor approaches an exhibit the Geofences microservice (via the Notifications microservice) sends a push notification. It notifies the visitor that they are near an exhibit (see Figure 7).

Figure 7. Push notification.

Figure 7. Push notification.

The visitor can open the notification or leave it without consequence. If the visitor opens the notification, the service returns only brief information about the exhibit, but begins proactively generating a personalized content. From the brief information, the visitor can find out exactly what the exhibit is, as well as use the chatbot to get more information. If the visitor chooses to receive additional information the service returns the already prepared personalized content.

When the notification is opened, a short content for the exhibit is generated. This content includes a name of the exhibit, a short text description and an image. If the exhibit is of interest to the visitor, they can get answers to their specific questions via the museum’s chatbot as well as request a full description of the exhibit depending on their current profile (see Figure 8).

Figure 8. Description of the exhibit: a) brief description; b) chatbot; c) full description (text); and d) full description (images).

Figure 8. Description of the exhibit: a) brief description; b) chatbot; c) full description (text); and d) full description (images).

The full description contains those types of information that the visitor prefers to receive. In this example, this is text, images, video, and audio files related to the exhibit. References to additional resources and links to the Internet sites are missing.

The mobile app is also used to implement explicit visitor profiling. For this purpose, they receive push notifications inviting them to answer one or more questions. Initially, it is checked how much time each visitor can spend browsing the exhibits in the museum. Depending on the answer, profiling module uses different strategies. If the visitor declares little time to spend in the museum, a survey form is sent to the visitor that contains all the questions related to explicit profiling. Otherwise, to make the visitor more likely to respond, the questions are sent in parts, each containing one to three questions. Figure 9 shows some of the questions sent to a museum visitor.

Figure 9. Explicit profiling.

Figure 9. Explicit profiling.

3.5. Interpretation of results

The number of participants in the test is 24. They were selected based on random sampling. The tests were conducted within one week. Each visitor who participated in the service testing completed an electronic survey after completing their visit within the museum. For each of the 10 questions on the survey, test participants are asked to give their rating from 1 to 5 on a Likert scale.

Figure 10 shows the results obtained after analyzing all surveys.

Figure 10. User’s experience with the service.

Figure 10. User’s experience with the service.

Most participants (92%) in the test felt that the service provided content that met their interests and expectations (Q1). 54% of respondents answered "strongly yes" and 38% answered "yes". Only 8% were undecided in their answer and answered neutral.

The service relies on delivering content about exhibits unobtrusively and only, when necessary, via push notifications. The results obtained for question 2 of the survey (Q2) are therefore very important - Were the notifications deployed at appropriate times and were they unobtrusive during your visit? As can be seen in Figure 10, most participants (67%) stated that the service provided content in an extremely unobtrusive manner and at the appropriate time (score 5), and 33% of them rated the service with a score of 4.

The user interface of the mobile app (Q3) scores is expectedly good. An equal number (50%) each of respondents gave a rating of 5 and a rating of 4. The simplified user interface of the mobile app, as well as the lack of need for settings to use the app, suggest similar results.

Responses to question 4, whether delivering content via push notifications improved the visitor experience at the museum (Q4), garnered 88% high marks (4 or 5). Only 12% of respondents said they could not judge.

Personalization of content contributed to improving knowledge of exhibits (Q5) for 58% of respondents (score 4). Another 37% responded with a score of 5. A small proportion of respondents (5%) said they could not judge.

Only 5% of respondents said they could not judge how relevant and useful the information provided by the museum chatbot was (Q6). Most respondents (54%) rated the chatbot a 4, with the remainder (41%) giving the highest rating a 5. These are very good ratings, and they suggest that the ChatGPT API can be successfully used for the purpose of delivering personalized content in museums. Similar are the responses to the question "How often did you use the museum’s chatbot during your visit (Q7). 29% of respondents used the chatbot very often (score 5), and 58% used it often (score 4). 13% gave a neutral response. There was no test participant who did not ask at least one question to the chatbot.

All respondents answered that they would recommend the service for providing personalized content to others (Q8).

Responses to question 9 (Q9) refuted our concerns that explicit profiling might fail if responses to survey questions were optional. 55% of respondents answered ’yes’ to this question and 42% answered ’strongly yes’. We assume that the reason for this is the way the survey questions are presented - in small portions, mostly after the visitor has viewed an exhibit.

Overall, respondents were satisfied (55%) or very satisfied (45%) with the service of providing personalized content through the museum’s mobile app (Q10).

All these results are very encouraging for the feasibility and usability of the developed service to provide personalized content to museum visitors.

Comparison of results from different age groups

An analysis of the survey responses by age group was conducted. In the first age group "Children and students" (6 to 20 years) the number of respondents was 5 (Mean=10.6, SD=5.54). In the second age group "Adults" (21 to 64 years old), 12 visitors participated (Mean=40.17, SD=12.90). In the third age group "Seniors" (over 64 years) 7 visitors participate (Mean=74.86, SD=6.31).

Figure 11 shows the scores of the respondents grouped by age groups.

Figure 11. Users’ experience in age groups.

Figure 11. Users’ experience in age groups.

The differences in estimates for question Q2 are also minimal. The highest score (4.80) was given by age group 1 participants and the lowest score (4.57) was given by age group 3 participants. This is to be expected as young people use push notifications from social networks daily.

When evaluating the user interface (Q3), the youngest and oldest gave slightly higher scores when compared to the scores of age group 2, 4.60 and 4.57 respectively. We assume that the intuitive interface, which displays the desired content without having to search for information through the mobile app menu, is equally convenient and usable for all age groups.

As expected, the youngest rated the effect of using push notifications to enhance their museum experience the highest (4.60). Visitors in age group 3 gave the lowest rating (4.29).

The scores for Q5 are very revealing. The difference in participants’ ratings across the three age groups was statistically insignificant. This suggests that the choice of personalized content delivery method was equally well rated by all age groups.

The highest score (4.80) for the use of the museum chatbot (Q6) was given by age group 1. In this age group, the number of questions asked through the chatbot for each exhibit ranged from 3 to 9. The scores of the other two age groups are also high, 4.25 for group 2 and 4.27 for group 3. Only one visitor from age group 3 asked fewer than 5 questions about all the exhibits they viewed. This can be seen from the ratings of question Q7 on how often visitors used the museum’s chatbot. All visitors from age group 1 declared that they used the chatbot very often.

The most satisfied with the service (Q10) are visitors from age group 1. Their average score was 4.80. The scores of visitors from the other two age groups were relatively high at 4.33 and 4.43 respectively.

4. Discussion

5. Conclusions

References

1. Cristobal-Fransi, E.; Ramón-Cardona, J.; Daries, N.; Serra-Cantallops, A. Museums in the digital age: An analysis of online communication and the use of E-commerce. J. Comput. Cult. Herit. 2021, 14. [Google Scholar] [CrossRef]
2. Ye, B.H.; Ye, H.; Law, R. Systematic review of smart tourism research. Sustain. 2020, 12. [Google Scholar] [CrossRef]
3. Robinson, A. The future of museums and a history of ignorance: Books in brief. Nature 2023. [Google Scholar] [CrossRef] [PubMed]
4. Ayala, I.; Cuenca-Amigo, M.; Cuenca, J. The Future of Museums. An Analysis from the Visitors’ Perspective in the Spanish Context. J. Arts Manag. Law Soc. 2021, 51, 171–187. [Google Scholar] [CrossRef]
5. Elizabeth Merritt, It’s Personal: one size does not fit all, in Trends Watch 2015, American Alliance of Museums, https://www.aam-us.org/2015/05/01/its-personal-one-size-does-not-fit-all/, (accessed on 1 August 2023). (accessed on 1 August 2023).
6. Kosmopoulos, D.; Styliaras, G. A survey on developing personalized content services in museums. Pervasive Mob. Comput. 2018, 47, 54–77. [Google Scholar] [CrossRef]
7. Fernández-Hernández, R.; Vacas-Guerrero, T.; García-Muiña, F.E. Online reputation and user engagement as strategic resources of museums. Museum Manag. Curatorsh. 2021, 36, 553–568. [Google Scholar] [CrossRef]
8. Argyros, A.; Kosmopoulos, D. The MuseLearn platform: Personalized content for museum visitors assisted by vision-based recognition and 3D pose estimation of exhibits. In Proceedings of the IFIP Advances in Information and Communication Technology; 2020; Vol. 583 IFIP, pp. 439–451.
9. Hijazi, A.N.; Baharin, H. The Effectiveness of Digital Technologies Used for the Visitor’s Experience in Digital Museums. A Systematic Literature Review from the Last Two Decades. Int. J. Interact. Mob. Technol. 2022, 16, 142–159. [Google Scholar] [CrossRef]
10. King, E.; Smith, M.P.; Wilson, P.F.; Stott, J.; Williams, M.A. Creating Meaningful Museums: A Model for Museum Exhibition User Experience. Visitor Studies 2023, 26, 59–81. [Google Scholar] [CrossRef]
11. Eke, C.I.; Norman, A.A.; Shuib, L.; Nweke, H.F. A Survey of User Profiling: State-of-the-Art, Challenges, and Solutions. IEEE Access 2019, 7, 144907–144924. [Google Scholar] [CrossRef]
12. Vrettakis, E.; Katifori, A.; Kyriakidi, M.; Koukouli, M.; Boile, M.; Glenis, A.; Petousi, D.; Vayanou, M.; Ioannidis, Y. Personalization in Digital Ecomuseums: The Case of Pros-Eleusis. Applied Sciences (Switzerland) 2023, 13. [Google Scholar] [CrossRef]
13. Antoniou, A.; Katifori, A.; Roussou, M.; Vayanou, M.; Karvounis, M.; Kyriakidi, M.; Pujol-Tost, L. Capturing the Visitor Profile for a Personalized Mobile Museum Experience: An Indirect Approach. In Proceedings of the CEUR Workshop Proceedings; 2016; Vol. 1618.
14. Norouzi, R.; Baziyad, H.; Aknondzadeh Noghabi, E.; Albadvi, A. Developing Tourism Users’ Profiles with Data-Driven Explicit Information. Math. Probl. Eng. 2022, 2022. [Google Scholar] [CrossRef]
15. Chu, W.; Park, S.T. Personalized Recommendation on Dynamic Content Using Predictive Bilinear Models. In Proceedings of the WWW’09 - Proceedings of the 18th International World Wide Web Conference; pp. 2009691–700.
16. Sparacino, F. The Museum Wearable: Real-Time Sensor-Driven Understanding of Visitors’ Interests for Personalized Visually-Augmented Museum Experiences. In Proceedings of the In: Proceedings of Museums and the Web (MW2002); pp. 200217–20.
17. Lewalter, D.; Phelan, S.; Geyer, C.; Specht, I.; Grüninger, R.; Schnotz, W. Investigating Visitor Profiles as a Valuable Addition to Museum Research. International Journal of Science Education, Part B: Communication and Public Engagement 2015, 5, 357–374. [Google Scholar] [CrossRef]
18. Sepe, F.; Marzullo, M. Making Smarter Museums Through New Technologies. In Handbook of Research on Museum Management in the Digital Era; 2022; pp. 75–98.
19. Karayazi, S.S.; Dane, G.; de Vries, B. Utilizing urban geospatial data to understand heritage attractiveness in Amsterdam. ISPRS Int. J. Geo-Information 2021, 10. [Google Scholar] [CrossRef]
20. Spachos, P.; Plataniotis, K.N. BLE Beacons for Indoor Positioning at an Interactive IoT-Based Smart Museum. IEEE Syst. J. 2020, 14, 3483–3493. [Google Scholar] [CrossRef]
21. Barsocchi, P.; Girolami, M.; La Rosa, D. Detecting proximity with bluetooth low energy beacons for cultural heritage. Sensors 2021, 21. [Google Scholar] [CrossRef]
22. Spachos, P.; Plataniotis, K.N. BLE Beacons for Indoor Positioning at an Interactive IoT-Based Smart Museum. IEEE Systems Journal 2020, 14, 3483–3493. [Google Scholar] [CrossRef]
23. Wohllebe, A.; Hübner, D.S.; Radtke, U.; Podruzsik, S. Mobile Apps in Retail: Effect of Push Notification Frequency on App User Behavior. Innovative Marketing 2021, 17, 102–111. [Google Scholar] [CrossRef]
24. Trunfio, M.; Lucia, M. Della; Campana, S.; Magnelli, A. Innovating the cultural heritage museum service model through virtual reality and augmented reality: the effects on the overall visitor experience and satisfaction. J. Herit. Tour. 2022, 17, 1–19. [Google Scholar] [CrossRef]
25. Verde, D.; Romero, L.; Faria, P.M.; Paiva, S. Architecture for Museums Location-Based Content Delivery using Augmented Reality and Beacons. In Proceedings of the ISC2 2022 - 8th IEEE International Smart Cities Conference; 2022. [Google Scholar]
26. Meng, Y.; Chu, M.Y.; Chiu, D.K.W. The impact of COVID-19 on museums in the digital era: Practices and challenges in Hong Kong. Libr. Hi Tech 2022. [Google Scholar] [CrossRef]
27. Giannini, T.; Bowen, J.P. Museums and Digital Culture: From Reality to Digitality in the Age of COVID-19. Heritage 2022, 5, 192–214. [Google Scholar] [CrossRef]
28. Choi, B.; Kim, J. Changes and Challenges in Museum Management after the COVID-19 Pandemic. Journal of Open Innovation: Technology, Market, and Complexity 2021, 7. [Google Scholar] [CrossRef]
29. Wang, B. Digital Design of Smart Museum Based on Artificial Intelligence. Mob. Inf. Syst. 2021, 2021. [Google Scholar] [CrossRef]
30. Pisoni, G.; Díaz-Rodríguez, N.; Gijlers, H.; Tonolli, L. Human-centred artificial intelligence for designing accessible cultural heritage. Appl. Sci. 2021, 11, 1–30. [Google Scholar] [CrossRef]
31. Gaia, G.; Boiano, S.; Borda, A. Engaging Museum Visitors with AI: The Case of Chatbots. In Springer Series on Cultural Computing; 2019; pp. 309–329.
32. OpenAi ChatGPT: Optimizing Language Models for Dialogue. Available online: https://online-chatgpt.com/ (accessed on 1 August 2023).
33. Elgarf, M.; Peters, C. CreativeBot: A Creative Storyteller Agent Developed by Leveraging Pre-Trained Language Models.; 2022; pp. 13438–13444.
34. Commission joins forces with Member States to launch a Collaborative Cloud for Europe’s cultural heritage. Available online: https://ec.europa.eu/commission/presscorner/detail/en/IP_22_3855, (accessed on 1 August 2023).
35. Martin, D.; Alzua, A.; Lamsfus, C. A Contextual Geofencing Mobile Tourism Service. In Information and Communication Technologies in Tourism 2011; 2011; pp. 191–202.
36. Elizabeth Merritt, For your radar: The Metaverse and Web 3.0, in Trends Watch 2023, American Alliance of Museums, Center for the Future of Museums, https://www.aam-us.org/2023/01/13/for-your-radar-the-metaverse-and-web-3-0/, (accessed on 1 August 2023).