Transient Narrative Networks and Information Landscapes for Enhancing Human Understanding Remi van Trijp1 1 SONY Computer Science Laboratories, 6, rue Amyot, 75005 Paris, France Abstract This demonstration paper presents work-in-progress on a system that explores the use of interactive narratives and information landscapes for helping people to make sense of complex topics, especially those who struggle with data overload, information anxiety and other illnesses of the Digital Age. The current prototype allows people to explore articles from the English Wikipedia in three ways: a news-like web interface for careful reading, a ‘transient narrative network’ for coherence-building which unfolds based on user interaction, and an information landscape for perceiving which information exists outside of the user’s current purview. 1. Introduction 2. Three Ways of Exploring The explosion of data in the Digital Age has so far Information mainly been treated as a technological problem that The first prototype of the system allows users to can be solved through more sophisticated data man- explore information from wikipedia, which has been agement techniques. However, an equally important selected as the prototype’s knowledge base because challenge is how to manage the impact of data ex- of its scale, free access, linguistic diversity, and plosion on people’s well-being. More specifically, evolving nature. The prototype currently uses the problems such as data overload [1] or information English wikipedia but localization efforts are made anxiety [2] make it increasingly difficult for people to extend the interface to access articles from other to make sense of the events that they experience languages as well. or perceive on a daily basis. Moreover, data and The following three subsections explain the sys- communication overload also have disastrous effects tem’s three ways of exploring information from the on interpersonal understanding, leading to more perspective of a fictitious person called Emily, a hostile interactions especially on social media [3]. young British woman who is very concerned about One of the most promising avenues for solving the war in Ukraine and who wants to learn more these problems is to use human-centric artificial about the major entities that are involved. intelligence that can help to enhance the human understanding process. This demonstration paper therefore presents work-in-progress on a system that 2.1. Careful Reading explores how interactive narratives and information It is March 24, 2022. Emily hears on the news that landscapes could be used for this purpose. The core British Prime Minister Boris Johnson is speaking of the prototype is implemented using the Babel at a press conference at the NATO headquarters in Toolkit [4], an open-source software platform that Brussels, where he accuses Russia of committing war includes an interactive web interface and cognitive crimes. Emily doesn’t remember well what NATO language technologies such as Fluid Construction is, or why the Prime Minister would go there, so Grammar [5], which will be used in future versions she decides to look for more information. of the prototype for deep linguistic analysis. The Upon starting the system, Emily is greeted by a core system is designed to be highly modular and ex- traditional-looking, news-like webpage as illustrated tensible through a RESTful architecture that allows in Figure 1. She then types “NATO” in the search the system to request and integrate information bar in the upper right corner, and a wikipedia sum- from external components. mary about NATO now appears under the header “focus” in the top left, with an option to read the IJCAI 2022: Workshop on semantic techniques for full article. Below the summary she finds links to narrative-based understanding, July 24, 2022, Vienna, Aus- related pages, while on the right she sees three more tria prominent suggestions under the header “related $ remi.vantrijp@sony.com (R. v. Trijp) © 2022 Copyright for this paper by its authors. Use permitted under entities.” Creative Commons License Attribution 4.0 International (CC BY 4.0). As their names suggest, both the “related pages” CEUR Workshop Proceedings http://ceur-ws.org ISSN 1613-0073 CEUR Workshop Proceedings (CEUR-WS.org) 43 Figure 1: The system’s web interface starts with a news-like homepage that invites careful reading, with a search bar in the top right corner that allows the user to query wikipedia. On the left, a summary of the corresponding hit is shown, with the option to read the full article. On the right, the three most similar wikipedia entities are shown based on entity embeddings. Below, related pages are shown that are obtained through the Wikimedia REST API. and “related entities” section aim to offer more tion sources. AI researchers have therefore become pieces of information that may be relevant to increasingly interested in incorporating the concept Emily’s quest for knowledge. They only differ in of a narrative into the design of human-centric AI where they are derived from: the related pages systems [9, 10], especially now that information is are obtained directly from Wikipedia’s linked data more scattered than ever across billions of connected through the Wikimedia REST API, while the re- devices. lated entities were suggested by a model of embed- Following studies in narratology [11], a narrative dings of entities (vector representations) trained is assumed to have a three-layered structure: with the Wikipedia2Vec tool [6], which not only learns from wikipedia’s link graph model but which • The fabula (or story) is the set of facts and also uses a word-based skip-gram model and an an- events; similar to the concept of ground truth. chor context model (using the neighbouring words In the current prototype, the set of wikipedia of a hyperlink that points to an entity as context). articles is assumed to represent the fabula. • The plot (or syuzhet) is a structure that ar- ranges the relevant facts and events into a 2.2. Transient Narrative Network causal chain or causal network. Emily clicks on a number of related pages, as she • The narration (or narrative presentation) usually does when browsing the web, but she soon concerns how the narrative is presented. loses track of which topics she has already covered The prototype has been observing Emily’s inter- or how they relate to each other. She therefore actions with the system, and keeps track of which switches to the narrative network view using the pages she visited (her personal plot) and which navigation bar on top. related pages and entities would make possible ex- Narratives are widely accepted to be a key el- cursions or new pathways for exploration (parts of ement of the human sense-making process [7, 8], the fabula that may become relevant).1 particularly as a vehicle for creating coherence from otherwise fragmented, disparate and noisy informa- 1 All the while respecting her privacy: all of her information is stored locally on her own computer. 44 Figure 2: The system builds transient narrative networks based on how the user interacts with the system. The user can traverse and manipulate the network, and switch back to careful reading mode to learn more about the currently selected node. Here, the larger nodes and the red directed path illustrates the path followed by the user, while smaller nodes and grey edges provide potential pathways. The resulting narrative, shown in Figure 2, is op- on something? erationalized as a transient narrative network [12] using vis.js (a dynamic browser-based visualization library). A transient narrative network is a graph that dynamically changes as new nodes and edges are added when there is additional input or when more information is requested from other knowl- edge sources. Here, the graph shows which nodes represent pages that Emily visited with directed links indicating the order in which she traversed the information space. Additional nodes and edges are automatically added from wikipedia’s linked graph model and shown as possible paths for explo- ration, with more or less prominence depending on where Emily is currently situated in the network (i.e. which node is selected). Emily can now choose Figure 3: Information landscapes allow users to get to continue exploring the network and expanding it a glimpse of the available information outside of their by selecting nodes of interest, or she can return to purview. Here, the 100 most similar entities to “NATO” the careful reading view to learn more details about are shown. a topic. She can also manually add, remove and modify her own nodes and edges in other to further She therefore switches to a third page, which of- personalize or complete the graph. fers a data visualization of all Wikipedia entities, where she can see at which location she currently 2.3. Information Landscapes finds herself. She also gets a glimpse about which information is out there, both close by and remotely While Emily now has a more coherent answer to located. In the current prototype, the only visualiza- her initial question, she is still somewhat distrustful tion that is available is a visualization of the embed- about the system. How reliable are the system’s dings of wikipedia entities by the Wikipedia2Vec suggestions? And what information is out there? tool [6], illustrated in Figure 3, which is not yet The narrative network is anchored to her own per- connected to the rest of the system. spective so she can only see the so-called “adjacent- possible” nodes [13, 14]. What if she is missing out 45 3. Conclusion and Future Work to my colleagues Inès Blin and Martina Galletti. Finally, I wish to thank Vittorio Loreto and Hi- This demonstration paper introduced the first pro- roaki Kitano for creating such a superb working totype of a system under development that aims environment. to enhance human understanding through human- centric artificial intelligence. More specifically, it combines careful reading with narrative networks References (for coherence building) and information landscapes (for purposefully navigating the information space) [1] A. Toffler, Future Shock, Random House, New through a web-based interface. York, 1970. Co-authored by Heidi Toffler (un- Future work is planned in all three areas. For credited). the careful reading page, additional components [2] R. S. Wurman, Information Anxiety: Towards need to be integrated that support different read- Understanding, Doubleday, New York, 1989. ing and comprehension monitoring strategies [15]. [3] M. Fan, Y. Huang, S. A. Qalati, S. M. M. These may include a combination of neurostatistical Shah, D. Ostic, Z. Pu, Effects of informa- models for keyword extraction (for quickly scanning tion overload, communication overload, and texts) or entity recognition and linking (for improv- inequality on digital distrust: A cyber-violence ing the recommendations); and symbolic models for behavior mechanism, Frontiers in Psychol- extracting more detailed semantic frames [16, 17]. ogy 12 (2021). URL: https://www.frontiersin. To enhance the narrative networks, additional org/article/10.3389/fpsyg.2021.643981. doi:10. knowledge sources need to be integrated such as 3389/fpsyg.2021.643981. knowledge graphs (e.g. WikiData). Such knowledge [4] M. Loetzsch, P. Wellens, J. De Beule, J. Bleys, sources can also be employed for offering different R. van Trijp, The Babel2 Manual, Technical narrations of the same plot. Promising work already Report AI-Memo 01-08, AI-Lab VUB, Brussels, exists on how knowledge graphs can be exploited for 2008. presenting events on a chronological timeline [10], [5] L. Steels (Ed.), Design Patterns in Fluid Con- which may further help users to build a coherent pic- struction Grammar, volume 11 of Construc- ture about topics that they might otherwise explore tional Approaches to Language, John Ben- in a more random fashion. Narrative networks also jamins, Amsterdam, 2011. URL: https://doi. need to be stored in a Personal Dynamic Network org/10.1075/cal.11. [9]. [6] I. Yamada, A. Asai, J. Sakuma, H. Shindo, Finally, more interactive visualization methods H. Takeda, Y. Takefuji, Y. Matsumoto, are needed that provide users with intuitive ways Wikipedia2Vec: An efficient toolkit for learn- to understand the information space. This atlas of ing and visualizing the embeddings of words semantic maps then needs to be connected to the and entities from Wikipedia, in: Proceedings of core system so that the system can propose more the 2020 Conference on Empirical Methods in interesting routes for navigating information. Natural Language Processing: System Demon- strations, Association for Computational Lin- guistics, 2020, pp. 23–30. 4. Ethical Statement [7] J. Bruner, The Narrative Construction of Reality, Critical Inquiry 18 (1991) 1–21. There are no ethical issues. URL: http://www.jstor.org/stable/1343711, publisher: The University of Chicago Press. [8] B. Boyd, On the Origin of Stories: Evo- Acknowledgments lution, Cognition, and Fiction, Harvard The work reported in this paper was carried out in University Press, Boston MA, 2009. URL: the context of MUHAI (Meaning and Understand- http://www.jstor.org/stable/j.ctvjf9xvk. ing in Human-Centric AI – https://muhai.org/), doi:10.2307/j.ctvjf9xvk. a project funded by the European Union’s Hori- [9] L. Steels, Personal dynamic memories are zon 2020 research and innovation programme under necessary to deal with meaning and under- grant No 951846. I wish to thank the workshop or- standing in human-centric AI, in: A. Saf- ganizers Lise Stork, Katrien Beuls, and Luc Steels; fiotti, L. Serafini, P. Lukowicz (Eds.), Pro- the anonymous reviewers; and the MUHAI partners ceedings of the First International Workshop for their invaluable feedback. Many thanks also go on New Foundations for Human-Centered AI (NeHuAI) co-located with 24th European Con- 46 ference on Artificial Intelligence (ECAI 2020), CEUR Workshop Proceedings, Santiago de Compostela, 2020. URL: http://ceur-ws.org/ Vol-2659/steels.pdf. [10] I. Blin, Building narrative structures from knowledge graphs, in: P. Groth, M.-E. Vi- dal, F. Suchanek, P. Szekley, P. Kapanipathi, C. Pesquita, H. Skaf-Molli, M. Tamper (Eds.), Extended Semantic Web Conference (ESWC) 2022, Springer Nature, Cham, 2022. [11] M. Bal, Narratology: Introduction to the The- ory of Narrative, University of Toronto Press, Toronto, 2017. Fourth edition, first ed. 1985. [12] L. Steels, Narrative art interpretation, in: L. Steels (Ed.), MUHAI Deliverable D1.1 White Paper. Foundations for Incorporating Meaning and Understanding in Human-Centric AI, The MUHAI Consortium, Bremen, 2022. [13] S. Kauffman, At Home in the Universe: The Search for Laws of Self-Organization and Com- plexity, Oxford University Press, Oxford, 1995. [14] B. Monechi, . A. Ruiz-Serrano, F. Tria, V. Loreto, Waves of novelties in the expan- sion into the adjacent possible, PloS one 12 (2017) e0179303–e0179303. URL: https: //pubmed.ncbi.nlm.nih.gov/28594909. doi:10. 1371/journal.pone.0179303, publisher: Pub- lic Library of Science. [15] Y.-F. Yang, Reading Strategies or Com- prehension Monitoring Strategies?, Reading Psychology 27 (2006) 313–343. URL: https: //doi.org/10.1080/02702710600846852. doi:10. 1080/02702710600846852, publisher: Rout- ledge. [16] V. Micelli, R. van Trijp, J. De Beule, Framing Fluid Construction Grammar, in: N. Taatgen, H. van Rijn (Eds.), Proceedings of the 31th Annual Conference of the Cognitive Science Society, Cognitive Science Society, 2009, pp. 3023–3027. [17] K. Beuls, P. Van Eecke, V. S. Cangalovic, A computational construction grammar approach to semantic frame extraction, Linguistics Van- guard 7 (2021). 47