Information systems enable users to represent, manipulate, and access data efectively. Currently, the development and availability of extensive knowledge bases receive a lot of attention for supporting a variety of intelligent applications like smart assistants, question answering, and knowledge discovery. But building such knowledge bases generally means breaking down information into manageable pieces (e.g., in the form of triples in RDF) that tend to lose connections and extraction contexts. In contrast, humans typically share and exchange knowledge by interweaving diferent pieces into narratives that are plausible and easy to grasp. In this position paper, we summarize the main research directions of narratives in computer sciences. Moreover, we propose basic design principles that narrative information systems should consider in practice. In particular, we take a closer look at narrative representations, possible bindings between narratives and real-world data, the context-compatibility of information, and finally a narrative's plausibility.
sentation of (mostly entity-centric) knowledge [
Our research is focused on the exchange of knowl- also knowledge bases that capture knowledge about other edge through meaningful patterns, i.e., we understand a types of things. In particular, there is an increasing internarrative as a possible form to share knowledge. In this est in knowledge bases focusing on events, e.g., Knowlypaper, we focus on the exchange and transfer of knowl- wood captures information about activities mined from edge through plausible narratives. Therefore, we first Hollywood narratives [4], and the EventKG captures all examine how established systems represent knowledge kinds of information about a variety of events [5]. To and discuss their limitations. We continue with a brief eficiently build such collections, methods have been prointroduction to narratives in computer science and move posed to extract events from textual sources, e.g., events forward to our position: Narratives need to be handled from news [6], temporal facts and events [7], or events as first-class citizens in information systems. from Wikipedia [8]. For a good overview on event extraction, see [9].
Knowledge Bases. So how do current systems rep- In brief, the resulting knowledge bases store factual resent knowledge? In contrast to classical relational knowledge about entities or events, e.g., an entity’s name, databases or warehouses for well-structured data, the age, and type or some event’s date, location, and particidevelopment of the Resource Description Framework pants. Moreover, it also reflects typed relations between (RDF) within the standardization eforts of the Semantic pairs of entities and events. So is this knowledge repreWeb has allowed a structured and simple-to-use repre- sentation suficient for information systems? From our perspective, knowledge bases indeed efectively repreIJCAI-ECAI 2022 workshop: Semantic Techniques for Narrative-based sent and allow access to factual knowledge, in the sense Understanding, July 24, 2022, Vienna, Austria of universally valid statements. In strict correspondence $ k00ro0l0l-@00ifis0.1cs-9.t8u8-7b-s9.d2e76(H(H.K.rKorlol)l;l)b;a0l0k0e0@-0ifis0.0c2s.-t5u4-4b3s-.d1e21(W5(.WB.alBkael)ke) to the real world, such statements can consistently co© 2022 Copyright for this paper by its authors. Use permitted under Creative exist within a knowledge base. But what happens to CPWrEooUrckReshdoinpgs IhStpN:/c1e6u1r3-w-0s.o7r3g CCoEmmUoRns LWiceonsrekAstthribouptionP4r.0oIncteerenadtiionnagl s(CC(CBYE4U.0)R.-WS.org) knowledge that is not universally valid? For instance, Bindings
Substory: “Invention”
developed Edgar Codd
Relational Model
CACM published in prerequisite for
Implementation of a
relational DB |timeframe=1980s
|timeframe=1980s Oracle participant turns into
Successful
tech-company
Context-Compatibility
some statements may either reflect subjective opinions and were collected in a respective schema database [14].
or they may be only valid in certain contexts or under Previous works generally assume an order and relation
certain conditions or even only with a certain probability. of events in texts that needs to be mined. Tang et al.
proThe resulting knowledge containing also such facts base posed neural methods to learn the respective relations
might easily become inconsistent or contradictory. Thus, between events from text directly [15]. Another approach
while knowledge bases are efective for factual knowl- proposed the mining of so-called plot graphs [
A good overview of the limitation of today’s knowl- optional, and mutually exclusive events. In brief, a plot
edge bases is given in [
Narrative Applications. Various applications have been proposed to utilize narratives in information systems. Narrative schemas and plot graphs are used for narrative planning, i.e., to write specific stories by computers [17, 18]. Another application is the alignment of narratives and movie scripts, i.e., linking narrative components (e.g., fall in love) to performed actions (e.g., A kisses B) in movies [19].
The digital library project Europeana utilizes narratives to compose events involving artists into timelines [20]. The narratives are then shown to users as Narratives. Narratives have been studied in computer overviews of certain artists, i.e., to tell a story about them. science for a while. Sequences of events or actions to The benefits of such a representation have already been model human behavior and situation awareness go back studied [21]. Moreover, knowledge bases like Wikidata to Schank and Abelson [11] in the form of scripts. [12] can be utilized to populate narratives with additional proposed the representation of narratives as a sequence factual information [22]. of events: Narrative schemas have been introduced to Finally, Shahaf et al. propose to connect the dots berepresent a set of events and their semantic roles [13] tween news articles in so-called metro maps [23]. Metro maps consist of articles as stations and topics as metro diferent kinds of relationships: factual and narrative lines. Users may then follow a line (topic) across diferent relationships. Like knowledge bases, factual relationarticles. In addition to news items, Shahaf et al. have ships describe properties involving entities and literals, demonstrated how scientific articles can be arranged as e.g., that Edgar Codd developed the relational model. These metro maps of sciences, too [24]. In summary, such Metro relationships can be put between entities and literals. We maps allow users to understand better the spatial, causal, compose the actual progress of a narrative by a set of or temporal structure of available resources. We under- narrative relationships, e.g., the implementation of a relastand such maps as a possible application of narratives tional database system by Oracle that has since turned into aligning similar textual sources in a meaningful way. a successful tech-company. In brief, narrative relations feature special, non-factual labels. They can be placed between events or between events and entities, but not 2. Narratives and Plausibility between entities. In this way, narrative can represent static knowledge and interweave such facts into dynamic Our main objective is to integrate narratives in infor- progress (e.g., some causal structure, some temporal flow, mation systems to allow a meaningful representation etc.). of knowledge. Therefore, we discuss four central de- The last thing a narrative representation must consign principles that narrative information systems must sider is nesting. With nesting, also called recursion or consider: narrative representation, bindings, context- induction, we refer to using narratives inside narratives, compatibility, and plausibility. e.g., a single narrative leads to another narrative. Nesting
When is a narrative meaningful? While we agree supports recursive elements as components, known from that meaningfulness is strongly connected to the actual spoken languages [26]. In our example, the development structure of a narrative, our research is focused on the of the relational model forms a narrative on its own that plausibility of narratives. We argue that plausible nar- is, in turn, the prerequisite for the actual implementation. ratives have a meaning. Consider our example in Fig. 1: While we agree that nesting makes the representation The development of the relational model by Edgar Codd more challenging, especially concerning precise semanwas a prerequisite for the implementation of a relational tics, we argue that nesting allows a higher expressiveness database system by Oracle. Oracle has then turned into of narratives, e.g., the occurrence of narrative a’s events a successful tech-company. The narrative is considered together finally lead to the whole narrative b. plausible because of two reasons: First, all of its parts have evidence in knowledge repositories. Second, the timeframe of the implementation matches the timeframe of Oracle’s stock value increase. With that in mind, we ifrst argue on narrative representations and then continue with key concepts to assess the plausibility of a narrative.
Narrative Representation. We understand a
narrative as a structure that describes some progress, e.g., of
temporal nature (one event might follow another event),
of conditional nature (one event might depend on the
other event being observed before), or of causal nature
(one event leads to another event). In previous work, we
designed narratives as logical overlays on top of
knowledge repositories [
Components of a narrative are entities (relevant things of interest, concepts, etc.), events (something which happens, some observed state, a change of some state, etc.) and literals (values). We made the distinction between entities and events to distinguish between rather static things from the real world and dynamic happenings that have a temporal dimension or represent some state or state change.
A narrative then puts these components in relation to each other. Therefore, we distinguish between two Narrative Bindings. For now, narratives are rather artificial structures. However, the question remains, how can we estimate whether the narrative is plausible? We argue that evidence for a narrative is one way to go: Evidence allows us to estimate if a narrative is grounded by some real-world data. If a structure is based on real-world data, the narrative itself is supported and has thus evidence. Such a grounding leads to a better understanding of whether the narrative is plausible.
The next step for a plausibility assessment is thus to
connect narratives with knowledge repositories. In our
understanding, a knowledge repository can be any
collection of knowledge, be it a collection of articles, a
structured database, a data set repository, etc. We call
the connection between a narrative’s relationship and a
knowledge repository a binding [
Definition 1. We call a narrative plausible, if the following conditions hold: 1. There exists a set of bindings NB that bind each relationship of against real-world data. 2. This set of bindings NB must be context-compatible, i.e., there is at least a single context to which all bindings agree (are valid in).
Context-Compatibility. Briefly speaking, bindings allow us to find evidence for a narrative. But can we just compute bindings for each relationship separately to ground the whole narrative? While some pieces of information are universally applicable and easy to connect to other pieces, such as the birth date of some person, there may be pieces that are only true within specific semantic settings. For instance, the capital of some country may change over time and thus should only be connected to pieces valid in the same time frame. Thus, when extracting pieces of information from knowledge repositories, it is necessary to consider potential contexts. Contexts are given by constraints on environment variables that describe under which condition a certain piece of information is valid.1
Narrative information systems must consider contexts to support a valid information fusion of pieces, i.e., to fuse only pieces that are valid under the same context conditions. In our example, the turning into a successful tech-company should be valid in the same timeframe as the implementation of the actual relational database. We call bindings context-compatible if they bind relation
processing [32]. The central idea is that if we can bind the whole narrative structure, and all bindings are also context-compatible, then there is decisive evidence that the narrative is plausible – again in the sense of having evidence.
However, although we defined narrative plausibility, there still remain challenges for a practical application.
1Please note that for the scope of this paper, we are abstracting We proposed four major dimensions that influence the from the actual truth or correctness of each piece of information, overall plausibility of a narrative [33]: which may rather reflect on the respective repository’s trustworthiness than provide a context in the sense of validity.
data management has become more apparent than ever.
Types of Evidence: Are we just looking for direct bind- But connecting published data sets and claims from sciings? What about counter-examples, i.e., state- entific publications remains challenging. If we could link ments that contradict a narrative’s relationship? both, we, on the one hand, could explain what a data set tells, and on the other hand, verify claims of new Confidence of Bindings: Each binding should be con- publications by existing data sets [28]. nected to confidence, i.e., a score that describes how certain we are that the relationship can be bound against some knowledge repository. Therefore, confidence is related to two properties: the trustworthiness of knowledge repositories and the quality of the binding process (e.g., the confidence of a retrieval or NLP method).
gumentation focused on the exchange of knowledge through narratives. How a narrative, or its events, are perceived by an audience also plays a central role today (think about fake news and social media perception here). Attributions and roles could further extend the narrative representation for richer semantics to better describe these events [34, 35].
Narrative Structure: A slight change of a narrative, Narrative Information Access. Narrative informae.g., adding, editing, or leaving out some compo- tion access allows user to formulate their information nent, can afect the result of a plausibility assess- need as a narrative. Then query processing takes place ment, i.e., diferent narratives might end up in to make the narrative plausible, i.e., finding contextdiferent assessments (plausible/not plausible). compatible bindings that cover the whole narrative. We already demonstrated the benefits of narrative informaValidation Approach: Which knowledge repositories, tion access to digital libraries in [32, 29]. and especially, which methods are used to compute the bindings? What do the methods guarantee? And thus, what could a user then expect?
that rules all purposes. Nevertheless, we have seen decisive properties that should be considered. From our perspective, a narrative information system must thus consider:
tion systems promise to support a wide range of novel and exciting applications to support human-centered Narrative Representation that makes a distinction workflows, e.g., by accessing and explaining knowledge between factual knowledge (e.g., entities and their through plausible narratives in the sense of storytelling. properties) and dynamic progress (e.g., causation Knowledge shared in this way is bound to allow for a and temporality). better and easier understanding of how a piece of knowledge is placed in the respective domain. And of course, Bindings that connect artificial narrative structures such models could even be extended further, for instance, with real-world data and events to give evidence by adding attributions also to support subjective aspects for the narrative. of knowledge in the sense of opinions, reflect special user intents in how knowledge can or should be used, or Context-Compatibility that embeds information express emotional stances towards entities, events, and pieces into contexts and defines rules when temporal or causal developments. contexts are compatible to support a valid In summary, the central argument of this paper is that information fusion. in order to build efective narrative information systems, Plausibility of narratives by binding the whole narra- we need to treat narratives as first-class citizens within tive with context-compatible bindings. these systems. Arguably there might not be a single model that applies perfectly to all use cases. Nevertheless,
So, why should we implement a narrative information there are some key properties and essential problems that system? In contrast to pure knowledge bases, narra- every narrative information system needs to consider, tive information systems allow us to interweave factual such as extraction methods, representational issues, or knowledge into plausible patterns. This way, the knowl- the possibility and validity of information fusion. This edge is represented in a pattern like humans would ex- paper opens up the relevant design dimensions and points change their thoughts. Reaching this target then allows to some early-stage solutions. a set of useful applications.
pp. 503–512. doi:10.1145/2911451.2911529. [9] W. Xiang, B. Wang, A Survey of Event Extraction
From Text, IEEE Access 7 (2019) 173111–173137.
doi:10.1109/ACCESS.2019.2956831.
[
42nd European Conference on Information Retrieval, Text2Story@ECIR 2020, Lisbon, Portugal, April 14th, 2020 [online only], volume 2593 of CEUR Workshop Proceedings, CEUR-WS.org, 2020, pp. 95– 104. URL: http://ceur-ws.org/Vol-2593/paper12.pdf. [11] R. Schank, R. Abelson, Scripts, Plans, Goals, and
Understanding, Routledge, 1977.
[12] N. Chambers, D. Jurafsky, Unsupervised
Learn[1] J. László, The Science of Stories: An Introduction ing of Narrative Event Chains, in: ACL 2008,
Proto Narrative Psychology, Routledge, 2008. doi:10. ceedings of the 46th Annual Meeting of the
As4324/9780203894934. sociation for Computational Linguistics, June
15[
W3C recommendation 10 (2004) 6. for Computer Linguistics, 2008, pp. 789–797. URL: [3] D. Vrandecic, M. Krötzsch, Wikidata: a free collab- https://aclanthology.org/P08-1090/. orative knowledgebase, Commun. ACM 57 (2014) [13] N. Chambers, D. Jurafsky, Unsupervised Learning of Narrative Schemas and their Participants, in: ACL 2009, Proceedings of the 47th Annual Meeting of the Association for Computational Linguistics and the 4th International Joint Conference on Natural Language Processing of the AFNLP, 2-7 August 2009, Singapore, The Association for Computer Linguistics, 2009, pp. 602–610. URL: https: //aclanthology.org/P09-1068/. [14] N. Chambers, D. Jurafsky, A Database of Narrative Schemas, in: Proceedings of the International Conference on Language Resources and Evaluation, LREC 2010, 17-23 May 2010, Valletta, Malta, European Language Resources Association, 2010. URL: http://www.lrec-conf.org/proceedings/ lrec2010/summaries/58.html. [15] J. Tang, H. Lin, M. Liao, Y. Lu, X. Han, L. Sun, W. Xie,
J. Xu, From Discourse to Narrative: Knowledge Projection for Event Relation Extraction, in: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), Association for Computational Linguistics, Online, 2021, pp. 732– 742. doi:10.18653/v1/2021.acl-long.60.
plexity in Digital Libraries, in: Proceedings of the Workshop on Digital Infrastructures for Scholarly Content Objects (DISCO 2021) co-located with ACM/IEEE Joint Conference on Digital Libraries 2021(JCDL 2021), Online (Due to the Global Pandemic), September 30, 2021, volume 2976 of CEUR Workshop Proceedings, CEUR-WS.org, 2021, pp. 24– 28. URL: http://ceur-ws.org/Vol-2976/short-3.pdf . [35] F. Plötzky, W. Balke, It’s the Same Old Story! Enriching Event-Centric Knowledge Graphs by Narrative Aspects, in: WebSci ’22: 14th ACM Web Science Conference 2022, Barcelona, Spain, June 26 - 29, 2022, ACM, 2022, pp. 34–43. doi:10.1145/ 3501247.3531565.