Making sense of an increasingly complex world is becoming harder in particular when facing the flood of information available in digital libraries. Hence, actively supporting users in sense making activities is a new challenge that digital libraries have to face. A popular way for humans to succeed at making sense of the world is the usage of analogies: Analogies transfer some high-level meaning from one concept to another which eases the cognitive burden to understand the concept. In brief, they connect a known concept to the unknown. A prime example is to find analogies between current and historic events to make sense of current events, i.e., place them in contexts, assess their impact, and draw conclusions on what may happen next. In this position paper, we outline a system architecture showing what an event analogy system could look like. We utilize a conceptual narrative model and define narrative roles and prototypes as a possible solution for finding event analogies.
1. Introduction ter in New York City, resulting in a collapse of the main towers and the death of nearly 3000 people. The aftermath of the 9/11 attacks is still palpable even after nearly 20 years have passed. Usually it does not take events of the size of 9/11 to invoke a need of making sense of what happened. In this context “making sense” means to understand an event s.t. one can build a mental model of it and draw conclusions regarding what happened, why did it happen and what might be happen next (similar to the verdict in [1]).
We often rely on history to find similar instances from may still happen based on the historic events. The media coverage of 9/11 for instance often drew a comparison to the attack on Pearl Harbor in 1941 by the Japanese army [2]. Moreover, street interviews shortly after the 9/11 attack also yielded the comparison to the Pearl Harbor attack (see for example [3]), therefore “Pearl Harbor” may be seen as an analogy to “9/11”. In contrast to literal similarity, an analogy tries to transfer some high-level meaning from a base to a target [4] but does not require common attributes or attribute values. But how would a digital library find and ofer good analogies?
Digital libraries (DL) ofer a wide range of information nEvelop-O (W. Balke) (W. Balke) 2. On Analogies, Events, and
Narratives computation of 4-term analogies up to a certain degree and have therefore been used in various works to capture these analogies since then (e.g. [14, 15]). 2.1. Analogy Models However, although the existence of linear relationships as shown in the Cuba example is proven for word2vec One theory on analogies is the Structure-Mapping Theory [16], they limit the scope to word pair analogies, making (SMT) introduced by [9]. SMT describes analogies as a them unsuitable to go beyond 4-term analogies. Addicomparison between relational predicates of two con- tionally, some works suggest that the workings of word cepts compared in the analogy (called base and target). embeddings in analogical querying highly depends on For instance, the 9/11 to Pearl Harbor analogy can be the corpora and the relations used (see e.g., [17]), limiting seen as an analogy, because central underlying relations the possibilities of simply relying on word2vec even fur(surprise attack on a country, national shock, relation to ther. Recently, more complex neural models have been a war following the event) are comparable. used to improve the quality of finding analogies. While
Another model of analogies is the High-Level Percep- some of them use language models to compute 4-term tion Theory (HLPT) [10] which states, that a spectrum analogies [18] others model a deep learning architecture between lower level perception (like detecting that a mimicking the SMT on synthetical data [19]. However, certain object is a tree) and high-level perception (i.e. Ushio et al. [18] stated, that the deep models overall did complex concepts like war) exist and analogies are de- not outperform the word embedding models. The work ifned by a mixture of high-level and low-level perception of Crouse et al. [19] was solely used on synthetical data, interweaving. For events this means that e.g. the pub- therefore no results for real-world applications exist yet. lic opinion on certain events play a role in perception and therefore Pearl Harbor is more likely to be seen as 2.3. Events and Narratives an analogy to 9/11 and not, for instance terror attacks outside of the USA.
Originally, HLPT was described as a competing theory to SMT [10] but it was later shown, that both theories can be seen as somewhat orthogonal [11]. SMT in that notion describes whether a given base and target tuple can be seen as an analogy, i.e. the structure mapping process between both succeeds. HLPT then goes beyond comprehension and may be used to produce analogies by building cognitive representations on the spectrum of perception for a given set of objects or concepts and then find matching representations. ing event chains without losing provenance information ture may therefore not be enough to capture analogies. regarding the validity of each relation also in the sense We argue, that events can be interpreted from diferent of information fusion over diferent repositories. In con- viewpoints. The SEM already allows changes in the roles trast, event schemas typically only model single events for participants in an event, but a viewpoint can also alter and SEM can not model inter-event relationships other event types and their connotation (e.g. the accession of than sub-event chains at all. Moreover, narrative struc- Crimea to the Russian Federation is usually interpreted tures in cognition have long been shown to be an eficient as annexation by western media while Russian sources way of transporting experiences and meaning between see it as a peaceful secession). Changes in viewpoints humans [24], which makes them suitable as an abstrac- therefore heavily influence the set of candidate analogies tion for events in the task of finding event analogies. for the target and need to be taken into account.
Combining aspects of both, ESS and EP, we can enhance narratives. Narratives as proposed in [8] can depict 3. Designing a DL System for events and their participants on diferent levels of granEvent Analogies ularity and at the same time combine knowledge from various repositories. However, viewpoints are not part As argued in Sec. 2.1, an analogy does not solely rely on of this conceptual model yet. Therefore we will extend attribute similarity, but on comparable mental structures. the narrative model for the usage as an event analogy Thus, we need an approximation of those structures for model in Sec. 3.1, before we introduce a conceptual sysevents to measure whether two events might be cast as tem architecture of an event analogy system in Sec. 3.2. an analogy or not. As mentioned above, analogies can be led back to diferent theories: While the SMT relies on 3.1. Narrative Roles and Narrative the similarity in underlying mental structures, the HLPT adds to the notion that analogies are made by connec- Prototypes tions between perception and cognitive concepts. We Determining whether a given base and target can be argue, that event analogies should rely on both theories seen as a fitting analogy is a hard task, even for humans. up to a certain degree and propose, that event analogies And as statistical models are likely to only find analogies can be determined by two components: Event Structure between events often observed together, we argue that Similarity (ESS) and Event Perception (EP). ifnding event analogies is not successful without provid
Event Structure Similarity: SMT needs base and target ing at least some human-modeled semantic elements. to share some common relation and also in the case of While the 9/11 ∼ Pearl Harbor analogy as often referevents humans tend to build analogies between events of enced example in literature might be easily discovered, similar type [2]. Yet, a strict requirement to be of exactly the analogy between the UEFA soccer cup finals of 2004 the same ontological type might be far too restrictive and 1992 may easily be missed: In 1992 Denmark won in everyday analogies, especially when using common against Germany while in 2004 Greece beat Portugal metaphors. Instead types should adhere to the same pro- despite the latter teams being the clear favorites. Examtotype. Given an event taxonomy, the prototype of some ples for this particular scenario are plenty: just recently, event is the most general superclass for the respective Switzerland won against the current soccer world chamevent that does not lose the core characteristics of the pion France during the UEFA cup in 2021. event. For instance, the event prototype “confrontation” All instances share a common narrative prototype might refer to both, an armed conflict or a football game. which might be called the “David vs. Goliath”-narrative Typically we assume confrontations to be some kind of based on the biblical figures. The weaker party of the (peaceful or violent) fight between a number of partici- confrontation (the underdog) won against the far stronger pants where one party wins, which describes both events party (the favorite) against all odds in an instance of an in an abstract way, even though war and football do not event of type confrontation. We call these patterns narhave other major attributes in common. In other words, rative prototypes and argue, that narratives that can be a competition between sport teams and a battle between matched on the same prototype can be seen as analogies. two countries, share the same prototype type of a conflict The most simple narrative prototype consists of a single situation. And while not being similar in their percep- event which allows a direct mapping from event schemas tion, they may still be used as analogies as is sometimes and event ontologies (specifically SEM) to a narrative the case e.g., in boulevard media during large football prototype. The left portion Fig. 1 illustrates this mapping tournaments.1 by trying to map the two Euro cup finals from above to
Event Perception: According to HLPT perception and the “David vs. Goliath” prototype. Greece and Portuconcepts are heavily interwoven. The same event struc- gal on the one hand and Germany and Denmark on the other hand are used as substitutions for the participants 1See [25] for an often cited example of this analogy. ?X and ?Y of the special confrontation prototype, where UEFA Euro Final (1992)
UEFA Euro Final (2004) Part. 1 wins
Part. 2 Portugal ?Y
Winner
Greece
Role: underdog
Query (narrative prototype)
List of event analogies Structured
Resources Unstructured Resources
Query Processor C1: Narrative Query Processor C2: Narrative
Role Matching C2.1: Viewpoint Approximation the respective winner must be substituted by ?Y. in the narrative to substitute the participant variables
Regarding event perception we propose narrative roles (?X, ?Y, etc.). for the modeling of narratives as shown here for under- Defining such bindings requires the operationalizadog and favorite, respectively. Both roles have a semantic tion of prototypes. In particular, algorithms to determine meaning and must be checked after the substitutions, i.e. which events fits which prototype must be able to deterthe narrative prototype only fits, if both participants do mine whether a superclass of some given event contains not violate any role constraints. Role constraints in the the inherent structure typical for the event or not. In case of Fig. 1 mean, that the participant substituted in ?X other words: a event belongs to the prototype , if must be significantly “stronger” than the one substituted the distance of its ontological concept does not surpass a in ?Y, where “strength” is defined by the concrete type given threshold. Thus, the system needs a semantically of event described by the prototypical event “David vs. curated view on the underlying ontological structures, Goliath” (in Fig. 1 international soccer matches). In this i.e. how far could one generalize an event without losing example the prototype calls for narrative roles like un- its inherent meaning. derdog and favorite but a role like “emissary” might not Regarding task b), component C2, the narrative role be a sensible role, the prototype should therefore restrict matching is used. As explained in Sec. 3.1, narrative roles the narrative roles that can be used in its context. carry a clear semantic meaning (as in the “underdog”). The manifestation of these semantics does not only de3.2. System Architecture pend on the prototypes, but also on the concrete type of the event. For instance, the confrontation prototype Building on the design in Sec. 3.1, we propose a novel ar- may define underdog as a drastically weaker participant chitecture for a event analogy system in the right portion in comparison to the favorite. However, the concrete of Fig. 1. Users are enabled to retrieve lists of analogies as features to measure the strength depend on the concrete results of respective queries. Queries generally contain event: For soccer games strength might be measured by narrative prototypes taken from a separate repository. current titles or recent successes (like world cup rankings Currently this repository is manually curated by domain or leaderboard positions) or the monetary net worth of experts, but it might become possible to automatically the team. In warlike confrontations the military strength derive suitable prototypes in later stages of the system needs to be determined. Of course, defining the relevant development. features for all event types in all prototypes manually is
The query processor then needs to perform two tasks: cumbersome. Therefore means of approximating those a) find suitable substitutions for the prototypes based features will be needed in future. on the events available in structured and unstructured Following the HLPT and the EP component of analogy formats and b) check whether the substitutions fit their detection, we take diferent viewpoints into the equation respective narrative roles or not. All narratives fulfilling with component C2.1. While net worth and champia) and b) are returned as analogies. To tackle task a), onships are objective measures, narrative roles may also the narrative query processor (C1 in Fig. 1), performs a include features which are heavily influenced by difernarrative binding against the data repositories following ent viewpoints. A viewpoint in this regard could be a the basic method of [23]. But in contrast to the narrative diferent event type in a narrative, i.e. an event analogy bindings in [23], C1 focuses on events in each prototype with more than one event. Take for instance the Crimea accession, which could be seen either as an annexation or [9] D. Gentner, Structure-Mapping: A Theoretical as a peaceful secession. Depending on the viewpoint, the Framework for Analogy, Readings in Cognitive respective event type changes and may lead to diferent Science: A Perspective from Psychology and Artififeatures and thus understanding of the narrative roles. cial Intelligence 7 (1983).
Viewpoints may be approximated by leveraging context [10] D. J. Chalmers, R. M. French, D. R. Hofstadter, Highinformation from mostly unstructured sources like news, level perception, representation, and analogy: A crihistoric documents or, if available, social media content. tique of artificial intelligence methodology, JEAIL This problem needs more refinement regarding how nar- 4 (1992). rative roles are connected to viewpoints and how these [11] C. T. Morrison, E. Dietrich, Structure-Mapping vs. viewpoints may be approximated. High-level Perception: The Mistaken Fight Over The Explanation of Analogy., in: CogSci, 1995. [12] T. Mikolov, K. Chen, G. Corrado, J. Dean, Eficient 4. Outlook estimation of word representations in vector space, In this position paper we presented a first design for an in: ICLR, 2013. event analogy system. The next step is the development [13] T. Mikolov, W.-T. Yih, G. Zweig, Linguistic regularof algorithmic ideas regarding the open questions, start- ities in continuous space word representations, in: ing with: HLT-NAACL, 2013.
[14] J. Santos, B. Consoli, R. Vieira, Word embedding • How can event prototypes be defined and how to evaluation in downstream tasks and semantic analomeasure distances between prototypes and events gies, in: LREC, 2020.
in respective taxonomies? [15] A. Drozd, A. Gladkova, S. Matsuoka, Word embed• How can narrative roles and event participants dings, analogies, and machine learning: Beyond be matched and substituted? king-man+woman=queen, in: COLING, 2016. • What is the exact connection between narrative [16] C. Allen, T. Hospedales, Analogies explained: Toroles and viewpoints in a narrative prototype? wards understanding word embeddings, in: ICML, Finally, a larger set of concrete narrative prototypes needs 2019. to be built up before a system prototype can be properly [17] J.-C. Kalo, C. Lofi, R. P. Maseli, W.-T. Balke, Semanimplemented to test the practical usefulness of event tic Query Processing: Estimating Relational Purity., analogy systems in digital libraries. in: LWDA, 2017. [18] A. Ushio, L. Espinosa-Anke, S. Schockaert, References J. Camacho-Collados, BERT is to NLP what AlexNet is to CV: Can Pre-Trained Language Models Iden[1] G. Klein, B. Moon, R. R. Hofman, Making sense tify Analogies?, arXiv preprint arXiv:2105.04949 of sensemaking 1: Alternative perspectives, IEEE (2021). a r X i v : 2 1 0 5 . 0 4 9 4 9 .
intelligent systems 21 (2006). [19] M. Crouse, C. Nakos, I. Abdelaziz, K. Forbus, [2] R. Axelrod, L. Forster, How historical analogies in Neural Analogical Matching, arXiv preprint newspapers of five countries make sense of major arXiv:2004.03573 (2020). a r X i v : 2 0 0 4 . 0 3 5 7 3 . events: 9/11, Mumbai and Tahrir Square, Research [20] A. Spitz, M. Gertz, Terms over LOAD: Leveraging in Economics 71 (2017). Named Entities for Cross-Document Extraction and [3] B. Harden, A DAY OF TERROR: VULNERABILITY; Summarization of Events, in: SIGIR, 2016.
Physical and Psychological Paralysis of Nation, The [21] W. Xiang, B. Wang, A Survey of Event Extraction New York Times (2001). from Text, IEEE Access 7 (2019). [4] C. Lofi, Analogy queries in information systems - [22] W. R. Van Hage, V. Malaisé, R. Segers, L. Hollink,
A new challenge, JKIM 12 (2013). G. Schreiber, Design and use of the Simple Event [5] S. Gottschalk, E. Demidova, EventKG: A Multilin- Model (SEM), J. Web Semant. 9 (2011). gual Event-Centric Temporal Knowledge Graph, in: [23] H. Kroll, D. Nagel, M. Kunz, W.-T. Balke, DemonESWC, 2018. strating Narrative Bindings: Linking Discourses to [6] D. R. Hofstadter, Analogy as the core of cognition, Knowledge Repositories., in: Text2Story@ECIR, The analogical mind: Perspectives from cognitive 2021.
science (2001). [24] C. Baber, D. Andrews, T. Dufy, R. McMaster, Sense[7] H. Prade, G. Richard, Analogical Proportions: Why making as narrative: Visualization for collabora
They Are Useful in AI, in: IJCAI, 2021. tion, VAW2011 (2011). [8] H. Kroll, D. Nagel, W.-T. Balke, Modeling Narrative [25] I. Thomsen, Oh, Sorry: Tabloids Lose the Soccer Structures in Logical Overlays on top of Knowledge War, The New York Times (1996). Repositories, in: ER, 2020.