D-Lib Magazine

July/August 2013
Volume 19, Number 7/8
Table of Contents

Model-Oriented Information Organization: Part 2, Discourse Relationships

Robert B. Allen
University of Tsukuba, Japan
rba@boballen.info

doi:10.1045/july2013-allen-pt2

Printer-friendly Version

Abstract

In Part 1 of this 2-paper series, we proposed developing a fabric of entities and events. In order to apply that fabric to rich content describing history we also need to consider discourse relationships. Here, we first examine binary discourse relationships such as causation, generalization, and conditionals. We then consider more complex discourse relationships such as narrative, exposition, and argumentation as well as the role of evidence. Initial applications would support interaction with relatively constrained well-defined content. Eventually, a common-format, open-access, and open-source repository based on the standards we have described here could be developed. Moreover, because our approach combines event descriptions with discourse relationships it may be possible to use it as the basis for highly interactive services and simulations.

Keywords: Argumentation, Causation, Community Models, Evidence, Explanation, Footnotes, Frame Semantics, History, Information Organization, Narrative, Personalization, Simulation Semantics

1. Enhancing the Entity-Event Fabric with Discourse Relationships

There are massive amounts of historical resources online but we are still using relatively simplistic tools for organizing that material. Hundreds of millions of pages of historical newspapers have now been digitized but interfaces for accessing them mostly use simple descriptive taxonomies and simple keyword search. To support more advanced types of access, there is a need for a systematic representation of a wide range of rich content.

Fortunately, there are many types of structure which could be extracted from this material. In [4] we proposed developing an interwoven fabric of entities and events albeit one with many gaps. However, even straightforward descriptions of history are more than chronologies of connected events. Presentations shift focus across issues, contexts, and levels of granularity. In natural language such shifts are managed with discourse relationships. In this paper, we describe the coordination of discourse relationships with the entity-event fabric.

Discourse has been widely explored in linguistics. Polanyi [25] considers the building blocks of discourse to be Discourse Constituent Units (DCUs) which she defines as describing a "single state of affairs". Two or more DCUs are coordinated with discourse relationships such as cause/effect, generalization/instance or simple narrative. Another well-known approach to describing basic discourse relationships is Rhetorical Structure Theory (RST) [21]. RST has a particularly comprehensive set of basic discourse relationships. However, because it proposes a strict hierarchical relationship among the discourse elements, RST does not seem well suited to large volumes of complex texts. Thus, we use Polanyi's approach for basic discourse elements and, as we describe later, we extend it with additional structures for complex discourse.¹

In Sections 2, 3, and 4 we discuss some discourse relationships which are easily cast as binary relationships while in later sections we discuss more complex structures. Specifically, in Section 5 we survey composite rhetorical relationships and in Sections 6 through 9 we consider narrative, exposition, argumentation, and evidence respectively. Section 10 discusses applications. Section 11 is the conclusion.

2. Causation and Causal Chains

Causation often seems natural and many simple statements imply causal relationships. Many verbs are explicitly causative [16]. Even for those verbs which are not causative, causal relationships are often apparent from the surrounding text and are incorporated into the FrameNet frames. However, the nature of causation is one of the most contentious issues in the study of history [12]. While causation by direct action seems unproblematic, attributing causation in complex situations (e.g., such as the American Civil War) can be fraught. Different individuals or groups may have very different beliefs about causes. Among the challenges in these situations is the difficulty of isolating one or a few of many influences as causes. We can identify unusual or non-normative factors as causes² but there may be little agreement about what is unusual or non-normative. With people we may attribute causation to human needs, tendencies, or beliefs; but those are also often subjective. In fact, there is considerable variability in attributions of causation even for physical causation [30].

Therefore, we allow different versions of causal claims and allow users to insert causal links for entities and events in the fabric. A claim of causation suggests that the claimant believes the user has a mental model for a mechanism which relates cause and effect. Thus users might also provide a specific mechanism which they believe is involved and might support that mechanism with an argumentation analysis (Section 8). Others could examine the plausibility of those claims and may assess whether they agree with the proposed mechanism.

3. Generalizations and Abstractions

In historical analysis it is common to propose generalizations from instances or to explore instances as examples of generalizations. In The History of the Decline and Fall of the Roman Empire, Gibbon proposed that the decay of civic virtue among the citizens of the Roman Empire led to its decline. Gibbon made a generalized causal claim which he substantiated by layers of other generalizations ultimately grounded in historical evidence. Generalizations are often heavily context dependent. In some cases, they may be validated across a range of contexts and are associated with a plausible mechanism, in which case they become theories.³

The notion of abstraction has been richly elaborated in computer science. Here, we use a schematic notation related to those from computer science, which highlights similarities in the abstractions of natural language and computer science. For example, there can be abstractions based on both entities and events/flows. In our approach, representation of generalizations and abstractions is relatively straightforward. Generalizations would be based on entity classes rather than entity instances or, perhaps, on general verbs. The evidence and reasoning supporting the generalization could be documented with the structures described in Sections 7, 8, and 9 below.

4. Conditionals, Rules, and Laws

Conditionals such as if/then statements are also a type of binary discourse relationship. They link a possible event with other events which are the consequence of the initial event. Some conditionals have a direct causal relationship; by comparison, rules and laws are conditionals in which the consequence is imposed by some external party. For example, a rule or law against theft would describe what counts as theft and then describe a consequence. Such rules are often associated with a set of assumptions and definitions which form an internal logic. Commitments and contracts are also process statements associated with conditions [9].

5. Complex Discourse

While simple causation and generalization are binary relationships, the analysis, explanation, and justification of these relationships may lead to much more complex discourse structures such as narrative, description, and argumentation. We discuss those complex types of discourse in the following three sections. In practice, these types of complex discourse may be interwoven. A story may include an explanation or an example as part of an explanation and an explanation may include a story or could be used as part of a story. In some cases, the story may, essentially, be the explanation while in other cases the two types of complex discourse are distinct.

Some long-form discourse presentations are stand-alone information resources (e.g., The History of the Decline and Fall of the Roman Empire). Those cases will incorporate coordination widgets such as tables of contexts and indexes [2]. Moreover, such information resources may follow institutional standards and sets of them may form genres [27, 31].⁴

6. Narrative

A simple narrative may be just a recounting of a thread of related events (e.g., [5]). Complex story-based narrative is often organized toward resolving a puzzle or conflict and is broken into scenes and episodes along with a setting and an ending. There have been many attempts to capture the formal structural descriptions in such complex narratives (e.g., [10, 21]). We could adopt the notations for complex stories and that coding could facilitate adaptive narratives. Episodes which were less central to the plot could be added or deleted depending on the interests of the user.

Some narratives are more concerned with character than plot. While character-driven stories still involve causation, the focus is often on understanding what causes the person to make certain decisions rather than on a coherent thread of events. In Section 10.4, we consider biographies, which are historical character studies.

7. Exposition, Description, Comparison, and Explanation

Exposition and description present the context for an event, situation, or scenario [7]. A notable aspect of exposition and description is that they are often relatively static and do not emphasize state changes. To the extent that state changes are mentioned, they may refer to long-term trends or, perhaps, constraints that maintain equilibrium of counterbalancing factors.

Comparisons illuminate similarities and differences between situations. Systematic comparisons might be usefully applied, for instance, in collections such as the Valley of the Shadow project.

Explanation typically extends description with details of causal processes (Section 2). We might, for instance, explain the processes associated with the functioning of a system. We might also explain how a given state of affairs came to be and we could provide a rationale and evidence for those causal assertions (Section 9).

Just as there are often different versions of causal attributions, there will be different explanations associated with those versions. The versions may be based on processes which involve different entities, entity states, or state changes and we need to keep the versions straight with metadata. Finally, we may contrast different versions with argumentation.

8. Argumentation

Highly-structured argumentation was originally developed for policy analysis. Argumentation is the process of critically examining, evaluating, and presenting claims. Argumentation may be applied to many different aspects of the fabric, such as critiquing whether one set of evidence is sufficient or a generalization is plausible, or contrasting multiple versions of evidence or generalizations. Computer-based programs supporting structured argumentation have been developed in the hypertext community. These programs can be considered composite hypertexts with typed nodes and links (e.g., [8, 14]). Several of these systems support community annotations (see Section 10.2).

However, hypertext argumentation systems generally do not support complex chains of reasoning.⁵ For example, Toulmin [28] proposed a structured conceptual framework for representing light-weight argumentation. In his model, arguments are macro-units which link a Warrant to a Claim as supported by Evidence. Toulmin illustrates the model with the components of a decision determining the citizenship of a person based on relevant laws, and on his place of birth. Because Toulmin's model is stronger for describing the process of argumentation than its complexity, his approach has been especially popular in science education for illustrating the relationship of evidence to hypotheses, but when [24] tried to apply it to transcripts of a complex court trial, they concluded that was "pushing Toulmin too far".

The value of formal logic in practical contexts remains debatable because there can be many versions of events and interpretations of causes. While formal logic can be occasionally be useful, judgments often appear to be based on an overall impression of the factors rather than on logical analysis. Nonetheless, while the schematic representations in [4] are derived bottom up we believe they are structured enough that they could support logic.

9. Evidence

Evidence is an important aspect of argumentation for evaluating claims about instances (e.g., whether an event actually occurred), causal attributions, or generalizations. The key questions are what counts as evidence, whether a given item fits the criteria, and the weight to be given to specific evidence. Both history [15, 22] and law recognize criteria in valuing evidence. For instance, testimony may be weighed by its consistency with well-established principles and observations as well as scientific theories. We might be particularly inclined to trust an eyewitness if there was corroborating evidence, if there was no motivation to distort the facts, or if the same eyewitness had proved reliable in other situations. Such criteria could be used to score the quality of the evidence for entities and events.

Authenticity is a substantial factor in weighing physical and documentary artifacts as evidence. Provenance is an important aspect of such authentication. Archeological evidence from a dig is often persuasive if it has been undisturbed since it was deposited. For information artifacts, provenance metadata such as PREMIS for archives and ISO23081 for electronic records are increasingly common. While systematic corruption of documents maintained in archives is rare, other threats to the validity of historical evidence range from either explicit or implicit selection bias in preserving the evidence. Archival workflows can be modeled with processes and, ultimately, archival activities can themselves be incorporated in the fabric (e.g., Figure 1).⁶

Figure 1: Inkwell used by Lincoln in drafting the Emancipation Proclamation. The "drafting" FrameNet frame includes a non-core frame element for Instrument. That association provides a direct link into the entity-event fabric associated with the Emancipation Proclamation (see [4]). We could link to the physical object with a CIDOC record. In addition to describing its heritage, we could also describe its current status as an artifact and as part of the ongoing entity-event fabric. It was recently displayed at a discussion honoring the centenary of the Emancipation Proclamation (Photograph from neh.gov)

10. Applications: Reuse, Presentations, and Authoring

Linguistics, cognitive science, and hypertext researchers have developed prototypes for some of the potential applications of schematic descriptions. We believe it is now possible to adopt and extend ideas from those projects to develop standards for describing rich content which would allow a new generation of services to interoperate and be widely applied.

10.1. Rich-Content Digital Library

Increasingly, full text and other rich content are available as information resources. A new generation of digital libraries is needed to take advantage of the capabilities that rich content allows. At the least, these resources could be indexed with traditional metadata. With rich content, footnotes in scholarly historical works could be cross-linked and even back-linked in ways that were not possible with document-level indexing (see [3]). More ambitiously, the content could be modeled as we describe above. Such modeling might be particularly useful for large national and international cultural repositories such as Europeana and the Digital Public Library of America (DPLA).

While a great deal of detailed knowledge is required for these models, rich-modeling should be a useful direction for supporting interaction with the millions of pages of digitized historical newspapers or with the wide range of materials held by historical societies. Given the complexity of the content, the development of standards for such models and their implementation in a complete system will be a long-term effort. Nonetheless, content management of large collections will continue to evolve rapidly and useful initial steps should be taken.

10.2. From Semantic Publishing to an Interactive Scholarly Commons

In our approach, the enhanced fabric could be used for a particularly rich form of semantic publishing based on the multiple levels of modeling of the content. It might also be seen as a type of text generation from schematic structures. Our approach to organizing historical information should also support schematic authoring with annotations of the extended fabric. It could be a social media platform for commentary by the public such as amateur historians and genealogists. And, it could support the archival commons suggested by [6].

10.3. Personalization and Tutoring

Because entities, events, and discourse elements are relatively distinct and clearly identified in the fabric, it should be possible to personalize their presentation to individual users though modeling that user's familiarity with them. To an extent, this is similar to the user models in adaptive hypertexts [17] that are often focused at the level of modes or pages of information. In addition to personalization for applications such as cyber-drama, personalization would also be useful in tutoring systems (e.g., [13]). In our approach, the entity-event fabric could be used as a knowledge-base. For instance, personalized summaries and feedback could be developed. Such tutoring systems could also include representations for course lessons and even complete curricula.

10.4. Biography

In historical writing, biography focuses on people. Readers may simply empathize with the subject of the biography or they may want to explore the reasons for that person's behavior. In the latter case, the biography may provide explanations (Section 7) for that behavior. In some cases, the reasons may simply be a formative set of experiences with minimal psychological analysis. In other cases, psychological states and traits combined with situational factors may be invoked as causes of behavior [23].

10.5. Narrative Interactivity

The narrative timeline (Fig 1 in [4]) provides a limited degree of interaction. The schematic approach to organizing historical information proposed in this paper and in [4] could support much richer narrative interaction. As described in Section 6 above, we could use the enhanced fabric for traditional narrative. If the fabric is richly populated with descriptions of entities and events, we could develop much richer interactive narratives. The fabric could be used as the knowledge base for interactive conversational agents and those agents could provide historical re-enactments. This might support games or interactive cyber-dramas. Many of the necessary technologies are being actively explored [26].

10.6. Threads of Events, Executable Process Descriptions, Community Models, and Simulation Semantics

While information organization has usually been thought of as being based on static knowledge-bases of documents and works, its procedures could be applied in other contexts. For instance, the coding of events might be optimized by implementing a procedure once and then applying it across a set of related situations. Moreover, once we have a corpus of entities and events, we might weave them together in novel ways to provide dynamic stories and explanations. Moving away from the fabric obtained directly from historical records, we can model historical fiction or our system could make its own simple inferences to fill in gaps in the event record. Such inferences are common in narrative history and historical fiction. In many cases, the inferences are un-exceptional but in other cases they may be controversial. In the paper, we have focused on description rather than inference because automated inference can be very difficult.

The entity-event fabric might also be used for simulations. For instance, multi-agent simulations could attempt to reproduce known events (e.g., those reported in historical newspapers). Agent-based simulations in which the roles of individuals are modeled may provide more nuanced insights about the behavior of individuals. Ultimately, the agents might be able to support conversational interaction with users and become agent-based historical re-enactors (see Section 10.5).

With rich data sources such as historical newspapers, we should be able to develop models of the community and evaluate the accuracy of those models by how well they fit the historical data. Those models could describe the range of individuals and institutions in the community and support a type of prosopography. Indeed, agent-based community models could be prosopographic simulations, though, once again, we need to caution that the assumptions of such models must be carefully documented to be useful for scholarly work.

Finally, there is evidence from cognitive science that people use mental simulations to interpret the meaning of complex situations. Feldman [11] has proposed that understanding of meaning is based on mental simulations and he calls this approach "simulation semantics". Our model-oriented approach to historical events is consistent with simulation semantics.

11. Conclusion

We have proposed that rich schematic coding of complex content will be useful for developing new ways of interacting with that content. We have taken a strong position that it is helpful to separate the entity-event fabric from the discourse applied to that material. Moreover, we have developed one approach to these descriptions which builds an entity-event fabric. Further, we propose that this fabric be overlaid with additional structures to present causation, generalization, explanation, argumentation, and evidence. Several ad hoc systems of information organization have been developed for research applications such as tutoring systems, hypertext argumentation systems, science knowledge bases, question answering systems, and even games.

We consider whether common frameworks can be developed for integrating those systems. Beyond common frameworks, we may explore strategies for populating the entity-event fabric. Digitized historical newspapers are some of the richest sources for the 19th and 20th centuries. Ultimately, sophisticated text mining of historical newspapers and other historical materials will be needed to fill in all the details. Moreover, the structures proposed here can provide a consistent standard by which the effectiveness of text mining programs can be evaluated.

Ideally, a common-format, open-access, and open-source repository based on the standards we have described here would be developed. Presumably, it would incorporate and extend linguistic tools such as FrameNet, controlled vocabularies and ontologies, and directories and records. It could include descriptive standards, abstractions (such as the periodic table), schemas, and detailed content. This would be a type of highly-enriched linked data, but it goes well beyond the current approaches to linked data in explicitly linking threads of events, in the extent to which schemas (rather than data) are linked, in allowing versioning, and in including presentation data such as session data. Such a set of resources could incorporate both scientific and historical information and could provide a common knowledge base for both fictional and non-fiction worlds.

Processes and causal relationships are fundamental to many fields. In these two papers, we explore how systematic descriptions of causal relationship might be applied to history. In other work we are examining the application of processes and causal models in science [2, 4].

Notes

¹ Another sense of the term "discourse" is concerned with the intentions and implications behind certain ways of speaking. We do not emphasize that approach because we are primarily concerned with structures. To permit different viewpoints to be expressed, we allow different versions of the content to be included.

² In a formal analysis, we might apply Mackie's INUS criteria [19] and use that structure to annotate the causal claim in the fabric.

³ In the case of history, some historians feel that history is too context-bound to consider any theories based on historical examples. Other historians (e.g., [18]) urge that the discipline of history adopt approaches and theories more like those of social science. We take no position in that debate but only use history as a potential example of the interplay of generalizations and theory.

⁴ Both [29] and [31] describe the role of Giddens' structuration applied to records. Structuration might even be extended to explain the evolution of natural language as reflected in the potential fluidity of FrameNet frames.

⁵ Probably the most highly developed composite hypertext system is SEAS. However, many of its capabilities are proprietary.

⁶ Several theorists have asked whether it is reasonable to consider historical objects and records "artifacts". Such items are not frozen when they are brought into an archive; rather, they are often part of the ongoing fabric. Even when an archival institution tries to isolate documents, those documents may be better considered as part of a "continuum" [29]. As they are used to support reinterpretations of history they acquire new significance. It is reasonable to view the entity-event fabric as an implementation of Upward's continuum.

References

[1] Allen, R.B., 2011, Model-Oriented Scientific Research Reports, D-Lib Magazine. http://doi.org/10.1045/may2011-allen

[2] Allen, R.B., 2011, Weaving Content with Coordination Widgets, D-Lib Magazine. http://doi.org/10.1045/november2011-allen

[3] Allen, R.B., 2013, Rich Linking in a Digital Library of Full-Text Scientific Research Reports. Research Data Symposium, Columbia University.

[4] Allen, R.B., 2013 (this issue), Model-Oriented Information Organization: Part 1, The Entity-Event Fabric, D-Lib Magazine. http://doi.org/10.1045/july2013-allen-pt1

[5] Allen, R. B. and Acheson, J. A., 2000, Browsing the Structure of Multimedia Stories, ACM Digital Libraries, San Antonio, TX, 11-18. http://doi.org/10.1145/336597.336615

[6] Anderson, S., & Allen, R.B., 2009, Envisioning the Archival Commons, American Archivist, 72(2), 383-400.

[7] Black, B.K., & Britton, J.B., (eds.), 1985, Understanding Expository Text: A Theoretical and Practical Handbook for Analyzing Explanatory Text, Erlbaum, Hillsdale NJ.

[8] Conklin, J., & Begeman, M.L., 1989, gIBIS: A Tool for all Reasons. Journal of the American Society for Information Science, 40(3), 200-213. http://doi.org/10.1002/(SICI)1097-4571(198905)40:3<00::AID-ASI11>3.0.CO;2-U

[9] Daskalopulu, A., 2000, Modelling Legal Contracts as Processes, International Conference and Workshop on Database and Expert Systems Applications, 1074-1079.

[10] Elson, D.K., 2012, Modeling Narrative Discourse. Ph.D. Thesis. Department of Computer Science, Columbia University, NY.

[11] Feldman J., 2006, From Molecule to Metaphor: A Neural Theory of Language. MIT Press, Cambridge, MA.

[12] Froeyman, A., 2009, Concepts of Causation in Historiography, Historical Methods, 42(3), 116-128.

[13] Graesser, A. C., Lin, D., & D'Mello, S., 2010, Computer Learning Environments That Support Deep Comprehension. In M. T. Banich & D. Caccamise (eds.), Generalization of Knowledge, 201-224, Erlbaum, Mahwah NJ.

[14] Halasz, F.G., 2001, Reflections on NoteCards: Seven Issues for the Next Generation of Hypermedia Systems, ACM Journal of Computer Documentation, 25, 71-87. http://doi.org/10.1145/48511.48514

[15] Howell, M., & Prevenier, W., 2001, From Reliable Sources: An Introduction to Historical Methods, Cornell University Press, Ithaca NY.

[16] Khoo, C., Chan, S., & Niu, Y., 2002, The Many Facets of the Cause-Effect Relation. In R. Green, C.A. Bean & S.H. Myaeng (eds.), The Semantics of Relationships: An Interdisciplinary Perspective, 51-70. Kluwer, Dordrecht.

[17] Kobsa, A., Muller, D., & Nill, A., 1994, KN-AHS: Adaptive Hypertext and Hypermedia Clients of the User Modeling System BGP-MS, Conference on User Modeling, Hyannis, MA, 95-105.

[18] Landes, D.S., & Tilly, C., 1971, History as Social Science, Prentice Hall, Engelwood Cliffs, NJ.

[19] Mackie, J.L., 1974, The Cement of the Universe. Clarendon Press, Oxford UK.

[20] Mandler, J., & Johnson, N.S., 1977, Remembrance of Things Parsed: Story Structure and Recall. Cognitive Psychology. 9, 111-151.

[21] Mann, W.C., & Thompson, S.A., 1988, Rhetorical Structure Theory: Toward a Functional Theory of Text Organization, Text, 8(3), 243-281.

[22] McCullagh, C.B., 1984, Justifying Historical Descriptions, Cambridge University Press, New York.

[23] Mischel, W., & Shoda, Y., 1995, A Cognitive-Affective System Theory of Personality: Reconceptualizing Situations, Dispositions, Dynamics, and Invariance in Personality Structure. Psychological Review, 102, 246-268.

[24] Newman, S.E., & Marshall, C.C., 1992, Pushing Toulmin Too Far: Learning From an Argument Representation. Xerox PARC Technical Report No. SSL-92-45.

[25] Polanyi, L., 1995, The Linguistic Structure of Discourse. CSLI Technical Report, CSLI-96-200.

[26] Riedl, M.O., & Zook. A., 2013, AI for Game Production. IEEE Conference on Computational Intelligence in Games, Niagara Falls, Ontario.

[27] Swales, J., 1990, Genre Analysis: English in Academic and Research Settings. Cambridge University Press, Cambridge UK.

[28] Toulmin, S., 1958, The Uses of Argument. Cambridge University Press, Cambridge UK.

[29] Upward, F., 1997, Structuring the Records Continuum, Part Two: Structuration Theory and Recordkeeping, Archives and Manuscripts, 25(1).

[30] Wolff, P., 2008, Dynamics and the Perception of Causal Events. In T. Shipley & J. Zacks (eds.), Understanding Events: How Humans See, Represent, and Act on Events, Oxford University Press, New York, 555-587.

[31] Yates, J., & Orlikowski, W.J., 1992, Genres of Organizational Communication: A Structurational Approach to Studying Communication and Media, The Academy of Management Review, 17(2), 299-326.

About the Author

Robert (Bob) B. Allen is now a Visiting Foreign Research Scholar at the University of Tsukuba, Japan. In the past several years, his research has emphasized specification and interaction with narrative chains such as those from stories, from scientific explanations, and from narrative history. Dr. Allen has been Program Chair or General Chair of several major conferences. He was Editor in Chief of the ACM Transactions on Information Systems for 10 years and Chair of the Publications Board of the ACM. He was a Visiting Professorial Fellow at Victoria University of Wellington and he has worked at the iSchools at Drexel and the University of Maryland. Earlier, he was Senior Scientist in the Information Science Research Group at Bellcore and was a Member of Technical Staff in Research at Bell Laboratories. He received his PhD in Social and Cognitive Experimental Psychology from UCSD.