Narrative Geospatial Knowledge in Ethnographies: Representation and Reasoning

Narrative descriptions about populated places are very common in ethnographies. In old articles and books on the migration history of Taiwan aborigines, for example, narrative sentences are the norms for describing the locations of aboriginal settlements. These narratives constitute a form of geospatial knowledge, and there is a need to develop knowledge representation and reasoning techniques to help analyze literatures, and to aid field works. In this paper, we outline the design of a formal vocabulary to represent and reason about geospatial narratives about populated places, keeping as close as possible to the phrases used in ethnographies. The vocabulary is implemented as OWL concepts and properties, and the rules for geospatial reasoning are expressed in SWRL. 1 Narrative Geospatial Knowledge In research and study about People and Place, it is necessary to acquire and analyze geospatial information about populated places. Such geospatial information — the location of a settlement relative to certain geographic features and other settlements, for example — often is described in a natural language, and the descriptions themselves are expressed in an everyday vocabulary that is intuitive to human but difficult to process automatically. As an example, let us look at the following sentence, which is taken and translated from an ethnohistorical article about the Atayal aborigine in Taiwan: Kanaongan Settlement . . . is located at the right bank of Dacingshuei River, where the river meets the ocean. [17, p. 137] ? This research is supported in part by National Science Council of Taiwan under grant NSC 98-2410-H-001-075-MY2 (“Ontology–based Collaborative Production of Geospatial Information”). 2 CL Chang, YH Chang, TR Chuang, DP Deng, and A WC Huang The location of Kanaongan Settlement is described by two statements about the place’s geospatial relations to others: 1) it is at the right bank of the Dacingshuei river, and 2) it is near the Dacingshuei estuary. Although we are not given the coordinates of the Kanaongan settlement, we have a good idea of the settlement’s position, even more so if we consider the facts that the Dacingshuei river is eastbound, and it flows into the Pacific Ocean. That is, we can conclude that the settlement is near by the southern river bank of the Dacingshuei estuary (as the right bank of an eastbound river is the southern bank of the river). Suppose that we learn from the same article or other sources that settlement X and Kanaongan Settlement are on the opposite sides of the same river, and they face each other. We will be able to conclude that settlement X is near by the northern river bank of the Dacingshuei estuary. Note that the conclusion is drawn without the need to know the location of the Dacingshuei estuary. Narrative descriptions about geospatial relations among Places — collected from ethnohistorical records or field interviews — constitute a kind of geospatial knowledge that is rich in domain semantics, difficult to acquire by other means, and defying easy assimilation and analysis in and by conventional geographic information systems (GIS). To systematically process and analyze large collections of these geospatial narratives, therefore, it calls for new methods and new techniques. If successful such developments will help shed new lights on several areas such as knowledge representation and reasoning, ontology and semantic web, humanistic GIS, etc. We list below several characteristics of narrative geospatial descriptions. These characteristics also magnify themselves as the main technical issues in processing narrative geospatial knowledge. – The narratives are expressed in everyday vocabularies that are rich (even diverse) in their linguistic and cultural interpretations (the “left bank” of a river, the “foot” of a mountain, etc.). – Directional and relative terms are used to describe the location of a place, as well as its spatial relations with others. The terms are approximative (“opposite to”, “about 100 kilometers away”, “near by”, etc.). – Places are identified by (common) names, not by coordinates. Their positions and footprints are left unspecified. If specified, they are imprecise or vague by definition. (For example, what is the definite spatial extent of a mountain?) – The descriptions about a place can be conflicting, incomplete, or missing. Often such places are of an ethnohistorical nature and cannot be identified by nowadays technologies. Nevertheless they can be used as the primary references for other places. (Where is Shangri–La?) 5 However, the map in the same article puts Kanaongan Settlement at the northern bank of the Dacingshuei estuary [17, p. 135]. (Also see Figure 6 in Appendix A.) This case illustrates that geospatial reasoning can be used to detect inconsistency in geospatial statements — in the article either the sentence or the map is wrong. Or it could be that our understanding of the term “right bank” is different from that of the author. Narrative Geospatial Knowledge in Ethnographies 3 2 A Vocabulary for Qualitative Geospatial Expressions The main purpose of qualitative spatial representation and reasoning is to make explicit common–sense knowledge, so that given appropriate reasoning techniques, a computer could make prediction, diagnose and explain the behavior of physical system in qualitative manner without resolving to an often intractable or perhaps unavailable quantitative model [6]. Therefore, qualitative spatial representation and reasoning not only acts as a model to clarify formal semantics of qualitative spatial objects and relations from narrative descriptions, but is also used to find new information from what is already known. There are numerous studies on this subject, on various aspects of spatial relationships such as topology, orientation, distance, size, and shape [4, 5, 9]. However, given the nature of geospatial information present in narrative descriptions, currently we only concern about the representation of certain topological, directional, and orientational information when specifying qualitative relations among geospatial entities. As a study on narrative geospatial knowledge, we have looked into ethnographies for descriptions about the locations of Taiwan aboriginal settlements, so as to use them as the sources of actual vocabularies for geospatial expressions. Take the following sentences about Sikilian Settlement as examples. Sikilian Settlement is located at the left bank of midstream Liwu River, and is about 1.6 kilometers to the northwest of the junction of its branch Wahei-er River. It is opposite to the small terrace slightly east of and below the Syuejiachang Station on Central Cross–Island Road. That is, it is at the mountain belly north of Mantou Mountain. [16, p. 178] One immediately notices geospatial phrases that may subject to different interpretations (“left bank” and “mountain belly”). The use of size and direction modifiers are also problematic (“the small terrace slightly east of”). There are pronouns and missing nouns to resolve too (“the junction of its branch Waheier River”). Nevertheless, from these sources we have identified a small set of phrases frequently used for geospatial references: – phrases for directional references such “is x kilometers to the northwest of” and “is north of”; – phrases for orientational references such as “left bank” and “opposite to”; – phrases for references to a part of an geospatial object, such as the “midstream” of a river, and the “belly” of a mountain; – phrases for different types of natural features (“river” and “junction”) and artificial landmarks (“road” and “station”). We list in Appendix A a set of nine sample paragraphs drawn from the same ethnography [16]. They are all about the locations of (historical) aboriginal settlements. Based on the geospatial narratives used in these paragraphs, we develop 6 These are direct quotations, but of our translations. The original texts are written in traditional Chinese. We take care in making accurate translations of the narratives, in particular about how geospatial references are used in the original texts. 4 CL Chang, YH Chang, TR Chuang, DP Deng, and A WC Huang a vocabulary for geospatial references, keeping as close to the phrases that are actually used in these paragraphs. Our goal is that, by using this vocabulary, we can represent and reason in a formal way the geospatial narratives in these ethnographies. Note that we are not aiming at new techniques which will automatically process natural language texts for the extraction of geospatial knowledge. Rather, we seek systematic methods to express geospatial knowledge in ethnographic narratives so that such knowledge can be aggregated and analyzed, and becomes more useful and reusable to human and machines. In this paper, we only deal with qualitative geospatial knowledge; we consider direction but not distance, for example. We have also avoided relying on numerical calculation in the representation and reasoning of geospatial narratives. 3 Framing A Geospatial Knowledge Domain Before setting out to acquire and process narrative geospatial knowledge, we need to decide on the scope and the level of details of narrative geospatial knowledge when it is represented in a system. The decision inevitably depends on the formalism we are adapting, and is constrained by the restrictions of the formalism. For this case study, we have settled on OWL DL (the Description Logic subset of the OWL Web Ontology Language) [1] as the representation formalism for the narrative knowledge, and use SWRL (a Semantic Web rule language combining OWL and RuleML) [2] for rule representation and inference. Compared to the decision on a representation mechanism (OWL DL in our case), the process of adapting and/or devising an appropriate ontology, and that of mapping the assumptions and vocabularies in a narrative knowledge domain to those in the formal ontology, is much harder. These decisions frame the domain of the narrative knowledge, hence, effect the kinds of facts to be admitted and to be reasoned with in the system. Let us take the term “mountain”as an example. In our study of geospatial narratives, “mountain” almost always refers to the area of a mountain. The terms “top”, “side” and “foot” are often used in combination with term “mountain” to refer to the top, side, and foot sub–areas of a mountain area, respectively. These areas and sub–areas may be inhabited by people. In this section, we will use this example to elaborate some of the considerations when framing the domain of geospatial areas, and use these areas as references to the places of settlements. One can choose to use Region Connection Calculus (RCC) [5] to model the various mountain sub–areas and their relations. An area is mapped to a region in the sense of RCC, and the relation between two areas is but one of the RCC–8 (or rather, RCC–5) relations. That is, each of the mountain top, mountain side, and mountain foot regions is a proper part of (PP) the mountain region, see Figure 1. Figure 1, however, expresses more. It illustrates that the mountain region is partitioned into three sub–regions of which the mountain side region surrounds the mountain top region, and the mountain foot region surrounds the mountain side region. Observing this, it could be useful to first define three regions m, s, t Narrative Geospatial Knowledge in Ethnographies 5 Fig. 1. Topological relations among the top, side, and foot areas of a mountain. of which m is the region for the entire mountain area, and PP(s, m) and PP(t, s). One can then describe the mountain top area as t, the mountain side area as the area of s − t, and the mountain foot area as m − s. However, in our case study geospatial narratives do not impose or require topological precision of this kind. The use of a region difference operator on top of RCC–5, we think, will also impose further demands to a system based on OWL DL and SWRL. One can choose instead to model the top, side, and foot sub–areas of a mountain as three distinct parts of a mountain area, but without explicit mentioning their topological relations. That is, a mountain area will have three separate object properties, each keeping the top, the side, and the foot sub–area, respectively. The classes for the three kinds of sub–areas, as well as that of the entire mountain areas, moreover, are kept distinct from one other, so there will be no confusion of mistaking one kind for another kind. In description logic, they are disjoint concepts. We have adapted this approach to modeling inhabited areas that are identified by common geographic names. That is, in our narrative geospatial domain, the part–whole relationship of a place is elaborated more than its topological relationship. For the area (and each of their sub–areas) identified by a common geographic name (e.g., “the foot of mountain X”), we require that it also possesses a “position” property indicating its geophysical location. For example, for the top, the side, and the foot areas of a mountain (and the mountain itself), they all have the summit as their position. By their unique position, the sub–areas can be related together, and associated to the mountain to which they belong. Here, a position serves two purposes — as the coordinates to be used when the geographic feature is visualized on a map, and as an (imprecise) identifier from which parts of the feature can be related together and associated with the whole. Currently we do not require an area to have a “footprint” (or, boundary) property, as such property is not easily quantifiable, even in non–narrative geospatial domains. By modeling areas referred to by common geographic names as above, we establish a framework of geospatial references for settlements, whose positions can be vague or lost. By translating narrative descriptions of settlement locations to directional, orientational, and topographical relations to existing (and persistent) geographic features and their positions, in our model, we therefore establish the relative positions of the settlements. 6 CL Chang, YH Chang, TR Chuang, DP Deng, and A WC Huang 4 Representation and Reasoning We first give a short introduction to an alternative syntax we devise for OWL DL, then proceed to describe our current implementation for the representation and reasoning of geospatial narratives in ethnographies. Figures 2 and 3 list some of the basic concepts and common properties we use in the implementation. The syntax we use is very close to those used in Description Logic (or OWL DL) but is more concise. Named concepts are introduced by the keyword Class. An anonymous concept can be constructed from a property with certain restrictions on its range, e.g., “originateFrom min 1” and “meet some Upstream”. In Figure 2, for example, a River is defined as a Place that originates from at least one Place, ends at one Place, has one Upstream, has one Midstream, and has one Downstream. (From now on, we will simply use the term ‘a river’ to refer to an individual/object in the concept/class River.) Upstream, Midstream, and Downstream, furthermore, are all subclasses of Place. An Upstream has at least one position (of class Point), a Downstream has one position (of class Point), and a Midstream must meet some Upstream and some Downstream. The intuition behind such definitions is that the place from which a river originates are associated to the sources of its upstream, and the place from which a river ends is associated to the sink of its downstream. Moreover, the up/middle/down streams of a river are pair–wise connected by the ‘meet’ relations as required in the definition of Midstream. Figure 3 shows common properties about places; these are relations among the classes we just define. The keyword Property declares a relation between two classes; it specifies the name of the relation, and the class names of its domain and its range. The square bracket immediately follows the declaration can specify the property to be “functional”, “inverseFunctional”, “transitive”, and/or “symmetric” if that is the case for the property. An inverse property f , if exists, is indicated at the end with the notation f−1. In Figure 3, for example, it declares that “southOf” is a transitive property of places with “northOf” as its inverse property. By using SWRL, we can define additional properties by composing them from existing properties. See Figure 4 for some basic rules for geospatial reasoning. This way, we are free from asserting in OWL DL those facts that can be deduced from SWRL rules, e.g., the fact that a is northeast of b given the facts that a is east of some individual x, and x is north of b. The rules in Figure 4 basically fall into two categories. The first category is about the transfer of property from the part to the whole. For example, if x is located at a mountaintop t which is belonging to a mountain m, then x is located at mountain m as well. The second category is abut common–sense geospatial reasoning such as the northeast rule we just mentioned. In our case, the geospatial narratives used in ethnographies are very concise, but to define the corresponding common–sense geospatial reasoning rules can be quite involving. For example, given the following geospatial narratives: Narrative Geospatial Knowledge in Ethnographies 7 Class People. Class Place. Class Settlement v Place u (populatedBy min 1). Class Geo. Class Point ≡ Geo u (hasHeight exactly 1) u (hasLatitude exactly 1) u (hasLongitude exactly 1). Class Landmark ≡ Place u (hasPosition exactly 1). Class Bridge v Landmark. Class Station v Landmark. Class Mountain ≡ Place u (hasPosition exactly 1) u (hasTop exactly 1) u (hasSide exactly 1) u (hasFoot exactly 1). Class Mountaintop v Place u (hasPosition exactly 1). Class Mountainside v Place u (hasPosition exactly 1). Class Mountainfoot v Place u (hasPosition exactly 1). Disjoint Mountaintop, Mountainside, Mountainfoot. Class River ≡ Place u (originateFrom min 1) u (endAt exactly 1) u (hasUpstream exactly 1) u (hasMidstream exactly 1) u (hasDownstream exactly 1). Class Upstream v Place u (hasPosition min 1). Class Downstream v Place u (hasPosition exactly 1). Class Midstream v Place u (meet some Upstream) u (meet some Downstream). Disjoint Upstream, Midstream, Downstream. Class RiverCross v Place u (hasPosition exactly 1) u (hasInflow min 2) Class PlaceGroup v Place. Class Pair v PlaceGroup u consistOf exactly 2. Fig. 2. Concepts for places and settlement in ethnographies. Sakahen Settlement is opposite to Basawan Settlement; they are separated by Mugua River. Basawan is located at the right bank of upstream Mugua River. we can infer that Sakahen is located at the left bank of the Mugua river. For this, we need to formulate and use the following rule: Pair(p) ∧ consistOf(p, x) ∧ consistOf(p, y) ∧ differentFrom(x, y) ∧ River(r) ∧ inBetween(r, p) ∧ rightBankOf(x, r)→ leftBankOf(y, r). Before the above rule can be applied, we will first need to pack the two settlements into a pair, as well as to assert that the Mugua river is in between the pair, in addition to the usual assertions that Basawan is a settlement, Mugua is a river, etc. We also need rules to enforce integrity constraints governing the facts admitted to a geospatial knowledge base. Earlier we mentioned the intuition behind 8 CL Chang, YH Chang, TR Chuang, DP Deng, and A WC Huang Property populatedBy(Settlement, People). Property hasPosition(Place, Point). Property locatedAt(Place, Place). Property hasTop(Place, Mountaintop). Property hasSide(Place, Mountainside). Property hasFoot(Place, Mountainfoot). Property originateFrom(River, Place). Property endAt(River, Place). Property hasUpstream(Place, Upstream). Property hasMidstream(Place, Midstream). Property hasDownstream(Place, Downstream). Property meet(Place, Place) [symmetric]. Property bankOf(Place, River). Property leftBankOf(Place, River) v bankOf. Property rightBankOf(Place, River) v bankOf. Property hasInflow(RiverCross, River). Property consistOf(PlaceGroup, Place). Property inBetween(Place, Pair) separatedBy−1. Property separatedBy(Pair, Place) inBetween−1. Property southOf(Place, Place) [transitive] northOf−1. Property northOf(Place, Place) [transitive] southOf−1. Property westOf(Place, Place) [transitive] eastOf−1. Property eastOf(Place, Place) [transitive] westOf−1. Property southWestOf(Place, Place) [transitive] northEastOf−1. Property northEastOf(Place, Place) [transitive] southWestOf−1. Property southEastOf(Place, Place) [transitive] northWestOf−1. Property northWestOf(Place, Place) [transitive] southEastOf−1. Fig. 3. Common properties about places. Narrative Geospatial Knowledge in Ethnographies 9 locatedAt(s, top) ∧ hasTop(m, top) → locatedAt(s, m). locatedAt(s, side) ∧ hasSide(m, side) → locatedAt(s, m). locatedAt(s, foot) ∧ hasFoot(m, foot) → locatedAt(s, m). eastOf(s, top) ∧ hasTop(m, top) → eastOf(s, m). eastOf(s, side) ∧ hasSide(m, side) → eastOf(s, m). eastOf(s, foot) ∧ hasFoot(m, foot) → eastOf(s, m). westOf(s, top) ∧ hasTop(m, top) → westOf(s, m). westOf(s, side) ∧ hasSide(m, side) → westOf(s, m). westOf(s, foot) ∧ hasFoot(m, foot) → westOf(s, m). southOf(s, top) ∧ hasTop(m, top) → southOf(s, m). southOf(s, side) ∧ hasSide(m, side) → southOf(s, m). southOf(s, foot) ∧ hasFoot(m, foot) → southOf(s, m). northOf(s, top) ∧ hasTop(m, top) → northOf(s, m). northOf(s, side) ∧ hasSide(m, side) → northOf(s, m). northOf(s, foot) ∧ hasFoot(m, foot) → northOf(s, m). locatedAt(s, down) ∧ hasDownstream(r, down) → locatedAt(s, r). locatedAt(s, mid) ∧ hasMidstream(r, mid) → locatedAt(s, r). locatedAt(s, up) ∧ hasUpstream(r, up) → locatedAt(s, r). eastOf(a, x) ∧ northOf(x, b) → northEastOf(a, b). eastOf(a, x) ∧ southOf(x, b) → southEastOf(a, b). westOf(a, x) ∧ northOf(x, b) → northWestOf(a, b). westOf(a, x) ∧ southOf(x, b) → southWestOf(a, b). northOf(a, x) ∧ eastOf(x, b) → northEastOf(a, b). northOf(a, x) ∧ westOf(x, b) → northWestOf(a, b). southOf(a, x) ∧ eastOf(x, b) → southEastOf(a, b). southOf(a, x) ∧ westOf(x, b) → southWestOf(a, b). Fig. 4. Basic rules for geospatial reasoning about places. 10 CL Chang, YH Chang, TR Chuang, DP Deng, and A WC Huang the definitions of River, Upstream, Midstream, and Downstream. This kind of intuition actually is formalized as a SWRL rule over a river r and its constituting parts up, mid, and down as follows (assuming river r originates from two sources m and n): River(r) ∧ hasUpstream(r, up) ∧ hasMidstream(r, mid) ∧ hasDownstream(r, down) ∧ originateFrom(r, m) ∧ originateFrom(r, n) ∧ endAt(r, e) ∧ hasPosition(m, p) ∧ hasPosition(n, q) ∧ hasPosition(up, p) ∧ hasPosition(up, q) ∧ hasPosition(e, z) ∧ hasPosition(down, z) ∧ meet(mid, up) ∧ meet(mid, down) → Checked(r). Similarly, for a mountain m, the part–whole constraints between the mountain top/side/foot and m itself are checked by the following SWRL rule: Mountain(m) ∧ hasTop(m, t) ∧ hasFoot(m, f) ∧ hasSide(m, s) ∧ hasPosition(m, p) ∧ hasPosition(t, p) ∧ hasPosition(s, p) ∧ hasPosition(f, p) → Checked(m). Note that in the above we use the concept Point to anchor the concept River: Individual points are used to relate a river to the river’s constituting parts. So is the case for Mountain. We can relate rivers to mountains in a similar way, by connecting the place from which a river originates to certain mountains. This allows us to say that a river originates from these mountains. Finally, we mention that some implicit facts can be derived directly by an OWL DL reasoner without the need of SWRL rules. For example, from the narratives: Kumuge Settlement is southeast of the river–cross of Mugua River and Cingshuei River. Kumuge is northwest of Tongment Settlement. one can infer that Tongment is also southeast of the Mugua and Cingshuei river– cross. This is because the inverse property of northWestOf is southEastOf, and southEastOf is a transitive property. In addition, certain queries to a knowledge base can be formulated as class definitions (i. e., new concepts with additional restrictions). By reasoning about whether the class is empty, an OWL DL reasoner will be able to deliver the results for us. For example, the query “what settlements are west of Sipan and are near by some rivers” can be answered by reasoning about the individuals in the following class: Class WestOfSipanNearByRivers ≡ Settlement u (westOf has Sipan) u (bankOf some River). We now return to the illustrative example we use in Section 2: Sikilian Settlement is located at the left bank of midstream Liwu River, and is about 1.6 kilometers to the northwest of the junction of its branch Wahei-er River. It is opposite to the small terrace slightly east of and below the Syuejiachang Station on Central Cross–Island Road. That is, it is at the mountain belly north of Mantou Mountain. Narrative Geospatial Knowledge in Ethnographies 11 Geospatial knowledge in the above narrative is now formalized by the following assertions in our implementation: Sikilian : Settlement. Liwu, Wahei-er, U : River. MidLiwu : Midstream. Liwu Wahei-er : RiverCross. Mantou : Mountain. MantouSide : Mountainside. Syuejiachang : Station. T : Place.