We propose to use dynamic Extreme Tagging Systems in combination with the classic approach of static KR models like ontologies, thesauri and gazetteers. We show the usefulness of the approach for local search by applying it to the notion of place defined as a location that supports a homogeneous affordance field.
We propose to use dynamic Extreme Tagging Systems in combination with the classic approach of static KR models like ontologies, thesauri and gazetteers. We show the usefulness of the approach for local search by applying it to the notion of place defined as a location that supports a homogeneous affordance field.
We propose to use dynamic Extreme Tagging Systems in combination with the classic approach of static KR models like ontologies, thesauri and gazetteers. We show the usefulness of the approach for local search by applying it to the notion of place defined as a location that supports a homogeneous affordance field.
KMi – The Open University KMi – The Open University Walton Hall, Milton Keynes Walton Hall, Milton Keynes MK7 6AA, United Kingdom MK7 6AA, United Kingdom +44 (0) 1908 655412 +44 (0) 1908 655014 v.tanasescu@open.ac.uk j.b.domingue@open.ac.uk vladtn@gmail.com ABSTRACT constitutes an important percentage of internet searches [1] [3]. For extracting the characteristics a specific geographic entity, and The text based nature of local queries imply the same semantic notably a place, we propose to use dynamic Extreme Tagging issues as non geographical text searches: the ambiguity of natural Systems in combination with the classic approach of static KR language, in the query, as well as in the indexed documents, models like ontologies, thesauri and gazetteers. Indeed, we argue combined with the lack of knowledge of search engines, leads to that in local search, the what that is queried is implicitly about suboptimal results. Moreover, geographic space, i.e. the places. However existing knowledge representation (KR) models, environment which supports geographic objects as opposed to such as ontologies based on logical theories, conceptual spaces, table-top ones, is a domain in which scale, orientation, affordance or other, cannot capture in isolation all aspects of the boundaries, and cultural conceptions matter to a great extent [4], meaning of a place. Therefore we propose to use a combination of making verbal consensus more difficult. Moreover, them based on the underlying notion of differences, linked representations in this domain have to alternate between the elements of meaning without commitment to any KR model. object and field models [5], between the scientific view of a Mapping to elements of different KR models can be made later to measured space and the experienced vision of our common follow the requirements of a given task, supported by a KR medium, studied as part of Naïve Geography [6], and must be representation of the elements that support this task. We show the able to present multi-representations depending on the activities usefulness of the approach for local search by applying it to the involved [7]. Hence, the introduction of the notion of locality in notion of place defined as a location that supports a homogeneous generic web search makes information retrieval even more affordance field, i.e. the spatial area which allows me the do a problematic. particular thing, while allowing the homogeneity of movement, To ease the integration of local data query oriented applications meaning that the previous field is not interrupted by any such as geographic web mashups present in an ad hoc way boundaries. information relevant to the user. These applications successfully handle the semantics of a particular domain in a limited context, Categories and Subject Descriptors e.g. looking for apartments in an area1. However, (i) handling data I.2.4 [Artificial Intelligence]: Knowledge Representation heterogeneity still requires considerable manual work as every Formalisms and Methods – semantic networks, representation additional data source has to be integrated manually, (ii) the lack languages, Web-based services, Spatial databases and GIS. of semantics limits the expressivity of queries and the precision of the results, as the applications generally only allow the filtering of answers using simple dimensions such as price for the house one General Terms is looking for etc., and (iii) in mashups the results only take into Algorithms, Reliability, Experimentation, Human Factors, account a single context and generally do not tackle the Theory. specificities of geographic space beyond point based representation. Keywords Alongside individual efforts of mashup developers, generic Local search, knowledge representation, differences, affordances, applications such as Google Local have been built to aggregate multi-representation, image schemata, conceptual spaces, extreme geo-located results in an automatic way. As opposed to mashups, tagging, Wordnet, similarity, AI. data is integrated automatically, but this comes at the expense of the quality of the information retrieved, and of the results’ 1. INTRODUCTION expressivity. Therefore improving the quality of local search in a Local search, i.e. the retrieval of geo-located elements for generic web search context is still a necessity. everyday use [2] – to be distinguished from text queries without In order to improve the quality of local search alternative efforts geographical scope as well as from professional GIS usage – based on higher semantics have been advocated in the NLP communities, the Semantic Web and the GIScience ones. Efforts Permission to make digital or hard copies of all or part of this work for have been deployed to describe geographic space semantically, personal or classroom use is granted without fee provided that copies are often by building ontological descriptions. However, the simple not made or distributed for profit or commercial advantage and that definition of geographic entities is more problematic in copies bear this notice and the full citation on the first page. To copy geographic space (e.g. a forest or the town centre are vaguely otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. LocWeb 2008, April 22, 2008, Beijing, China. 1 Copyright 2008 ACM 978-1-60558-160-6/08/04…$5.00. e.g. http://www.housingmaps.com/ defined, context dependent, or can spread over multiple areas) Major web actors are aware of the growing importance of local than in table-top space (e.g. a chair or an apple). Therefore, in search alongside pure document search and local search engines order to build more practical solutions, logic based ontologies have been developed by Yahoo!3, Google4, and Microsoft5. All have been merged with allegedly cognitively sounder approaches: offer a yellow page style geographic aggregation of business paradigms originating from psychology, cognitive science or addresses with several map visualization features. They also philosophy, such as affordances, image schemata, conceptual provide a simple natural language query system introduced spaces and multi representation models, have been advocated. through examples showing query patterns. To illustrate the However these models have often been applied at the exclusion of difficulties encountered when using these systems we tried others, revealing only a few aspect of geographic space but failing Google local with the example query “hotels near lax” (lax to handle all its characteristics. standing for Los Angeles International Airport). The query is successful in that it returns a list of hotels situated at what appears The notion of place is essential in geography and, we believe, in as a walking distance from the airport. However the second result local search, not only because the location of the where aspect of page displayed hotels which were closer than the first result set, a query is often expressed as a place, named or not [8], but also while in the following pages start to appear items which do not because the what aspect itself is essentially about places that appear to be hotels at all (e.g. the company “Westchester Travel allow users to do things. Therefore the notion needs to be Service”, which indeed displays Los Angeles hotels on its home elucidated. We will argue that the special character of geographic page, and is classified in a yellow page website as “Hotel Motel & space is only due to the over fitness of KR concepts to table-top Resort Reservations in Los Angeles, California”). We also notice space elements and to domains such as controlled experimental that trying “restaurant near lax”, returns precise results, but that environments (i.e. “toy worlds”) or scientific ones (e.g. biology switching from “restaurant” to “bistro” or “eatery” returned classifications), and are therefore not adapted to the description of results which were as close or even closer but which were not embodied meaning such as the one relevant to geographic space. included in the “restaurant” result set: the system did not take into Therefore alternative knowledge representation techniques need account that an “eatery” was the same as a “restaurant” and that a to be used in order to represent the dynamic aspect of real world “bistro” was similar in essential aspects as well. meaning, a meaning typically human and social, rather than purely logical. In previous work we developed the notion of Another query, “sushi near lax”, returned good results, and even difference, elements of meaning able to assume different roles in the query “great sushi near lax” did not return restaurants situated different KR models, and therefore allowing to move from one to more than 22 miles from the airport, and traditional American another [10]. In order to capture meaning outside of any food restaurant as it did in 2005 [15]6. However, by switching the predefined KR model, and by doing so to express differences, we query to “wine near lax” we were oriented toward specialized extended the technique of tagging to allow tagging the tag wine shops – some of which more than 42 miles away from the themselves as well as the relationships between them [11]. airport – neglecting the fact that, given the cultural importance of wine in western culture, and in California, a bottle of wine could First we are going to describe the challenges of local search, and probably be obtained in any drinking place or restaurant. its link with geographic space and the notion of place. Then we will introduce differences, followed by Extreme Tagging systems Finally, the attempt of improving the results by adding keywords and present an example applied to the notion of place. We will to the query, e.g. “buy wine near lax”, was unsuccessful as the conclude with a summary of the advantages of the techniques results only returned pages that mentioned the word “buy”. used as well as indications of future directions. It seems obvious that with a bit of semantics, e.g. the knowledge that a restaurant serves wines, and that eateries are restaurants too, 2. LOCAL SEARCH, SPACE AND PLACE the results could be greatly improved. Even without changing the 2.1 The State of Local Search retrieval system, adding relevant keywords to the query, e.g. If literature exists about the nature of geographic terms used in “restaurants and eateries and bistros near lax”, or sending several local search queries [3], to our knowledge, a complete typology of queries in parallel and aggregating the results, would lead to local queries is still missing. Therefore, to describe local search, a better precision. In order to achieve this in the general case, basic query model is used here, identical to the ones used in [12] however, we need to understand the semantics of geographic and [13], and to the one proposed by the GeoCLEF 2007 query space, and of place, and to represent it in a machine-processable parsing task proposal 2 . It consists in a (what, spatial-relation, way. where) triple. Hence examples queries are: 2.2 Representing Spatial Semantics • “food near lax” Currently, the most common means of expressing high level semantics is the use of logic based ontologies. In this sense, an • “buy food in lax” ontology is “a logical theory accounting for the intended meaning • “eat in LA” of a formal vocabulary” [9]. The apparent universality and
• “restaurant near me”
3 http://local.yahoo.com Less obvious examples – such as “the closest restaurant”, “rivers 4 with vineyards”, etc. [14] –, can be converted to this normal form http://local.google.com – i.e. as “restaurant near my location” and “vineyards near 5 http://www.live.com rivers”. 6 Such improvements for queries involving very specific keywords are certainly due to the increase of the quantity of 2 http://ir.shef.ac.uk/geoclef/2007/Query-Parsing.htm available documents versatility of this definition has lead many researchers to advocate Therefore alternative KR models are often used in isolation (as their use in a geospatial context (e.g. [16][17]). However some for example conceptual spaces in [25] or in [26]). This expressed scepticism (e.g. [18]), originating from the fact that, if unfortunately reveals their own expressive limits, as it is for ideally “meaning” expressed by ontologies would provide the example quite difficult to emulate is-a hierarchies with conceptual long sought for “glue” between geospatial communities by spaces. capturing their practices and conceptualisations, and ease the Given the usefulness, as well as the intrinsic variety of the notion alignment of heterogeneous elements, the actual use of ontologies of place, several of these models have been applied to it in order for spatial knowledge representation quickly leads to issues which to present a coherent description. seem difficult to overcome. These issues are related to object representation in geographic space when trying to handle 2.3 The Notion of Place vagueness [19] as well as cultural and subjective discrepancies There is a consensus on the fact that a place is related to a (e.g. [6] [20]). location, a spatial footprint, and that places, as locations, can be To overcome these difficulties alternative KR models such as nested. However, once this basic fact has been established, places affordances, image schemata, conceptual spaces and multi appear to be used in very diverse ways in natural language [32], representation models, have been advocated. Each of these probably too many to be categorized. For some authors, the roots models have specific components that are used to describe aspects of this variety can be traced to the origins of our sense of place, of knowledge. We are going to describe them in turn, starting that meaning that each particular place has for an individual. with logic based ontologies. Jordan et al. [28] for example, quoting Curry [29], argue that places do not have predefined boundaries but that a place is a Logic based ontologies describe a domain by using categories and “location that has been given shape and form by people”. The by defining relations between them. Categories have attributes diversity of the meanings of place simply reflects the complex and are related by is-a relationships to form taxonomies. Instances genealogy of the sense of place, i.e. the ways by which we are individuals belonging to a category. Moreover, some ontology transform a simple undifferentiated location into a place: languages provide procedures, functions, axioms and (production) rules. • Naming operations: giving a name to a new place is a process of appropriation (e.g. New Hampshire in the Multi representation models promote the notion of context- USA). dependent identity. Indeed, instead of linking an entity to a unique set of categories with fixed sets of attributes, multiple • Applying Typologies: applying geographic categories to representations are provided depending on the context. These things (e.g. stream, suburb, etc.) makes them place-like. representations are not only visual (e.g. an airport being • Making/Picking up a Symbol: a metonymical operation represented geometrically as a point at world scale and a complex (e.g. the pyramids for Egypt). structure at a closer scale) but also concerns the attributes, which can change depending on the particular representation dictated by • Telling Stories: a place described by what happened the context [21]. there (e.g. battles, past life events, etc) Conceptual spaces, as introduced by Gärdenfors [22], present • Doing Things: places become tangible through activities individuals as points in geometric multi dimensional quality (e.g. everyday routines, or things we do when arriving spaces. As an example, a particular colour is represented as a to a new town) point in the three dimensions of the primary colours. This It is reasonable to think that the sense of place comes from a representation allows to easily measure similarity between two combination of these activities. For example a territory which has individuals as the distance between two individual points. been given a name from a distance, will be more a place if a Image schemata are described as “a small number of parts and typology can be applied to it – e.g. if it is rocky –, even more a relations, by virtue of which it can structure indefinitely many place if stories are told about it (e.g. ghost ones), and even more if perceptions, images, and events” [23], and are basically a small human start using it for activities such as commerce. If this notion set of structures which support relations. Examples are container, of place seems a bit remote from the common experience of place balance, compulsion, blockage, counterforce, etc. Schemas can be it is because most places have already been combined to create complex structures. For example a crossroad named/categorized/used for us in the process of building a culture is composed of two paths, which are themselves based on the and a society and that most individuals do not have a word to say image schemata links and support. in this process [30]. However, these elements are reactivated when we find ourselves in alien environment, travelling or after a Affordances are an essential concept of ecological psychology life event. [24]: they are possibilities for action latent in an environment. As actions have to be carried by someone, the users’ capabilities This genealogy of the sense of place is useful in order to have an important role in the definition of an affordance (e.g. a understand why a place can be vague, why it does not always wall does not afford to be climbed if I am not tall or fit enough or have to be based on a geographic category, that it can be created do not have a ladder). ad hoc as categories can, depending on the context. Given this genealogy, places also appear as essentially human, naming being Most of these paradigms are only superficially compatible with a way of encompassing what is named, at least indirectly on a ontology languages (e.g. quale in DOLCE [27] are inspired by map, applying typologies requiring a particular culture and maybe conceptual spaces, but constitute a minimal part of the underlying language [20], picking up symbols implies political or cultural theory) and integrations tends to disregard the specificities of the interests, as do telling stories, and doing things implies human paradigm (Gibson’s affordances for example, are often activities and the human ability to move around objects. Given assimilated to functions, or their meaning is limited to an HCI these various origins difficulties that arise in categorizing the environment, or the concept of direct perception is disregarded). notion of place in an ontology, as described in [31][32] seem more than understandable. Even when these origins are not taken movement and affordances is essential for local search since if the into account, existing ontology languages such as OWL seem to where part of a local search is closely related to the notion of be inappropriate [33] to the description of existing places and place, the what is part of the description of a place as well. need to be complemented in order to capture the human and Indeed, when I query for something local, “food” or a inherent dynamism of the notion. “restaurant”, I am implicitly looking for a place which provides To include all these elements Jordan et al. [28] have proposed the food, or for a place which is a restaurant. Indeed, asking for food description of a place as a network of affordances, as the location or drinks nearby does not usually mean where can I find food on where given affordances are presented to us, but the extreme the streets, which would be accidental, and as such could not be variety of affordances to take into account has been found localized, but means I am looking for a place that serves food or problematic. A multi paradigm approach to the description of drinks. geographic object has been advocated by Kuhn [34], combining image schemata as supports for affordances, approach which can 3. REPRESENTING MEANING be applied to the notion of place. We will build on these two The meaning of place must embrace all the dimensions that approaches in the next section to present an alternative notion of constitute the sense of place, and, to be operational, still allow to place. be represented in the various models of which components can then be algorithmically manipulated. The notion of differences 2.4 Local Search as Search for Places aims to fulfil the role of grounding for generic meaning without According to the previous genealogy one can see that a place is a) the need to commit to a particular KR model. Tagging, and more experienced, which implies that all aspects of human experience particularly Extreme Tagging Systems, i.e. tagging systems can get attached to it, through direct experiment, cognitive allowing the tagging of tags themselves, allow to represent picking of categories or landmarks, or indirect narrations, etc. and meaning with minimal commitment. A tag can then be projected b) explored, geographic space and places are only accessible onto elements of any KR model in order to become operational. through movement, and therefore everything that modifies the In the next section we briefly explain where the notion of capacity to move, such as public transportation in a town, or difference originates from and how it is defined. Then we disability in a home, has an impact on the sense of place. Our introduce Extreme Tagging Systems (ETS) as a way to describe hypothesis is that the spatial footprint of a place is to be linked difference networks and as a knowledge capture tool. ETS are with the experiences a location allows, establishing a tension then used to describe a place. which is the place itself. In other words the location is constitutive of a place in so far that it is recognized as a homogeneous 3.1 Meaning as Differences affordance field and that it allows the homogeneity of movement. Work on differences as a grounding for meaning originates from For example, a restaurant allows to eat as long as you are inside Gilles Deleuze’s study of the concepts of Difference and Identity, its walls, or sitting immediately outside at a table. The separation from the emergence of collaborative tagging systems in Web 2.0 made by the walls of the restaurant building are secondary applications, as well as from work of Jaques Derrida on the compared to the notion of affordance field and to the fact that I dynamic aspects of his concepts of différance. According to can easily move from inside to outside, but as soon as I move Deleuze, differences precede identities, as they allow us to make a beyond a given limit, e.g. out of sight, I find myself in a place distinction between things giving them an identity a posteriori. It which presents different affordances, i.e. which is not a restaurant is the differences we can establish between elements of the world anymore. Another example is the unity of a country as a place, and ideas that allow us to identify them and to establish similarity even if it’s geographic extension is constituted by multiple islands statements, and not the opposite. Derrida is interested by the (e.g. Italy). The sense of place is maintained by the fact that my relations between elements of meaning, and, following the model movement is homogeneous, i.e. that there exist relatively of the linguistic sign, describes the process of becoming homogeneous ways to go from one island to another, or from something else in order to signify, coined différance. Indeed, as mainland to the island, in terms of speed, frequency, price, etc, the nature of the sign is to refer to something else in order to and that I can expect similar affordances, such as infrastructures, acquire meaning, any element of meaning itself is linked to other habits, or maybe languages, in all parts of the country. We believe meanings and only makes sense in this movement towards the combination of both factors is important as only movement something else. In Web 2.0 applications, collaborative tagging allows to provide the feeling of affordances as a field, while on gives the user absolute freedom in the elicitation of a label, or tag, table top space, affordances are supported by quasi punctual independently of any pre-established categorisation, or indeed KR objects. It also provides us with an articulation between the two model which would be based on categorisation. spaces. For example my house is a place where I can sleep since it contains (punctual) objects which allow me to lie down, and that Hence, a difference [10] is defined an element of meaning, atomic my movement is only marginally interrupted by boundaries inside in that it has a specific meaning by itself, but which can also be its walls. The feeling aspect of the notion of place, for example reduced to the meanings he is linked to. It is anything that makes the one of safety, which certainly contributes to my perception of sense, emerging from a background and which can be isolated by affordances – as I would not lie down in a potentially hostile an actual process. A difference cannot be dissociated from the environement –, is not explored here but supposed to be implicit process which allows it to signify by itself, unlike symbol systems in the way people tag things, therefore choosing their affordances. which signify only thanks to relations between symbols. For example particular colours, shapes, a distinct word, a sound, an The notion of place, and a definition of this notion which involves action, an event, a town etc., exhibit meanings I can understand Copyright is held by the author/owner (s). immediately if I have the appropriates processes to grasp them, LocWeb2008 held at WWW2008, April 22, 2008, Beijing, China ACM but it is still can be reduced to other differences. The meaning of a 978-1-60558-085-2/08/04 “restaurant”, once this meaning is acquired, becomes present immediately, but other differences are linked to it and, for example, the presence of a building, with tables and a menu in Since the notion of difference is dynamic rather than static, the display will certainly evoke the meaning “restaurant”. Only criterion of applicability of a KR paradigm to the flux of meaning differences carry meaning, as, for example, I may not be able to that are difference becomes the emergence of reasoning distinguish as many shades of a colour as an artist, which means capabilities, and not a representation that may or may not support they have no meaning for me, and that they cannot be attached to some reasoning. Therefore differences have to be projected onto a other meanings either. particular KR model in a punctual, dynamic and task oriented Hence difference-based KR is distinct from symbolic KR in that way. differences are not static symbols, but always the result of an active process. The processes that produce differences, or rather 3.2 Extreme Tagging Systems and Differences that are differences, themselves process other differences, other Extreme Tagging Systems (ETS) are an attempt to use the free elements of meaning. These processes can be algorithms association processes common to tagging, in order to acquire operating on data in an information science context, or chemical knowledge [11]. In ETS tags can themselves be tagged, as well as processes operating on modifications of the physical environment the relations between them. For example, some image of a car in a biological context, or actions operating on a changing would be tagged with “car”, and “car” itself with “wheels” and environment in an ecological psychology context. Independently the relationship between the two with “part-of”. These operations of the nature of the process, the result, if detected as meaningful, can be done collaboratively, online, by several users on the same is called a difference, and is the production of a snippet of tag network, or individually on one’s own tag network. Tags are meaning that can be processed further. Therefore, as differences, URIs and can contain for example location data. Filtering one can follow the trace of meaning from chemical processes operations allow to query the whole network or to limit it to a detecting molecular variations to mathematical, logical, cognitive particular user or community of identified users. The tagging can or even social processes, as an uninterrupted chain of actions and be done by users on separate websites and the results retrieved reactions which transport meaning through differences. They through RDF. To summarize ETS are Semantic Networks [35] provide the link which makes logical actions meaningful, but also with no particular constraints on the vocabulary during the social signs, particular actions, designs, environments, etc. creation phase. However some control mechanisms are applied to the tagging operation itself. Differences, as processes, cannot be represented directly: they are a transformation of something into something else, and the simple Indeed, when tagging tags, the operation of tagging is meaningful, description of preconditions/postconditions, inputs or outputs, i.e. the fact that an intelligent being chooses to tag a tag with tells us little about the actual transformation process. However, another tag is not arbitrary, and this operation can be linked to the being meaningful, differences can be reflected in a particular KR laws of association of ideas [36]. Connectionists argue that in paradigm, in the similar way that, for example, an AND gate analogy with the neural structure of our brain, the stream of executes a process on an electronic board, process which can be thought, i.e. the way one thought leads to another, seems to follow described by a logical formula, and which can be emulated by a certain rules. Therefore, in ETS, taggers are asked to describe the computer, but not represented directly as a process acting on meaning they give to their tagging activity, i.e. why the electrons, as the process just is. In a constrained environment such movement from one tag to the other occurs, independently from as an electronic system, meaning can be captured by one or two the relation they may choose to add after that between two tags. paradigms, such as logical and analogical. Less constrained Association laws that have been selected for ETS are: environments however, such as those experienced in everyday life or the ones related to geographic space are complex fluctuant • similarity: a tag has been applied because the concepts mosaics of meaningful processes, differences, of which KR are similar (e.g. from pub to bar), although the paradigms can only partially and temporarily map in their similarity is not explored further at this level. diversity. • co-occurrence: a tag has been applied because of the Finally, if differences, being processes, cannot be represented, we presence of the two concepts in the real world (e.g. from can represent the links between them as a graph. As an example, a car to wheels). We do not investigate further the nature network of differences describing aspects of Milton Keynes (UK) of the relation, e.g. we do not ask if it is topological, City Centre (MK CC) is presented in Figure 1. mereological, temporal, etc. • sound-co-occurrence: the words are often found together (e.g. “Japanese restaurant” has a particular meaning, a tag in itself, but the words “Japanese” and “restaurant” are sometimes found together, and therefore I may tag “restaurant” with “Japanese”). • Cause-effect: e.g. “weapon” and “death” • sound-similarity: e.g. “pink” and “link” • opposition: e.g. “life” and “death”, “something” as opposed to ”nothing”, etc. The Extreme Tagging activity, consisting in following a stream of thought while exploring a particular meaning, going from an element to others, without caring about categories, or being Figure 1. A network of Differences constrained in any other way by any KR model, appears as a good candidate to the capture of differences, meaningful relations One can see that ETS networks do not have to be complete. For before any inclusion in a KR model. A tag is a word, here an example pizza is also a kind of food, although the diagram does English one, which tries to express the meaning of a particular not indicate it: as the network is evolving constantly this is not a difference. It is a natural language representation of a difference permanent and therefore not a worrying issue. network, not to be mistaken with differences themselves – which ETS are work in progress, however we believe that useful results are dynamic processes of meaning production –, but both ETS can already be extracted from their use, and notably in networks and difference networks share 1) the characteristic of conjunction with the notion of place. avoiding to commit to a KR model, and 2) a network topology. Moreover by asking users for the association laws used during the tagging process, we avoid the case where, eventually, “everything 3.3 The Meaning of Place is somehow related to everything”, as the systems allows to know We can now build a representation of place not by trying to fit in what sense something is related to something. elements of place in a given KR model, but by freely describing the meaning of a place before projecting it on various KR models Extreme Tagging has been inspired by social networks. depending on the context. Tagopedia, the first implementation of an ETS is based on the Facebook application framework [11]. This allows introducing, if As previously described, our notion of place is a location which needed, further regulation mechanisms based on collaboration. allows homogeneous movement and affordance fields. Therefore, For example weighting allows to increase the importance of tags from our KR toolbox we are going to transform tags, which are which have been applied by many users, as well as to implement a non committed representations of differences, into committed recommendation system. Also tags from others can be ranked, ones, i.e. into elements of a particular KR model. The choice of and the author of one tag can ask a friend about the meaning of these elements will be dictated by our ability to extract them from his or her tag. In this way we try to create some social meaning, an ETS, as well as by our task. Here a place will be an ontological non only a multitude of individual independent ones. geographical category (or a group of categories) which provides affordances. Once some knowledge has been captured by a collaborative process, several filtering levels are present: One solution to the problem of deciding what a geographical category or an affordance is, is to use social mechanisms to • Timeline: groups of tags added during a same session annotate the network. Indeed, as we ask each user about the are considered to be more related than the others (e.g, process of association which makes him or her move from one tag in one session “tank” can be tagged with “fish”, while to another, we can also ask taggers to select the type of a tag from in another “tank” can be tagged with “weapon”). a list. However as we plan to map to different KR model elements depending on the context (here, places and local search, mapping • Personal: only my own tags are considered to entities and affordances), this process can be tedious for the • Group: I choose a group and only the tags of this group users, and we may have to wait for this information after the are considered introduction of every new application. Another solution, as only broad categories are needed, is to use • Friends: only my contacts’ tags are presented. comprehensive pre-existing ontologies, such as the almost formal • Global: all tags are shown. Wordnet to look up the meaning of each term. Our test queries will be “food near Lax”, and “eat near Lax”. We assume that the Here is the representation of the previous difference network, where term, “Lax”, can be discovered from a gazetteer like in where nodes represent tags and arcs represent the action of [38], which also allows us to handle nearness relations. Hence, for tagging (cf. Figure 2). Only the association relations are shown, the first query, “Food near LAX”, we start with the what term, tagged relations are here not needed and therefore not Food and notice that it is similar to {sushi} and co-occurs with represented. {MK CC, restaurant, pub} (cf. Figure 3). Food is not a geographical category according to Wordnet, i.e. there is no hypernym which can be related to a geographical entity, and therefore is it considered as an affordance: what we are looking for are things that provide food. MK CC cannot be found in Wordnet and therefore is ignored. Restaurant has building in its list of hypernyms, which is considered by our algorithm a geographical entity. Note that the notion of scale is trivially handled, e.g. if food is linked to buildings but also larger areas such as named places, both can be returned. Moreover restaurant has synonyms ({eating house, eating place, eatery}) and hyponyms ({bistro, brasserie, brewpub,...}) which can be used in the query. Food also co-occurs with pub, which is similar to bar. Pub and bar are buildings and have their own synonyms and hyponyms. Therefore, when looking for food, which is implicitly looking for a place that provides food, the query can be enriched with additional terms, the synonyms and hyponyms of the linked geographic entities, which may be mentioned on the indexed web pages. Figure 2. An Extreme Tagging network automatically from textual or other knowledge sources. Moreover, ETS allow the extraction of subjective knowledge. For example if the relationship between food and restaurant is somehow generic and can be found in an ontology, queries about “Fun in LA”, in which the system would be unable to return relevant results for fun without the help of a personalised Extreme Tagging network. We also believe that the use of ETS for local search and the noticeable improvements will help motivate users to improve their own tag network, and, by doing so, contribute to the shared knowledge of their community. We have described a notion of place based on a new approach to knowledge capture and representation that avoids the commitment Figure 3 "Food near LAX" query with extensions shown for to a particular model before the apparition of a context. In future “restaurant”. work we would like to experiment with more complex cases and larger ETS. We are interested in using tags networks to produce Regarding the next query, “Eat near Lax”, we also start with the ontologies, rather than to only link to them. Indeed we believe what term, eat, which according to Wordnet is a verb, and that a tagging network can give some insight into the meaning of therefore, in our context, an affordance. It is linked to restaurant a concept, such as restaurant, in a given context, but and pub//bar, which are buildings, as well as to pizza, which is in independently of any task. our context an affordance (as it is not a geographical category). Note that here, as we are only looking for indications of KR 5. ACKNOWLEDGMENTS elements matching, and only have two categories, affordances Thanks to the anonymous reviewers which comments helped and geographical categories, ambiguity is easier to handle improving the quality the paper, and to Olga Streibel for support automatically. For example, if we had a query about drink, which and for the always interesting chats and discussions. can be a verb and a noun, it is not a geographic category, and therefore would be considered as an affordance. 6. REFERENCES [1] Sanderson, M., Kohler, J. 2004 Analyzing geographic queries. In Proceedings of SIGIR-GIR-2004, the workshop on geographical IR held at the 27th conference on research and development in information retrieval. [2] Himmelstein, M. 2005 Local Search: The Internet Is the Yellow Pages. Computer 38, 2 (Feb. 2005), 26-34. [3] Sanderson, M., Han, Y. 2007 Search words and geography. In Proceedings of the 4th ACM Workshop on Geographical information Retrieval (Lisbon, Portugal, November 09 - 09, 2007). GIR '07. ACM, New York [4] Smith, B., Mark D. 1999 “Ontology with Human Subjects Figure 4 "Eat near LAX" query Testing: An Empirical Investigation of Geographic Categories.” American Journal of Economics and Sociology This approach of using a theory of place to facilitate GIR is 58:245-272. related to [37] in which, for a representation of a place, a set of associations are requested from the users. However, ETS can be [5] Couclelis, H. 1992 People Manipulate Objects (but Cultivate used for various KR projection. For example if it was useful to Fields): Beyond the Raster-Vector Debate in GIS. In use categories and affordances for the notion of place, a way- Proceedings of the international Conference GIS - From finding task would probably need image schemata providing Space To Territory: theories and Methods of Spatio- structure to a path. This minimal amount of semantics, easy to Temporal Reasoning on theories and Methods of Spatio- capture, can improve the quality of local search by providing Temporal Reasoning in Geographic Space (September 21 - more semantically related words to search for. 23, 1992). A. U. Frank, I. Campari, and U. Formentini, Eds. Lecture Notes In Computer Science, vol. 639. Springer- Verlag, London, 65-77. 4. CONCLUSION AND FUTURE WORK Advantages of a technique based on differences expressed [6] Egenhofer, M., Mark, D. 1995 Naive Geography. In: Spatial through ETS over the choice of a particular KR model are Information Theory—A Theoretical Basis for GIS, multiple. First, by using differences, and therefore delaying any International Conference COSIT ’95, Semmering, Austria commitment to a KR model, one can easily adapt the KR (A. Frank & W. Kuhn, eds.), Lecture Notes in Computer elements used for a particular task. In this way we hope that ad Science, Vol. 988. Springer, Berlin, pp. 1-15. hoc knowledge based applications, developing ontologies in order [7] Vangenot, C., Parent, C., Spaccapietra, S. 2002 Modelling to handle a particular task only, will tend to disappear. Secondly, and Manipulating Multiple Representations of Spatial Data. by using tags, knowledge can be acquired very quickly, and future Proc. of the Symposium on Geospatial Theory, Processing work will investigate ways of acquiring part of the tag network and Applications, Ottawa, 2002. [8] Santos, D., Chaves, M. S. 2006 The place of place in [25] Raubal, M. 2004 Formalizing Conceptual Spaces Formal geographical IR, GIR’06, August 10, 2006, Seattle, Ontology in Information Systems, Proceedings of the Third Washington International Conference, FOIS 2004, A. Varzi and L. Vieu, Editors, pp. 153-164 [9] Guarino, N. 1998 Formal Ontology in Information Systems: Proceedings of the 1st International Conference June 6-8, [26] Ahlqvist, O. 2004 A Parameterized Representation of 1998, Trento, Italy. 1st. IOS Press., Uncertain Conceptual Spaces Transactions in GIS, Blackwell Synergy, pp. 493-514 [10] Tanasescu, V. 2007 Spatial Semantics in Difference Spaces, COSIT 2007, Melbourne, Australia . [27] Gangemi, A., Guarino, N., Masolo, C., Oltramari, A., Schneider, L. 2002 Sweetening Ontologies with DOLCE [11] Tanasescu, V., Streibel, O. 2007 Extreme Tagging: Emergent Proceedings of EKAW, Springer, pp. 166-181 Semantics through the Tagging of Tags, ESOE 07, Busan, Korea [28] Jordan, T., Raubal, M., Gartrell, B., Egenhofer, M. 1998 An Affordance-Based Model of Place in GIS 8th Int. [12] V. Cardoso, Nuno, Mario J. Silva. 2007 Query expansion Symposium on Spatial Data Handling, SDH, pp. 98-109, through geographical feature types. Proceedings of the 4th ACM workshop on Geographical information retrieval, pp. [29] Curry M. R. 1996 The Work in the World - Geographical 55-60. Practice and the Written Word, University of Minnesota Press, Minneapolis/London f [13] Delboni, Tiago M., Karla A. V. Borges, Alberto H. F. Laender. 2005 Geographic web search based on positioning [30] Lefebvre, H. 1991 The Production of Space. Blackwell expressions. Proceedings of the 2005 workshop on Publishers. Geographic information retrieval, pp. 61-64. [31] Agarwal, P. 2002 Contested Nature of Place: Knowledge [14] Santos, D., Chaves, M. S. 2006. The place of place in Mapping for Resolving Ontological Distinctions Between geographical IR, GIR’06, August 10, 2006, Seattle, Geographical Concepts. Geographic Information Science: Washington SLIDES Second International Conference, GIScience 2002, Boulder, CO, USA, September 25-28 [15] Tanasescu, V. 2006 Toward User Oriented Semantic Geographical Information Systems, 2nd AKT Doctoral [32] Bennett, B., and P. Agarwal. 2007 Semantic Categories Symposium, Aberdeen Univeristy, UK Underlying the Meaning of Place., COSIT 2007, Melbourne, Australia [16] Fonseca, F.T., Egenhofer, M.J. 1999 Ontology-driven geographic information systems. GIS '99: Proceedings of the [33] Abdelmoty, A. I., P. Smart, and C. B. Jones. 2007 Building 7th ACM international symposium on Advances in place ontologies for the semantic web: issues and geographic information systems, ACM Press, pp. 14-19 approaches. Proceedings of the 4th ACM workshop on Geographical information retrieval [17] Agarwal, P. 2005 Ontological considerations in GIScience International Journal of Geographical Information Science, [34] Kuhn, W. 2007 An Image-Schematic Account Of Spatial Taylor & Francis, 19, pp. 501-536 Categories, COSIT, 2007, Melbourne, Australia. [18] Winter, S. 2001 Ontology: buzzword or paradigm shift in GI [35] Sowa, J.F. 1992 Semantic networks, in: S.C. Shapiro (Ed.), science? International Journal of Geographical Information Encyclopedia of Artifcial Intelligence, second ed., Wiley, Science, Taylor & Francis, 15, pp. 587-590 New York, pp. 1493-1511. [19] Smith, B., Mark, D. 2003 Do Mountains Exist? Towards an [36] Young, R. M. 1968 Association of ideas. In Dictionary of the Ontology of Landforms Environment & Planning B: History of Ideas, Wiener, P. P. (ed.), Charles Scribner's Sons, Planning and Design, 30, pp. 411-427 New York, vol. 1, pp. 111-118. [20] Mark, D., Turk, A. 2003 Landscape Categories in [37] Edwardes, A. J. and Purves, R. S. 2007 Eliciting concepts of Yindjibarndi: Ontology, Environment, and Language Spatial place for text-based image retrieval. In Proceedings of the information theory: Foundations of geographic information 4th ACM Workshop on Geographical information Retrieval science, Springer, pp. 28-45 (Lisbon, Portugal, November 09 - 09, 2007). GIR '07. ACM, New York, NY, 15-18 [21] Vangenot, C., Parent, C. amd Spaccapietra, S. 2002 Modeling and manipulating multiple representations of [38] Henrich, A. and Luedecke, V. 2007 Characteristics of spatial data Proc. of the Symposium on Geospatial Theory, geographic information needs. In Proceedings of the 4th Processing and Applications ACM Workshop on Geographical information Retrieval (Lisbon, Portugal, November 09 - 09, 2007). GIR '07. ACM, [22] Gärdenfors, P. 2000 Conceptual spaces: the geometry of New York thought, MIT Press [23] Johnson, M. 1987 The body in the mind: the bodily basis of meaning, imagination, and reason University of Chicago Press, [24] Gibson, J. 1979 The Ecological Approach to Visual Perception, Lawrence Erlbaum Associates