In natural language, people often describe location using locative expressions [1] that include three key elements: the place or object the location of which is being described (known as the locatum or trajector), a place which is the reference object (known as the relatum or landmark) and a relationship between these two (spatial relation term, often a preposition), e.g., theatre [locatum] near [spatial relation term] Trafalgar Square [relatum]. Koja 1 is an interactive web map which retrieves a set of candidate objects that correspond to the described locatum type, using machine learning to predict the distance between the locatum and relatum based on a range of linguistic and contextual features from the three elements in a given expression.
Web maps have been created to answer a wide range of research questions. Environmental applications such as ClimateCharts, which provides an interactive web map showing temperature and precipitation of most places on earth [2], are common, and enable researchers to explore various physical interactions. Web mapping applications that incorporate individual perspectives have also been developed. For example, eImage which focuses on the selection of an area based on user interests to understand how people feel about different areas [3]. However, web maps that allow users to find the location of places in response to a query rarely take account of vague and context sensitive terms, and in production systems such as Google Maps2 and OpenStreetMap3 , places are most commonly found simply by their name, or a type of feature (e.g., hairdressers). While it is possible to include spatial relation terms (e.g., in, near, outside, north of) in Google Maps searches, they have little impact on the results returned. There have been many studies of the use of spatial relational language, e.g., [1,4,5,6,7,8,9], some of which focus specifically on contextual factors, e.g., [10]. The concept of spatial templates was introduced in [6] where the template models the applicability or acceptance of locations relative to a reference object, for a given spatial relation. Only limited progress has been made in automating the interpretation of spatial language in order to model or predict the locations referred to in geo-spatial locative expressions e.g. [11,12,13], though some progress has also been made in 'table-top' space, such as [14] in which spatial templates were modelled with deep learning. Some notable examples of considering uncertainty in the interpretation of multiple spatial relationships in locative expressions include [15,16,17] which focus on the geometric characteristics of objects, with limited attention to the geographical types. In the GeoLocate4 service complex locality descriptions are interpreted by determining displacements from a reference object (relatum) using a mainly rule-based approach, but it does not take context into account and is limited in the range of spatial relations that can be interpreted. In an early use of multiple spatial templates to predict geo-spatial location [13] predicted location but did not consider feature types and only distinguished between urban and rural contexts, while [12] interpreted relative location descriptions taking account of the contextual factors of place size and prominence. Logan et al. [6] predicted distance and direction based on spatial relation term and relatum, using contextual factors but did not consider an explicit locatum. An example of an experimental search system interface that allows the user to specify one of a small set of spatial relations is that of [18] in which a search buffer adapts to the size of the reference object, taking account of predetermined parameters relating to the locatum and relatum types, and enabling search expansion if no object is found initially. The method adopted here differs from previous studies in retrieving locations of candidate locata, for a query that specifies locatum type, spatial relation term selected from a fairly large number of options, and reference object, based on a regression model that is trained on the combination of spatial relation and context features.
As an interactive web map, Koja uses location and language features in a machine learning model to predict the distance between locatum and relatum for particular spatial relational terms, in order to select a set of possible features of specified type that lie within a range of the inferred distance from the relatum. It is intended as a simple prototype for future spatial search systems that attempt to interpret the meaning of relative spatial relational terminology in a query. We used a regression model for this prediction, with SMO a regression version of support vector machine (SVM). The training data contains about 700 location descriptions extracted from Geograph 5 and Foursquare6 websites which include locatum, geospatial preposition and relatum. For a structured query that specifies locatum type, spatial relation and a named relatum, the derived input features for the regression model include the following:
• GloVe embeddings 7 of the feature types of locatum and relatum.
• The characteristics of the locatum and relatum including geometry type (point, line, polygon), scale, area, elongation and whether it is liquid or solid. These characteristics can affect the interpretation of a spatial relation term (e.g., "church beside post office" vs "road beside river"). • The density of objects in the area and surrounds.
• The semantics of the spatial relation term based on matching the query spatial relation with spatial relations in training data. We exploit semantic similarity of spatial relation terms to support improved learning from training data with similar spatial relations (e.g., "beside", "next to").
The spatial relation term in combination with the other features are used by the classifier to infer a distance which will vary for an individual spatial relation depending on the types of feature (e.g., "park bench near post office" vs "village near London") [19]. The distance is converted to a buffer zone to retrieve objects of the specified locatum feature type that lie inside that zone. Koja supports search based on the specified feature type of the locatum that is selected from OpenStreetMap types, a choice of 24 different spatial relation terms (e.g. "near", "on", "at", "beside", "opposite") [20] and, for the reference objects, named places in the City of London, UK. It is implemented with Leaflet with OpenStreetMap as a base map, and uses models trained on a set of <locatum-spatial relation term-relatum> triples extracted from Geograph and Foursquare.
The interface allows the user to select the relatum by typing the first three characters of the name and then selecting from the available places, and then to select the spatial relation term and locatum type from a drop down list (again, the initial letters can be typed in to narrow the selection). Once these values are selected and the query submitted, the system determines the input features and runs them against the model created from training data to predict the distance that best reflects the specific context of those three terms.
Figure 1 shows the results of the query theatre near Trafalgar Square. The doughnut shape shows the area that the locatum might be located in, calculated using the distance predicted by the machine learning model for the expression (in this case, 366.86m), with a margin of error on either side (based on the mean absolute error divided by average of all distances across all the expressions in the dataset expressed as a percentage). The orange circles identify the theatres that fall within the doughnut, and thus that may be considered to fulfil the query theatre near Trafalgar Square. As can be seen, the distance that is predicted for the expression using at (163.66m) is much smaller than for that using near, since the preposition at usually refers to a closer distance than near.
Figure 3 shows results of the query theatre near St-Martin-in-the-Fields, illustrating the effect of different context. St-Martin-in-the-Fields is a church, quite close to Trafalgar Square, but the predicted distance is 149.20m, much smaller than for Trafalgar Square. This is because the characteristics of the relatum, which is a church (St-Martin-in-the-Fields) differ from the characteristics of the relatum of Trafalgar Square (in Figure 1), which is a square.
The Koja web mapping application offers an interactive interface for structured spatial queries that incorporates context into the interpretation of vague spatial relation terms, providing distance predictions that use machine learning models with multiple semantic and linguistic input features. It was trained using examples of georeferenced locative expressions from the volunteered data sources Geograph and Foursquare. This provides some significant advances beyond current web mapping applications that either do not cater for spatial relation terms, or use much simpler models for them.
In future work, the authors plan to work with unstructured queries and to further advance the models of spatial language by exploiting deep learning approaches, as well as using richer sets of training data that reflect the geographic context that humans use when formulating natural language location descriptions.