When people talk about place, there are a few key parts of the produced sentences that are particularly useful for GIS tasks, spatial relations being the primary concern in this paper. Used to describe the position of an object in relation to another, spatial relations play a crucial role in the proper understanding of spatial communication. In the paper related to this analysis, a set of algorithms for generating polygons that match spatial relations used in daily communication are proposed. An experiment is then conducted, not only evaluating the proposed algorithms but also shedding some light on the way people understand spatial relations. This report is an analysis of the data collected in the experiment.
## options: GEOM_POSSIBLE_NAMES=geometry
## Reading layer `drawings' from data source `/home/jslucassf/workspace/mestrado/geoinfo/data/drawings.csv' using driver `CSV'
## Simple feature collection with 509 features and 6 fields
## geometry type: GEOMETRY
## dimension: XY
## bbox: xmin: -35.88474 ymin: -7.24181 xmax: -35.86695 ymax: -7.221268
## CRS: EPSG:4326## options: GEOM_POSSIBLE_NAMES=geometry
## Reading layer `drawings' from data source `/home/jslucassf/workspace/mestrado/geoinfo/data/drawings.csv' using driver `CSV'
## Simple feature collection with 509 features and 6 fields
## geometry type: GEOMETRY
## dimension: XY
## bbox: xmin: -35.88474 ymin: -7.24181 xmax: -35.86695 ymax: -7.221268
## CRS: EPSG:4326Through the usage of a web app, participants were told to picture the following scenario:
Imagine that a friend will give you a ride and tell you over the phone where the car stopped and is waiting. Based on the description he gave you, we ask you to draw on the map the area where you think the car might be.
The web app then shows up a map with a highlighted landmark and a sentence that describes the location of the car. Participants drew the regions by clicking on the map and creating points and lines. Each person had to draw five relations for each of the four landmarks.
The experiment was conducted in the Brazilian Portuguese language, so in order to better understand the analysis, here are the translations of the names of the spatial relations:
First, lets get a general view of the data. We’ll start by exploring the types of geometries drawn by participants.
As was expected, most people used a single polygon, and the relation that includes more multipolygons is AO LADO DE, as it can include the left and right side of the landmark. One interesting finding is that there are some lines. Only 2 of the participants chose to draw lines.
Now lets visualize the data in maps.
Let’s examine drawings by area
As expected, the relations BETWEEN and AT STREET - NEAR contain the drawings with more area. Lets have a closer look at the really big ones.
It seems that there is nothing weird about them.
Our assumption is that for landmarks located at street corners, participants will draw more multipolygons, since they might feel uncertain to which of the streets correspond to the relation.
The usage of geometry types seems to be the same. Now let’s examine by relation.
This pattern continues for all the relations. We didn’t expect to see many multipolygons for the relation IN FRONT OF in non corner places. Our assumption was that some people wouldn’t know the landmark, and therefore not know to which street its facade was turned to. This would led them to draw more than one polygon, one for each street.
Turns out our assumption was partially correct, for this also happened for non corner places. We suppose that this could be due to proximity to the streets, a more thorough analysis needs be done to allow any conclusion on the subject.
What exactly does one mean, when is said that one object is “to the right” of another? To the right might be interpreted as at the rightside from the object’s perspective, or from the perspective of an observer, that is positioned in front of the object. The former interpretation, uses a Frame of Reference (FoR) named “Intrinsic”, while the latter uses the “Deictic”.
We are interested in finding out, which of these two FoRs participants had in mind while drawing the relations and we can do this by analysing the relation “TO THE RIGHT OF”.
We do this by comparing the drawings with two buffer geometries. One that encompasses the right side of the landmark, and another for the left side.
We assume the egocentric FoR for our geometries, thus, if a drawing is found to be in the right buffer, it means that the reasoning of the participant fits in the deictic category. If however the drawing is in the left buffer, the intrinsic FoR better represents the participants’ thought process.
## Reading layer `OGRGeoJSON' from data source `{"type":"Polygon","coordinates":[[[-35.8698164543337,-7.23354689555164],[-35.8700657719415,-7.23375171630509],[-35.870310901112,-7.23345415553491],[-35.8700615836631,-7.2332493348928],[-35.8698164543337,-7.23354689555164]]]}' using driver `GeoJSON'
## Simple feature collection with 1 feature and 0 fields
## geometry type: POLYGON
## dimension: XY
## bbox: xmin: -35.87031 ymin: -7.233752 xmax: -35.86982 ymax: -7.233249
## CRS: EPSG:4326## Reading layer `OGRGeoJSON' from data source `{"type":"Polygon","coordinates":[[[-35.8700427811703,-7.23377962463979],[-35.8697934635476,-7.23357480387589],[-35.8695483338933,-7.23387236452473],[-35.8697976516749,-7.23407718539997],[-35.8700427811703,-7.23377962463979]]]}' using driver `GeoJSON'
## Simple feature collection with 1 feature and 0 fields
## geometry type: POLYGON
## dimension: XY
## bbox: xmin: -35.87004 ymin: -7.234077 xmax: -35.86955 ymax: -7.233575
## CRS: EPSG:4326For the Niscar landmark, most drawings (55%) are on the left buffer, which represents the intrinsic FoR.
One interesting observation is that many (25%) participants chose to draw on both sides.
## Reading layer `OGRGeoJSON' from data source `{"type":"Polygon","coordinates":[[[-35.87940238416194,-7.223747285824582],[-35.879239439964294,-7.224033334573388],[-35.878924280405045,-7.223838422306055],[-35.87908521294594,-7.223549712513339],[-35.87940238416194,-7.223747285824582]]]}' using driver `GeoJSON'
## Simple feature collection with 1 feature and 0 fields
## geometry type: POLYGON
## dimension: XY
## bbox: xmin: -35.8794 ymin: -7.224033 xmax: -35.87892 ymax: -7.22355
## CRS: EPSG:4326## Reading layer `OGRGeoJSON' from data source `{"type":"Polygon","coordinates":[[[-35.87943322956562,-7.223757264272369],[-35.879783257842064,-7.223936211065018],[-35.87960354983806,-7.224271486649396],[-35.87926559150219,-7.224051960997253],[-35.87943322956562,-7.223757264272369]]]}' using driver `GeoJSON'
## Simple feature collection with 1 feature and 0 fields
## geometry type: POLYGON
## dimension: XY
## bbox: xmin: -35.87978 ymin: -7.224271 xmax: -35.87927 ymax: -7.223757
## CRS: EPSG:4326For the second landmark (Café Poético) the values are closer to each other (7 drawings in the left buffer and 8 in the right one).
## Reading layer `OGRGeoJSON' from data source `{"type":"Polygon","coordinates":[[[-35.87012529373169,-7.23396975742232],[-35.8697310090065,-7.233635819466916],[-35.87004616856575,-7.233251324861703],[-35.87044849991798,-7.233582602240051],[-35.87012529373169,-7.23396975742232]]]}' using driver `GeoJSON'
## Simple feature collection with 1 feature and 0 fields
## geometry type: POLYGON
## dimension: XY
## bbox: xmin: -35.87045 ymin: -7.23397 xmax: -35.86973 ymax: -7.233251
## CRS: EPSG:4326## Reading layer `OGRGeoJSON' from data source `{"type":"Polygon","coordinates":[[[-35.87009847164154,-7.23399104429556],[-35.86986780166626,-7.234271764842552],[-35.86948022246361,-7.233923192383596],[-35.869717597961426,-7.233649123772656],[-35.87009847164154,-7.23399104429556]]]}' using driver `GeoJSON'
## Simple feature collection with 1 feature and 0 fields
## geometry type: POLYGON
## dimension: XY
## bbox: xmin: -35.8701 ymin: -7.234272 xmax: -35.86948 ymax: -7.233649
## CRS: EPSG:4326For the third landmark, the most drawings are in the left buffer.
## Reading layer `OGRGeoJSON' from data source `{"type":"Polygon","coordinates":[[[-35.879565328359604,-7.223766577490084],[-35.87934136390685,-7.224131123290164],[-35.879019498825066,-7.223914258488325],[-35.87923809885979,-7.223569669417386],[-35.879565328359604,-7.223766577490084]]]}' using driver `GeoJSON'
## Simple feature collection with 1 feature and 0 fields
## geometry type: POLYGON
## dimension: XY
## bbox: xmin: -35.87957 ymin: -7.224131 xmax: -35.87902 ymax: -7.22357
## CRS: EPSG:4326## Reading layer `OGRGeoJSON' from data source `{"type":"Polygon","coordinates":[[[-35.879617631435394,-7.2237945171421085],[-35.879974365234375,-7.224023356131708],[-35.8797624707222,-7.224405197676178],[-35.87940037250519,-7.224165715212135],[-35.879617631435394,-7.2237945171421085]]]}' using driver `GeoJSON'
## Simple feature collection with 1 feature and 0 fields
## geometry type: POLYGON
## dimension: XY
## bbox: xmin: -35.87997 ymin: -7.224405 xmax: -35.8794 ymax: -7.223795
## CRS: EPSG:4326In the last landmark, the left buffer again has the majority of drawings, but by a smaller difference.
Upon examining the combination of results, in most of the cases drawings of the relation “RIGHT OF” intersected only with the left buffer, therefore the intrinsic FoR seem to be used more often by participants. However, the higher intersection with the right buffer for both Café Poético and Maria Pitanga raises questions. Might this result be affected by the geographical direction towards which the landmarks are facing (Both Café Poético and Maria Pitanga are neighbors and face the same direction)? What happens if we factor in people’s problems in differencing left from right or even map reading difficulties?
Another possible future work, would be to repeat this experiment, but showing participant’s the actual images of landmarks facades.
In the paper associated with this analysis we propose algorithms for automatically generating geometries representing different spatial relations. Now we try to evaluate their performance by comparing them to the drawings made by participants of the experiment.
Here we verify if each drawing intersects the corresponding area (same landmark and relation) produced by the implementation of the proposed algorithms.
The geometries generated by the algorithm intersect more than 50% of the drawings in all relations, with the exception of the allocentric version of “TO THE RIGHT”.
During data collection, participants were told to draw the region they expected a car to be in, after reading a describing sentence. However some participants, produced areas in regions that are not streets.
The proposed algorithms, work in the scope of streets. So now we remove the drawings that do not intersect with streets.
If we now only consider the drawings that intersect streets, the percentage of drawings that our algorithm generated areas intersect is bigger for all spatial relations.
A common metric used to access the similarity between sets is the Jaccard’s Similarity Coefficient. It expresses how similar two sets are in a scale of 0 to 1.
Jaccard (A, B) = Area Of Intersection(A, B) / Area Of Union(A, B)
We can use this metric, to evaluate how similar are the geometries produced by the algorithms and the drawings made by the participants.
If we again consider only the drawings that intersect the streets, we can now crop the drawings and discard the portions that do not intersect streets to make a more fair comparison of areas.
Comparing the areas with only the parts of drawings that intersect streets yields better results even though most similarity values continue to be bellow .5, the exception being the Between relation, this can be explained by the fact that the region produced by the algorithm is large, for this reason it also intersects all the drawings in the same relation. This might be an indicative of how hard this problem is.
Lets examine the median similarity amonst the drawings themselves. To do this, we separate drawings by landmark and relation and then for each drawing, compute the jaccard similarity index between it and all the others in the same relation and landmark. The inner jaccard is the median value of these indexes.
In fact, we can observe that the drawings made by different participants are not very similar. In the future, different experiment setups might allow for a better use of this metric.
Nevertheless, the similarity indexes coupled with the intersection percentages show that the proposed algorithms represent a good approach to the implementation of spatial relations that are in accordance with the way people understand spatial language.