In previous work, di erent methods have been proposed in order to semi-automatically mine geospatial information and opinions in documents [3]. In this paper, we present the Web application, SentiAnnotator, based on NLP methods to extract and visualize geospatial information with the associated entities. The evaluation of our application shows good results on a French corpus, i.e. F-measure of 0.74 and 0.75 respectively for the identi cation of spatial features and organizations.
Researchers and experts of land-use planning are looking for decisional tools
for helping them to have an overview of user's awareness on territories. In this
context, we de ned the Opiland method that enables to semi-automatically
analyze sentiments related to land-use planning documents [
In this demo paper, we present the SentiAnnotator web application4 to extract di erent features related to land-use planning. Hereafter, in Section 2, we focus more precisely on the deployment of natural language processing methods to extract geospatial information, i.e. spatial features and organizations. The web application for uploading, indexing, and marking textual documents is detailed in Section 3. Finally, after a quick look to our system evaluation in Section 4, future work related to our project is drawn in Section 5. 2
Named Entity Recognition (NER) methods identify di erent types of Named
Entities (NE): dates, people, organizations, themes, numeric values, as well as
locations. There is a signi cant number of available systems, such as OpenNLP5,
OpenCalais6, and CasEN [
Moreover we propose a new type of patterns to identify Organizations (OE) which is a speci c NE useful for land-use planning domain. The addition of speci c rules enables to identify OE which could be confused with SF in documents. Such rules are: (1) an OE is followed by an action verb; (2) an OE is proceeded by prepositions: with, by, for, on behalf of, etc.
In order to manage these geospatial information, we developed the web application SentiAnnotator (http://siso.teledetection.fr/viewer.jsp). A screenshot is presented Figure 1. The web services use the Gate system7. After uploading a corpus (in French for this current version), Spatial Features and Organizations are extracted using the implemented rules. Moreover other concepts are
A webtool for analyzing land-use planning documents
extracted using Gate: (i) thematics based on a lexicon using Agrovoc thesaurus8,
(ii) Opinions related to land-use planning domain [
The web application (Figure 1) allows users to upload corpora, to index documents with speci c web services in order to mark di erent kinds of information (spatial features, organizations, opinions, and themes), to visualize, to correct the results, and to download validated results in XML format. More speci cally, it is possible to upload corpora (frame 1), each marked corpus is saved on the server and automatically available in the web application (frame 2). After having downloaded documents, users can select the marked features (frame 5), see the results on the selected documents in frame 3. In this frame, spatial features are in blue color, organizations in purple color, the positive opinions in green color, negative ones in red color and neutral in yellow color. By selecting di erent categories from frame 5, the related marked information will be kindles in frame 3 and listed by type in frame 4. In case of nding any mistakes, users can unselect marked information (frame 4). Finally expert can export the selected corrected documents by clicking the top right bottom. The downloaded corpus consists of selected documents with the marked information except those were removed by the user. The administration page allows users to upload, edit, and delete pipelines de ned in the Gate format. It is also possible to remove processed corpora and to edit the uploaded pipeline rules and the available lexicons.
Farvardin et al.
Three experts of the project evaluated the process for extracting geospatial information by using SentiAnnotator application. We use a French corpus composed of 4328 words (71 spatial features and 117 organizations). The evaluations (with classical measure, i.e. Precision, Recall, and F-measure) have been investigated by comparing the manual extraction done by experts with the web service results. For SF, we obtain an excellent recall (0.91) and an acceptable precision (0.62), the F-measure is 0.74. We extract the great majority of SF but the rules still return some errors. The rules to identify OE are very e cient and return high precision (0.85) but the value of recall is lower (0.67). The F-measure for organization identi cation is 0.74. The rules for organization extraction seem well-adapted to the domain but they have to be extended in order to improve the recall that remains low. 5
In this paper, we have presented a Web application called SentiAnnotator including web services (1) to annotate corpora with features related to landuse planning, and (2) to evaluate achieved approaches with experts. Experts are using this tool for analyzing the construction project of a road around Villeveyrac (France). Future work will be dedicated to the improvements of the de ned linguistics patterns for discovering NE in order to tackle the issues related to the land-use planning speci cities and the multilingual aspects. We also plan to extend our approach to di erent types of textual contents such as tweets.
Acknowledgments: The authors thank Midi Libre (French newspaper) for its expertise on the corpus and all partners of the Senterritoire project for their involvement (MSH-M, Geosud Equipex, Numev Labex, and Tectoniq PEPS project).