User-contributed content represents a valuable information source provided one can make sense of the large amounts of unstructured data. This work focusses on geospatial content and specifically on travelblogs. Users writing stories about their trips and related experiences e ectively provide geospatial information albeit in narrative form. Identifying this geospatial aspect of the texts by means of applying information extraction techniques and geocoding, we relate portions of texts to locations, e.g., a paragraph is associated with a spatial bounding box. To further summarize the information, we assess the opinion (“mood”) of the author in the text. Aggregating this mood information for places, we essentially create a geospatial opinion map based on the user-contributed information contained in the articles of travelblogs. We assessed the proposed approach with a corpus of more than 150k texts from various sites.
Crowdsourcing moods and in our specific case opinions from user-contributed data,
has recently become an interesting field with the advent of micro-blogging services
such as, e.g., Twitter. Here, blog entries reflect a myriad of di erent user opinions that
when integrated can give us valuable information about, e.g., the stock market [
The specific contributions in this work are as follows. In an initial stage several travelblog Web sites have been crawled and over 150k texts have been collected. Figure 5 shows such an example travelblog entry. The collected texts are then geoparsed and geocoded to link placename identifiers (toponyms) to location information. With paragraphs as the finite granularity for opinion information, texts are then assessed with the OpinionFinder tool and are assigned a score for each paragraph ranging from very negative to very positive. Scores are linked to the bounding box of the paragraph and are aggregated using a global grid, i.e., the score of a specific paragraph is associated with all intersecting grid cells. Aggregation of opinions is then performed simply by computing the average of all scores for each cell. Finally, the score can be visualized by assigning colors to each cell.
While to the best of our knowledge there exists no work aiming at extracting
opinions about places from travel blogs, we can cite the following related work. The
concept of information visualization using maps is gaining significant interest in various
research fields. As examples, we can cite the following works [
The remainder of this work is organized as follows. Section 2 describes the information extraction techniques employed in our approach dealing specifically with the aspects of geoparsing and geocoding travel blog entries. Section 3 outlines a method for computing user sentiment scores from travel blog entries. Section 4 outlines how such scores can be aggregated and visualized based on geospatial locations. In addition some specific examples are shown to give an initial validation of the proposed approach. Finally, Section 5 presents conclusions and directions for future work. 2
In what follows, we describe in detail the processing pipeline, which overall uses an HTML document as input (travel blog article) and produces a structured XML file containing the various entities and their respective attributes (toponyms and coordinate information for the text).
The pipeline consist of four parts (cf. Figure 1), (i) the HTML parsing module, (ii) the linguistic pre-processing, (iii) the main IE engine system (semantic analysis) and (iv) the geocoding-postprocessing part. In the next section, we describe the first part of our processing pipeline, i.e., the collection of HTML texts, the parsing and their conversion to plain text format, in order to prepare the documents for the forthcoming step of linguistic-preprocessing. For collecting travel blog articles containing rich geospatial information, we crawled Web sites providing traveblog authoring services. Each Web site has its own HTML layout and isolating text of interest from crawled and parsed HTML pages is done by hand. Thus, there was a need for Web sites with massive amounts of such type of documents. For this purpose we crawled travelpod.com, travelblog.org, traveljournals.net and worldhum.com, resulting in more than 150,000 documents. For crawling the web sites, we used Regain crawler1, which creates a Lucene2 index for indexing the documents’ information, while for HTML parsing and the extraction of useful plain text narratives, we used Jericho HTML parser3.
2.2
To prepare the input for the core IE engine for extracting objects of interest, the parsed plain text documents must be prepared accordingly. Such preparation includes linguistic pre-processing tools that analyze natural language documents in terms of distinct base units (i.e., words), sentences, part-of-speech and morphology . We are using the ANNIE tools, contained in the GATE release4, to perform this initial part of analysis. To this task, our processing pipeline comprises of a set of four modules: (i) the ANNIE tokenizer, (i) the (ANNIE) Sentence Splitter, (iii) the ANNIE POS Tagger and (iv) the WordNet Lemmatiser.
The intermediate processing results are passed on to each subsequent analysis tool
as GATE document annotation objects. The output of this analysis part is the analyzed
document and it is transformed in CAS/XML format5, which will be passed to the
subsequent semantic analysis component as input, Cafetiere IE engine [
The first step in the pipeline process is tokenization, i.e., recognizing in the input text basic text units (tokens), such as words and punctuation and orthographic analysis and the association of orthographic features, such as capitalization, use of special characters and symbols, etc. to the recognized tokens. The tools used are ANNIE Tokenizer and Orthographic Analyzer.
Sentence splitting, in our case the ANNIE sentence splitter aims at the identification of sentence boundaries in a text.
Part-of-speech (POS) tagging is then the process of assigning a part-of-speech class,
such as Noun, Verb etc. to each word in the input text. The ANNIE POS Tagger
implementation is a variant of Brill Transformation-based learning tagger [
Lemmatisation is used for text normalisation purposes. With this process we retrieve the tokens base form e.g., for words: [travelling, traveler, traveled], [are, were], the corresponding lemmas are: travel, be. We exploit this information in the semantic rules section. For this purpose we implement the JWNL WordNet Java Library API6 for accessing the WordNet relational dictionary. The output of this step is included it in GATE document annotation information. 2.3
Semantic analysis relates the linguistic pre-processing results to ontology information, as we will see in the next subsection about ontology lookup and applies semantic analysis grammar rules, i.e., documents are analyzed semantically to discover spatial concepts and relations.
5 CAS is an XML scheme called Common Annotation Scheme allowing for a wide range of annotations, structural, lexical, semantic and conceptual. 6 http://sourceforge.net/projects/jwordnet/
For this purpose we used Cafetiere IE engine, whose objective is to compile a set of semantic analysis grammar rules in a cascade of finite state transducers so as to recognize in text the concepts of interest. Cafetiere IE Engine combines all previously acquired linguistic and semantic information with contextual information. We modified Cafetiere and implemented it as a GATE pipeline module (GATE creole) for the purpose of performing ontology lookup and rule-based semantic analysis on information acquired from previous pipeline modules, in the form of GATE annotation sets. The input to this process are the GATE annotation objects resulted from the linguistic preprocessing stage stored transformed in Cafetiere needed format, in CAS/XML format for each individual document.
Cafetiere Ontology Lookup The use of knowledge lexico-semantic resources assists in the identification of named entities. These semantic knowledge resources may be in the form of lists (gazetteers) or more complex ontologies providing mappings of text strings to semantic categories, such as in general male/female person names, known organizations and known identifiers of named entities. In our case, the named entities we want to extract with IE methods are location based information. For example, a gazetteer for location designators might have entries such as “Sq.”, “blvd.”, “st.” etc. that denote squares, boulevards and streets accordingly. Similarly there are various sorts of gazetteers available for given person names, titles, location names, companies, currencies, nationalities etc. Thus, the named entity (NE) recognizer can use gazetteer information so as to classify a text string as denoting an entity of a particular class. However, in order to associate specific individual entities object identifiers are required as well as class labels, enabling aliases or abbreviations to be mapped to a concrete individual. For example, for an entity such as “National Technical University of Athens” the respective abbreviation “NTUA” could be included in the knowledge resource as an alias for the respective entity. Thus, more sophisticated knowledge resources than plain gazeteers in the form of ontologies may be used to provide this type of richer semantic information and allow for the specification and representation of more information, if necessary, than identity and class inclusion.
In this way, Cafetiere Ontology lookup module accesses a previously built ontology to retrieve potential semantic class information for individual tokens or phrases. All types of conceptual information, related to domain specific entities, such as terms or words in general that denote spatial concepts or properties and relations of domain interest are pre-defined in this ontology. For example, consider the partial ontology shown in Figure 2. Class “LOCVERB” stores verbs that when matched to a text phrase are likely to indicate a spatial relationship between the corresponding referenced concepts. We label as semantic any classification of tokens according to their meaning in the field of the application, in our case, geosemantics. This could be done, on a broad coverage level, by reading information from a comprehensive resource such as WordNet lexicon about most content words. However, the practice in information extraction applications as discussed in previous paragraph, has been to make the processing application-specific by using lists of the semantic categories of only relevant words and phrases, done by hand. The ontology used in our experimentation was created by manually analyzing a large number of texts and iteratively refining the ontology with words (e.g., verbs) that when matched to a text phrase are likely to indicate a spatial relationship between the corresponding referenced concepts. Summarizing, the lookup stage of analysis: 1. Supplies semantic classes (concepts) corresponding to words and phrases. 2. Supplies object identifiers for known instances, including where aliases and abbreviations name the same instance (For example “National Technical University of Athens”, “NTUA”). 3. Supplies properties of known instances, for example the country of which a city is the capital. 4. Uses verbs of interest to the application in order to identify inside the phrase potential unknown instances.
tion with IE methods can be divided into two main categories:
– Linguistic/rule-based approaches: in these approaches the Named Entity
recognition is based on linguistic/semantic rules defining the possible linguistic patterns
denoting Named Entity concepts, such as for example the approaches adopted by
ANNIE7, and Cafetiere [
– Machine learning/statistics-based approaches: these approaches view Named Entity recognition as a classification problem and, they have gained increased popularity due to their relatively rapid development/ domain customization and the reduced amount of human e ort required.
Cafetiere is a rule-based system for IE. A set of linguistic patterns (i.e., extraction rules) is written taking into account the lookup ontology and all previously acquired information from linguistic pre-processing. The semantic analysis rules, are developed as a set of context-sensitive/context-free grammar (CSG/CFG) rules and are compiled in a cascade of finite state transducers so as to recognize the concepts of interest in plain texts. 2.4
In this part, all information regarding each object of interest for each document in the collection is imprinted as a GATE annotation set object. For each document, we have collected information about all extracted entities, along with their respective paragraph, sentence and character o set in this document. During the HTML parsing process we keep the scope that each document is referred to in order to use this information for geocoding each extracted entity. For geocoding, we initially implemented YAHOO! Placemaker8 and used in combination with Cafetiere’s output, in order to deliver better results. We observed that PlaceMaker worked well for disambiguating some entities, but it identified significant fewer place entities than our IE engine. Thus, in the remaining entities extracted by Cafetiere, we applied YAHOO! Placefinder9 to geocode this place information passing the scope information described below for delivering more accurate results.
Finally, for each HTML travel blog entry (narrative), we created a collection of extracted referred geo-entities, some of them not being able to geocode. For each of these entities there is specific information (acquired from each of the previous pipeline steps) about where they were encountered in the respective document, namely, sentence, paragraph and o set character. Additionally, for each document, we calculated the mean coords and standard distance from all geocoded points extracted. All this information, along with the local parsed text file path and the respective URL of the document, are stored lastly into XML format for each corresponding plain text narrative. Samples of plain text narrative and the corresponding structured XML file are shown in Figure 3 and Figure 4 respectively. The XML tags in Figure 4 are denoting either statistical information, like the mean center and the standard distance of all geocoded locations for each document, or information related with each extracted entity, i.e., the o set characters, the sentence and paragraph ID. 3
Having geocoded the travel blog entries, we, in the following step, want to assign
sentiment information (“mood”) to text. To this e ect, we use OpinionFinder [
tem that performs subjectivity analysis, automatically identifying when opinions, sentiments, or speculations are present in text. It aims to identify subjective sentences, as also marking various aspects of subjectivity in these sentences, including the source (holder) of the subjectivity and words that are included in phrases expressing positive or negative sentiments.
OpinionFinder operates as one large pipeline. Conceptually, the pipeline can be divided into two parts. The first part performs mostly general purpose document processing (e.g., tokenization and part-of-speech tagging). The second part performs the subjectivity analysis. The results of the subjectivity analysis are returned to the user in the form of SGML/XML markup of the original documents.
For the first part, OpinionFinder takes any incoming text source and removes HTML
or XML meta info. Sentences are split and POS tagged using OpenNLP10, the open
source solution providing a variety of java-based NLP tools which perform sentence
detection, tokenization, pos-tagging, chunking and parsing, named-entity detection, and
coreference using the OpenNLP Maxent machine learning package. Next, stemming
is accomplished using Steven Abneys’ SCOL v1K stemmer program11. SUNDANCE
(Sentence UNDerstanding And ConceptExtraction) [
At this point, for the second part, a Naive Bayes classifier identifies subjective
sentences. The classifier is trained against subjective and objective sentences generated by
two additional rule-based classifiers drawing from large corpora [
OpinionFinder produces sentiment information assigned to paragraphs of texts. In the following, we describe how this information can be aggregated for specific locations. 4.1
OpinionFinder was applied to all texts of our collection of 150k travel blog entries assigning sentiment data to each paragraph of the collection. In the analysis that follows, only paragraphs containing geospatial data were retained. For each of these paragraphs we keep the total referred positive and negative sentiment scores as computed by OpinionFinder.
Each paragraph contains zero, one or multiple geographic entities that were suitably geocoded. In order to show the spatial extent of a paragraph, we chose to spatially visualize only paragraphs in which the MBR of the contained toponyms does not exceed 0.5 degrees in either dimension (e.g., Max latitude Min latitude 0:5 AND Max longitude Min longitude 0:5). Consequently, only paragraphs of limited and focused spatial extent are visualized, thus preventing paragraphs that refer to larger geographic entities (e.g., Europe) to dominate in the results.
We used five di erent categories for mapping opinion scores. The categories and respective color are given in Table 1, where each category scales from negative (red) to positive (green).
The proposed approach is clarified by the following example. A sample document14 (Figure 5) contains several paragraphs mentioning Washington D.C. and its landmarks. For each of this document’s paragraphs, a MBR covering the discovered toponyms was created and each paragraph was assigned a category according to Table 1. Therefore, this document may be spatially visualized on a map as shown on Figure 6.
Although this approach is viable when there is a limited number of documents and paragraphs, we need to overcome the following problem. Multiple paragraphs from different documents and di erent scores may partially target the same area, e.g., we need 12 http://wordnet.princeton.edu/ 13 http://nrrc.mitre.org/NRRC/publications.htm. 14 http://www.travelpod.com/travel-blog-entries/drfumblefinger/1/1269627251/tpod.html to visualize partially overlapping MBRs with di erent scores (colors). To do that, we split each paragraph MBR into small cells of a regular grid of 0.0045 degrees (corresponding to 500m) in each dimension. For each of those cells we sum up the sentiment score from all the containing paragraph MBRs. With this approach, instead of trying to visualize overlapping paragraph MBRs with di erent scores (colors), we visualize distinct small cells with each being assigned a unique score (and color). Consequently, it is easy to visualize the overall sentiment scores independent of how many paragraphs target the same area. 4.2
Further examples shown in the following include the geospatial opinion map of Amsterdam of Figure 7. It is interesting to observe that while most of the city is shaded green, the area around the train station and the Red Light district are shown in red, i.e., expressing rather negative sentiment.
Figure 8 gives a geospatial opinion map of Central Europe indicating the areas mentioned in the travel blogs. What can be observed is that positive sentiments are associated with areas in Switzerland and also Italy, while urban areas such as Brussels overall attract more negative sentiments. 4.3
Our initial experiments with the creation of geospatial opinion maps derived from subjective travelblog entries show that there is a clear bias for certain geographic areas shared by people. However, since in this work we only performed a simple aggregation of the scores generated by the OpinionFinder tool, it will require more in-depth analysis of the results to generate accurate statements and trends. Aggregating opinions is important for utilizing and assessing user-generated content. This work provides a means of visualizing sentiments for specific geographic areas as derived from travel blog entries. To demonstrate the approach, several travel blog sites were crawled and a total of more than 150,000 pages/articles were processed. Using (i) geoparsing and geocoding tools the content was geo referenced and (ii) sentiment information was derived using the OpinionFinder tool. In the proposed approach, sentiment information from various articles relating to the same geographic area is aggregated and visualized accordingly by means of a geospatial heat meat. Directions for future work are as follows. The current approach for aggregating user sentiment for geographic areas is rather simple and a more in-depth analysis of the results is needed to generate accurate statements and trends. An obvious improvement will also be to examine/include microblogging content streams. Here, sentiment information will be updated live and thus represent an accurate picture of the situation of a specific geographic area over time. Finally, OpinionFinder is a general purpose tool for deriving user sentiment. More involved approaches exist and need to be examined/developed for the case of geospatial data.
The research leading to these results has received funding from the European Union Seventh Framework Programme - Marie Curie Actions, Initial Training Network GEOCROWD (http://www.geocrowd.eu) under grant agreement No. FP7-PEOPLE-2010-ITN-264994.