Content-location relationships: a framework to explore correlations between space-based and place-based user-generated content

Abstract

The use of social media and location-based networks through GPS-enabled devices provides geospatial data for a plethora of applications in urban studies. However, the extent to which information found in geo-tagged social media activity corresponds to the spatial context is still a topic of debate. In this article, we developed a framework aimed at retrieving the thematic and spatial relationships between content originated from space-based (Twitter) and place-based (Google Places and OSM) sources of geographic user-generated content based on topics identified by the embedding-based BERTopic model. The contribution of the framework lies on the combination of methods that were selected to improve previous works focused on content-location relationships. Using the city of Lisbon (Portugal) to test our methodology, we first applied the embedding-based topic model to aggregated textual data coming from each source. Results of the analysis evidenced the complexity of content-location relationships, which are mostly based on thematic profiles. Nonetheless, the framework can be employed in other cities and extended with other metrics to enrich the research aimed at exploring the correlation between online discourse and geography.

Keywords:

• Content-location relationships
• UGC
• geo-tagged activity
• topic modeling

1. Introduction

Cities are multi-layered systems, hosting complex human-environment interactions in the form of activities, functions, flows, places and meanings embedded into the surrounding urban landscape (Gao et al. 2017, Iranmanesh et al. 2022). Today, the widespread use of online platforms through mobile phones and location-based services provides fast and voluminous georeferenced data in urban areas. Geospatial big data from user-generated content (UGC) are a major data source for urban studies, with applications such as identifying regions of interest (Shang et al. 2016), examining urban perception and functional structure (Hu et al. 2021), unraveling mobility patterns, mapping sentiments, among many others (Belcastro et al. 2021, Gao et al. 2021). When location is attached to published content, users are regarded as social sensors and their footprints are often used as a spatial proxy for obtaining place-based information (Goodchild 2007, Papadakis et al. 2020). Therefore, exploring the degree of correlation between space-based thematic information and the surrounding place-based information is crucial when geo-tagged UGC is an ubiquitous source of data in the literature.

Platforms such as Twitter, with approximately 200 million daily active users and more than 500 million tweets per day (Jay 2022), provide extensive location-based data in densely populated areas. Textual information linked to geographic coordinates, however, do not necessarily reflect the thematic signatures associated with the geographic context from where the user has posted (McKenzie and Adams 2017). Another widely exploited type of location-based UGC data in the literature is represented by points of interest (POI), sourced from platforms such as OpenStreetMap (OSM), Foursquare and Yelp (Niu and Silva 2020). Attributes such as thematic tags, user reviews and number of check-ins are more enriched in place-based information and better mirror the encompassing spatial context, portrayed by place names, functions and affordances (Psyllidis et al. 2022). The discrepancies between content and location are naturally more evident in geo-tagged activity from social networks (eg Twitter and Instagram), where relationships between what is said and where it is said are not obvious.

Previous research has addressed the content-location relationships through different lenses. One such avenue is the inference of geography in both non-georeferenced and geo-tagged textual content by extracting topononyms, geocoding as well as investigating ‘geo-indicativeness’ – the degree to which lexica semantically indicates geographic features (Adams and Janowicz 2021, Melo and Martins 2017, Qiu et al. 2022). In place semantics, natural language processing (NLP) methods have been applied to geo-tagged UGC to obtain thematic and cognitive dimensions of places (Hu 2018a, 2021). In NLP, efforts to incorporate location into topic modeling algorithms are examples of how researchers have acknowledged that location is not just another attribute, but is often intertwined with content (Bo and Martin 2013, Wang et al. 2020). However, studies that focus specifically on addressing the extent to which location and content are related in geo-tagged social media activity are still scarce. In addition, most of the works employing topic modeling rely on the latent Dirichlet allocation (LDA), which has been outperformed by more recent algorithms, making room for improvements regarding the NLP methods of choice (Egger and Yu 2022).

The extent to which everyday conversation in social networks is geo-indicative may vary depending on temporal and spatial scales, as well as the thematic signatures of the text (Gao et al. 2017, Fu et al. 2018). As these data sources continue to support research in urban studies, we need to outline reproducible and straightforward steps aimed at assessing the correlation between text and the urban environment for a given city (de Oliveira and Painho 2021). Limitations found in previous works include employing outdated topic models, relying on manual classification steps, performing content analysis based on individual point-based short-text activity and restricting context information to place types (Hahmann et al. 2014, Herfort et al. 2014, McKenzie and Adams 2017). Furthermore, traditional bag of words topic models, such as LDA, do not consider the syntactic and semantic relationships between words within a document, but recent algorithms are supported by methods that enable the contextualized representation of words (Yang et al. 2023). As content-location correlations are the bridge between spatial context and the content of online activity, the efforts to investigate these relationships should invest in up-to-date topic modeling techniques.

In this paper, we introduce a framework for modeling and comparing similar thematic signatures derived from space-based and place-based online activity. The content-location relationships are better represented as the relations between topics originated from geo-tagged social media text and those from POI reviews and tags. Since georeferenced social media data reveals information that is attached to space while not necessarily being thematically related to it, we refer to these sources as space-based. In contrast, POI information and reviews are considered place-based sources as they are better at disclosing urban functions, affordances and perceptions that describe and are related to space. Our contribution relies on providing a methodological framework that can be employed in other cities to enhance the content-location discussions and that is based on more recent methods for topic modeling which have not been applied for this task, more specifically the Bidirectional Encoder Representations from Transformers topic model (BERTopic, Grootendorst (2022)). We also attempt to improve previous efforts by aggregating textual content based on a grid, extracting statistically significant thematic regions, using metrics to objectively assess spatial and thematic similarity, as well as using place reviews as our proxy of the urban landscape. The framework is employed using geo-tagged Twitter posts as our space-based source and reviews and names from Google Places and OpenStreetMap as our place-based reference. All platforms provide large datasets from extensive activity in the majority of urban regions across the globe, including the city of Lisbon (Portugal), where we chose to test the framework.

The remainder of this paper is organized as follows. In Section 2, we present the literature that covers theoretical and methodological aspects of our study. Sections 3 and 4 bring forward our data and methods respectively, from which we obtained our results, found in Section 5. Section 6 is dedicated to our interpretation and discussion of the findings, and lastly, we present our concluding remarks in Section 7.

2. Background and related work

2.1. The relationships between content and location in social networks

Natural language in explicitly geo-tagged social media activity can either disclose information about a place or merely from a place (Hu 2018b). In both cases, content may be influenced or caused by features and events from users’ origins at different scales, such as the locale, neighborhood, city and country. This is particularly exploited in previous works that analyze geo-tagged social media data for situational awareness and emergency response in natural disasters including floods, earthquakes and typhoons (Herfort et al 2014, Huang and Xiao 2015, Suwaileh et al. 2022). Although extreme circumstances might generate a higher correlation between content and location, geo-tagged user-generated content can also reflect everyday urban life. In GIScience, these have become common sources for spatially assessing urban thematic characteristics derived from websites, digital gazetteers, social media (Twitter, Foursquare, Instagram, Flickr), Wikipedia, among others (Hobel et al. 2015, 2016, Chen et al. 2019, Twaroch et al. 2019, Belcastro et al. 2021, Gao et al. 2021). In social media, textual content can act as location-based proxies for urban life in regard to activities (eg shopping, working, eating out, recreation) and functions (eg commercial, transportation, residential) that cities can support in different places and regions (Gao et al. 2017). However, we need to be aware of the limitations in relying on social media posts with coordinates, as its attachment to space might not necessarily indicate correlation with the neighboring settings (Fu et al. 2018).

The vast number of works that explore urban dynamics from geo-text data is evidence that correlation between content and location is generally assumed to be high. Using tweets and POI classes, Hahmann et al. (2014) demonstrated that content-location correlation is often low and varies according to place types, arguing that studies should acknowledge this relationship in their applications while also discussing the need to critically consider the link between a piece of information to a pair of coordinates. With that in mind, McKenzie and Adams (2017) used place labels from Foursquare in a supervised topic modeling of geo-text data from social media platforms, showing that content related to built-up places seem to have a lower correlation while content characterized by physiographic features exhibit a higher alignment between data sources. Their theoretical underpinnings stem from the discussion between space and place, which is in fact fundamental in content-location relationships. Other similar examples in the literature seem to focus on the space-based aspects, such as extracting user positions based on tweet meta-data and matching to correspondent locations found in GeoNames and OSM data (Zohar 2021). More place-oriented approaches for discussing the relationships between content and location are timid: while Lansley and Longley (2016) revealed the influence of land-use and urban activities on the content of tweets, Yu et al. (2022) standpoint was to consider POI reviews as adequate spatial proxies of place-based information. Therefore, content-location relationships must be seen through an extended perspective, where comparisons are based not only on positions but also on meanings, functions, activities and affordances of the urban landscape.

2.2. Natural language processing and geo-tagged user-generated content

NLP consists in several techniques that aims at structuring, extracting information and making sense of human natural language (Lamurias and Couto 2019). As geo-tagged UGC carries information on people’s in-space activities, opinions and experiences, it provides discursive information that can be used to explore different thematic attributes related to the urban landscape (Dunkel 2015, Martí et al. 2019). According to Twaroch et al. (2019), UGC does reflect people’s experiences, focus, opinions and interests to a significant degree, and therefore NLP is a crucial tool to find relevant patterns in unstructured text data. The most prevalent NLP method found in the literature is topic modeling, which is able to reduce the complexity of massive geo-text datasets to extract thematic signatures linked to places, activities and perceptions (Fu et al. 2018).

From the wide range of available topic models, the LDA algorithm became pervasive in the literature (Liu et al. 2019). LDA is an unsupervised probabilistic model based on word co-occurrences (Blei et al. 2003, Jenkins et al. 2016). Some of the countless examples include extracting cognitive regions of Northern and Southern California (Gao et al. 2017); identifying urban functional regions in cities with check-in information (Gao et al. 2017); estimating geographic regions from unstructured text (Adams and Janowicz 2021); as well as the previously mentioned works of Lansley and Longley (2016), McKenzie and Adams (2017) and Yu et al. (2022). Nonetheless, research on topic modeling methods has empirically demonstrated the disadvantages of LDA, including careful tuning of hyper-parameters for generating cohesive topics, the requirement of detailed assumptions, overlapping topics, user-defined number of topics and restrictions in assessing the correlation between topics as word correlations are ignored (Egger and Yu 2022).

Although LDA has been one of the best-known and widely used models, other methods for text representation have been developed in the last years. In particular, algorithms that use word or sentence embeddings have been applied in more recent topic models such as the Top2Vec (Egger 2022). Word embeddings are vector representations of text data that enable semantic properties to be encoded whereby similar pieces of text information are nearer in vector space (Naseem et al. 2021). Therefore, by embedding words in a continuous vector space, words with similar semantic and syntactic meaning can be mapped to nearby points (Comber and Arribas-Bel 2019). Embeddings have been used within GIScience for tasks such as address geocoding, fine-scale land-use identification from POI data and even for building algorithms aimed at reasoning the complex spatial semantics of place types (Place2Vec), among others (Yao et al. 2017, Yan et al. 2017, Zhang et al. 2022). However, works in the field that employ topic models supported by word embeddings are still not commonplace, especially for exploring location-content relationships.

As embedding-based models are able to generate contextual representations, relationships that emerge in the vector space might be related to context emerging from the geographic space. Therefore, even without inserting spatial variables, the use of embedding-based topic modeling is more effective in unraveling latent geographic topics of interest and in the separation of geographic and non-geographic clusters (Yang et al. 2023). Among recent algorithms, Grootendorst (2022) has developed the BERTopic, a model that combines BERT embeddings (Bidirectional Enconder Representations From Transformers, developed by Devlin et al. (2019)) and other methods that enable higher flexibility for different use cases. The model works by first creating embeddings that use a pretrained language model, followed by reducing the dimensionality of documents and grouping semantically similar documents into clusters that represent distinct topics. Lastly, the model employs a class-based TF-IDF (term frequency-inverse document frequency) to compare the importance of terms and retrieve the most representative words per topic (Grootendorst 2022, Egger and Yu 2022).

BERTopic has been employed in social media text analysis such as investigating public sentiments regarding the monkeypox outbreak (Ng et al. 2022) and detecting cognitively distorted thinking patterns in Twitter (Alhaj et al. 2022). BERT embeddings have also been implemented in methods aimed at extracting geospatial information and toponyms in unstructured text (Chu et al. 2022, Berragan et al. 2023). In addition to outperforming other topic models, BERTopic is able to generate more interpretable topics, allows multilingual analysis and automatically finds the number of topics (Egger and Yu 2022, Egger 2022). In this paper, we have opted for implementing the model not only because the vector space might reflect the spatial context better than traditional approaches such as LDA, but also because the use of BERTopic in exploring location-content relationships in UGC has not been carried out in the literature.

3. Data and preprocessing

Using the Twitter Search API, we retrieved all georeferenced tweets posted roughly within the metropolitan area of the city of Lisbon, Portugal. Our search query collected tweets that lay within a 40 km radius around the centroid of Lisbon’s municipality without time constraints. The following filtering and selection are exemplified in Figure 1.

Figure 1. Twitter data filtering and preprocessing steps prior to topic modeling.

First, tweets without explicit coordinate-based geo-tagging were removed to best represent users’ active location sharing. Then, we selected those tweets whose assigned language field was Portuguese as the high number of tourists in the city might influence the data distribution. The next steps were to remove tweets with duplicated text entries to reduce contamination of spams, followed by clipping the data to Lisbon’s municipality extent. Based on a 200 m-spaced hexagonal grid, we filtered user contribution in space by allowing up to 10 tweets per user per cell with the objective of reducing users that might skew data distribution in specific locations. The chosen spatial unit of analysis has an area of approximately 0.03 km², which is able to embed most city blocks but not enough to cover neighborhoods. We believe that this resolution is reasonable for our analysis based on the city’s urban fabric and similar grid-based implementations (Andrade et al. 2020). As for limiting user contribution, our goal was to reduce the effects of potential dominating spatial bias from the most active users (Gao et al. 2017). We also removed the cells containing less than the median value of tweets per cell across the study area. In combination, the previous tasks were aimed at both spatially leveraging user contribution and ensuring that areas with reduced user activity would not contribute to the topic modeling. Since there are no standards on these procedures, our choice of thresholds per cell was done arbitrarily both for the cell removal and for limiting user contribution.

The publishing dates of filtered tweets ranged from 2010 to 2021 and thus we assume that more than a decade of space-based online activity might have substantially contributed to shaping thematic information regarding different aspects of the city. After obtaining the final tweet distribution, we spatially aggregated their textual within each cell of the hexagonal grid covering the city. Therefore, each hexagonal cell represented a document in our topic modeling analysis. Throughout the paper, we will use the word ‘document’, ‘hexagon’ and ‘cell’ interchangeably depending on the context, although they are the same in our analysis. Lastly, we processed the text for the model by removing unwanted text such as special characters, emojis, urls and stop words.

Representing the thematic place-based counterpart, we sourced data from Google Places API and OSM (Figure 2). We gathered all POI from Google Places within the city, as well as POI and building centroids across Lisbon from OSM. Features extracted from Google Places consisted of user reviews and place names, whereas we retrieved non-empty place names from OSM. We opted not to use place type tags from both sources as we intended to mainly focus on textual information generated by users (place reviews) and place names that act as specific information linked to places. Place type tags not only might not represent specific locations in the city but also are not necessarily defined by users. Following feature extraction, we aggregated the text-based data based on the previous hexagonal grid, succeeded by the same text processing prior to topic modeling. Most text data originated from users’ reviews on Google Places, where publishing dates ranged from 2011 to 2022.

Figure 2. Google Places and OSM data filtering and preprocessing steps prior to topic modeling.

By having similar temporal distributions, both datasets from Twitter and Google may thematically reflect consolidated place-based urban dimensions, even though POI might appear or cease to exist. Figure 3 shows the data distribution of instances from the space-based and place-based data sources prior to cell-based aggregation.

Figure 3. Point locations of data instances from Twitter, Google Places and OSM prior to hexagonal cell aggregation.

4. Methods

The framework we introduce is visually described in Figure 4. Our spatial unit of analysis are the cells that compose the hexagon-based grid across the city. The main components of the framework include: setting the aggregated textual data from Twitter and place-based sources (Google and OSM); employing the BERTopic transformer-based topic modeling for each source; comparing topics emerged from each source using the cosine similarity metric; carrying out Getis-Ord $G_i^{*}$ hotspot analysis for retrieving statistically representative topic-based cells; applying the Jaccard similarity index aimed at ultimately comparing thematic and spatial similarities that support the discussion on content-location relationships for the case study.

Figure 4. Methodological framework developed in the study.

4.1. Topic modeling

In order to extract thematic signatures from our space-based and place-based sources of textual information, we applied the BERTopic algorithm developed by Grootendorst (2022). Each cell of our hexagonal grid covering the city of Lisbon contained aggregated text-based data, acting as our documents for topic extraction. The BERTopic algorithm uses pre-trained transformer models, rooted in neural network architectures and able to encode words in vector-based representations (Saidi et al. 2022). In addition, it merges machine learning approaches to both reduce dimensionality through UMAP (Uniform Manifold Approximation and Projection for Dimension Reduction) and cluster similar embeddings for topic identification through HDBSCAN (Hierarchical Clustering and Density-Based Spatial Clustering of Applications with Noise).

We employed the BERTopic model for each data source independently, although we have set the same hyper-parameters to reduce any model-driven variations in the originated topics. Whereas most parameters were kept as default given the lack of similar frameworks that use BERTopic, we did modify others for our implementation. For the HDBSCAN hyper-parameters, we set the minimum cluster size to 10 documents while keeping the minimum number of samples as 5 to potentially minimize the number of outliers (Grootendorst 2022). Since topics are generated through density-based clustering, documents are not forced to fit clusters and hence those that fail to belong to a topic are considered outliers, which helps reducing noise and generating more meaningful topics (Egger and Yu 2022). In addition, this also means that although hyper-parameters can be tuned to reduce outliers or change the minimum number of documents for topic generation, there is no prior selection regarding the number of topics. The embedding-based model reduces the dimensions and clusters documents into an optimal number of topics given the input parameters and data.

As for the UMAP, we set the number of neighboring sample points to 5 to constrain local neighborhood size and focus on local as opposed to global patterns. Increasing the number of neighbors provides a more global view of the embedding structure whereas lower values output a more local perspective (Grootendorst 2022). As UMAP is stochastic in nature, we also set a random state to guarantee the reproducibility of the model. For each topic, we retrieved the top 15 words that contributed the most in representing the information for the topic cluster. Lastly, we chose a multilingual embedding model, as not only Google Places reviews might be in different languages, but also in tweets, as languages assigned by Twitter are not always accurate. For each data source, the final output is the topic probability distributions across the grid cells, which are the input of the hotspot analysis, whereas word probability distributions for each topic are compared using the cosine similarity between topics.

4.2. Cosine similarity

To objectively compare the topics identified in the model between data sources, we used the cosine similarity metric. The similarity metric represents the angle between vectors. As the output topic information from BERTopic consists of the 15 most important words that form the topic cluster and their respective values of importance or probability, we treated topics as 15-dimensional vectors. The smaller the angle between vectors, the more similar the topics are in the vector space (Liu et al. 2019). The cosine similarity is defined as follows: (1) $$\text{similarity}=\hspace{0.17em}\cos \hspace{0.17em}(\mathbf{A},\mathbf{B})=\frac{\mathbf{A}\cdot \mathbf{B}}{\Vert \Vert \mathbf{A}\Vert \Vert \hspace{0.17em}\hspace{0.17em}\Vert \Vert \mathbf{B}\Vert \Vert }$$(1)

Where A and B are vectors, and similarity is given by calculating the product between vectors divided by the cross product of their lengths (Fu et al. 2018). With values ranging from 0 to 1, we computed the cosine similarity for all topics retrieved from Twitter against those from the place-based sources. We then filtered the output pairwise matrix to select the highest values for each Twitter-based topic, showing the most similar corresponding place-based topics within the vector space. Following the selection, we assess the spatial relationship between corresponding topics to characterize and visualize the content-location relationships.

4.3. Getis-Ord $G_i^{*}$

The last two steps consist in assessing the spatial relationship between the space-based and place-based topics across the city, ultimately aimed at providing insights regarding the content-location relationships. In the first stage, we carried out a hotspot analysis to retrieve statistically significant cells in regard to topic distributions, represented by the probability values assigned to documents or hexagons of belonging to each topic retrieved by the algorithm. For this task, we chose to calculate the Getis-Ord Gi* statistic, part of the G family of statistics developed by Getis and Ord (2010) aimed at characterizing pronounced local clusters of high and low values. In a study area with n points and $X=\left[x_1,\dots ,x_n\right]$ measurements, and assuming weights $w_{i,j}$ to be defined between all pairs of points i and j (for all $i,j\in \{1,\dots ,n\}$), the Getis-Ord $G_i^{*}$ is denoted as: (2) $$r{G}_i^{*}=\frac{\sum _{j=1}^n{w}_{i,j}{x}_j-\overline{X}\sum _{j=1}^n{w}_{i,j}}{S\sqrt{\frac{n\sum _{j=1}^n{w}_{i,j}^2-{\left(\sum _{j=1}^n{w}_{i,j}\right)}^2}{n-1}}}$$(2)

Where $\overline{X}$ is the mean of all measurements and S is the standard deviation of all measurements (Kumar and Parida 2021). In our implementation, we ran the hotspot analysis for all topics extracted in the previous stage and identified cells with z-scores higher than 1.65, which are samples with standard deviations that have 90% or higher confidence or significance in regard to not responding to a random spatial distribution (Rossi and Becker 2019).

4.4. Jaccard index

In the second stage, we computed the Jaccard index metric for the identified hotspot areas corresponding to the pairwise comparison of similar topics derived from space-based and place-based sources. In other words, after selecting significant spatial distributions of the topics that yielded higher values of cosine similarity between sources, we computed the spatial similarity between these distributions. The metric is defined as: (3) $$J(A,B)=\frac{|A\cap B|}{|A\cup B|}$$(3)

Where the result corresponds to the intersection divided by the unions of two sets A and B. Ranging from 0 to 1, the metric measures the similarity of two sets, represented in our case by the hotspot areas computed previously, similar to the approach of Heikinheimo et al. (2020). More precisely, sets A and B are the areal extent composed by cells identified as hotspots with z-scores higher than 1.65 from each source respectively. We then discussed the spatial and thematic similarities between the topics derived from Twitter and the corresponding ones derived from POI names and reviews.

5. Results

5.1. Topics

As described in the previous section, we ran the BERTopic algorithm using the same hyper-parameters for the cell-based documents derived from our space-based (Twitter) and place-based (Google and OSM) data sources. While Twitter data yielded 31 topics with 376 outlier documents, Google and OSM data yielded 35 topics with 381 outliers. We present a selection of interesting topics, their words and probabilities as well as word translations from the space-based and place-based sources in Tables 1 and 2 respectively. We have listed all identified topics and their information in the appendix (Tables A1–A4), including the ones we do not mention or discuss throughout the paper. The topic order is based on the descending count of documents (hexagonal cells) that were assigned by the algorithm as belonging to the topics. Topic belonging corresponds to the dominant topic of each hexagon or the topic with the highest probability for the document, as each cell yielded probability values ranging from 0 to 1 for all topics. In addition, topic information also includes the corresponding number of instances that were originally aggregated in the documents: tweets, OSM features and Google Places POI. In total, the 31 Twitter topics were modeled based on the aggregated text of 610,593 tweets and the 35 topics of place-based sources originated from 11,392 Google and OSM instances.

Table 1. Selected sample of interesting topics from Twitter.

Football (Topic 2)			Alfama (Topic 4)
Hexagons: 49/tweets: 41079			Hexagons: 43/tweets: 26754
Word	Translation	Prob.	Word	Translation	Prob.
Estádio	Stadium	0.1462	Duetos	Duet	0.0698
Benfica	Benfica football team	0.1183	Bar	Bar	0.0472
Sport	Benfica football team	0.0995	Amp	Organization in Alfama	0.0434
Alvalade	José Alvalade stadium	0.0604	Alfama	Alfama neighborhood	0.0420
José	José Alvalade stadium	0.0568	Gastronomia	Gastronomy	0.0400
Slbenfica	Benfica football team	0.0470	Restaurant	Restaurant	0.0375
xxi	–	0.0414	Café	Café	0.0341
Sporting	Sporting football team	0.0374	Praça	Plaza/square	0.0299
Luz	Luz stadium	0.0301	Mercado	Market	0.0297
Carregabenfica	Benfica football team	0.0267	Ribeira	Area in Lisbon	0.0277
Campo	field	0.0266	Fado	Fado music	0.0269
Bairro	Neighborhood	0.0235	Sobremesa	Dessert	0.0268
Alto	Tall/high	0.0232	Comércio	Business	0.0260
Slb	Benfica football team	0.0214	Música	Music	0.0252
sportingcp	Sporting football team	0.0204	Concerto	Concert	0.0248
Airport (Topic 22)			University (Topic 28)
Hexagons: 16/tweets: 14991			Hexagons: 11/tweets: 3580
Word	Translation	Prob.	Word	Translation	Prob.
Aeroporto	Airport	0.3283	Faculdade	Faculty/university	0.1310
Lis	Lisbon airport	0.1976	Ciências	Sciences	0.1254
Others	–	0.0559	Universidade	University	0.1218
Delgado	Lisbon airport	0.0467	Colombo	Colombo mall	0.0557
Humberto	Lisbon airport	0.0466	Cinemas	–	0.0552
Chegadas	arrivals	0.0446	Campus	–	0.0530
Airport	–	0.0432	Justiça	Justice	0.0505
Arrivals	–	0.0431	FCUL	University of Lisbon	0.0381
Terminal	–	0.0423	Medicina	Medicine	0.0355
Departures	–	0.0367	Dentaria	Dental	0.0343
Partidas	Departures	0.0367	Holmes	Local gym chain (Holme’s place)	0.0265
Comunidades	Communities	0.0332	Campo	Field	0.0264
Lisboalis	Lisbon airport	0.0321	Place	Local gym chain (Holme’s place)	0.0208
Internacional	International	0.0318	Filme	Film	0.0168
Portuguesas	Portuguese	0.0312	IMAX	IMAX cinema	0.0160

Table 2. Selected sample of interesting topics from Google Places and OSM.

Health (Topic 4)			Education (Topic 7)
Hexagons: 41/OSM: 198/Google POI: 232			Hexagons: 35/OSM: 140/Google POI: 96
Word	Translation	Prob.	Word	Translation	Prob.
Good	–	0.0145	Escola	School	0.0800
Atendimento	Service/treatment	0.0144	School	–	0.0652
Excelente	Excellent	0.0140	University	–	0.0312
Farmácia	Pharmacy	0.0133	Faculdade	Faculty/university	0.0218
Service	–	0.0129	Universidade	University	0.0216
Saúde	Health	0.0127	Azulejos	Portuguese tiles	0.0197
Clínica	Clinic	0.0124	Registo	Registration	0.0181
Centro	Center	0.0114	Teachers	–	0.0169
Hospital	–	0.0112	José	José Fontana square	0.0166
Simpatia	sympathy	0.0103	Ensino	Teaching/education	0.0164
Café	–	0.0103	professores	Professors	0.0154
Great	–	0.0103	Fontana	José Fontana square	0.0152
lda	company	0.0102	Faculty	–	0.0142
Appointment	–	0.0092	Superior	Higher (education)	0.0137
Rua	Street	0.0091	campus	–	0.0131
Shopping mall (Topic 18)			Sports (Topic 30)
Hexagons: 20/OSM: 107/Google POI: 156			Hexagons: 12/OSM: 23/Google POI: 26
Word	Translation	Prob.	Word	Translation	Prob.
Atendimento	Customer service	0.0215	Futebol	Football	0.0554
Colombo	Colombo mall	0.0161	Campo	Field	0.0516
Serviço	–	0.0147	Musgueira	Musgueira sports complex	0.0451
Good	–	0.0144	Bandeiras	Flags	0.0451
lda	Company	0.0139	Desportivo	Sports	0.0449
loja	Store	0.0131	Ténis	Tennis	0.0407
Empresa	Company	0.0130	Park	–	0.0387
Centro	Center	0.0127	Universitário	University	0.0368
Excelente	Excellent	0.0119	Tennis	–	0.0346
Really	–	0.0111	Clube	Club	0.0327
Service	–	0.0099	Ferreira	Portuguese surname	0.0311
Profissionalismo	Professionalism	0.0096	Condições	Conditions	0.0311
Preço	Price	0.0095	Desportiva	Sports	0.0306
Telheiras	Telheiras neighborhood	0.0094	Court	–	0.0302
Equipa	Team/staff	0.0093	Amaral	Portuguese surname	0.0283

Alfama, a historic neighborhood in Lisbon known for Fado – a famous style of Portuguese folk music (Cocola-Gant and Gago 2021) – was the theme identified in Topic 4. The words thematically characterize the neighborhood as common in-situ activities include concerts (fado music) and dining out. Words that build the football topic (Topic 2) include mostly references to the two largest stadiums in Lisbon and their respective football teams, Benfica and Sporting (Borges 2019). The location-specific airport topic (Topic 22) mainly consists of references to Lisbon’s airport, whereas the university topic (Topic 28) contains words that are both generally related to higher education as well as specific faculties of the University of Lisbon. Overall, interpretable topics emerged from the social media network yielded thematic profiles mostly related to neighborhoods, locations and areas of interest, rather than general place-mediated activities.

Interesting topics from the place-based perspective included health (Topic 4), education (Topic 7), shopping mall (Topic 18) and sports (Topic 30). Topics originated from documents based on Google Places and OSM also contained words related to specific places and areas within the city, yet overall to a lesser extent in comparison with topics from Twitter.

5.2. Topic similarity

We computed the cosine similarity for all Twitter topics against those originated from Google and OSM data. For each topic, we selected the highest value of similarity using the pairwise matrix to obtain the most similar corresponding place-based topic. In Table 3, we listed each Twitter topic and their matching topics alongside their cosine similarity values.

Table 3. Most similar topic pairs based on the highest cosine similarity values yielded when comparing Twitter topics against those from Google Places and OSM.

Topics			Topics
Twitter	Google and OSM	Cosine similarity	Twitter	Google and OSM	Cosine similarity
−1 (outlier)	5	0.9367	15	5	0.9136
0	5	0.9367	16	20	0.8646
1	29	0.9162	17	6	0.927
2	19	0.9527	18	21	0.9163
3	29	0.9451	19	6	0.9396
4	14	0.9338	20	29	0.9286
5	29	0.8268	21	5	0.9038
6	6	0.9018	22	15	0.9275
7	6	0.8898	23	5	0.9218
8	5	0.9225	24	5	0.9221
9	21	0.8486	25	6	0.8911
10	26	0.915	26	14	0.8936
11	5	0.9105	27	29	0.9338
12	28	0.9242	28	7	0.906
13	29	0.9153	29	6	0.9213
14	20	0.8813	30	10	0.9039

In Table 3, we highlighted the highest values of similarity that represent the most similar comparisons in the vector space. The most similar topics (0.9527) were represented by the football topic from Twitter (Topic 2) and Topic 19 from Google and OSM, whose thematic signatures are related to football as well as Lisbon-based football teams and stadiums. The second highest similarity (0.9451) did not yield easy interpretations regarding the topics’ semantic relationships. While Topic 3 from Twitter is mostly related to landmarks and POI located in downtown Lisbon, the corresponding Topic 29 thematic signatures are characterized by shopping-related activities. However, the topic from Twitter has ‘comércio’ (business or commerce) as its first representative word, although likely related to a main landmark in the city named ‘Praça do Comércio’ (Comércio plaza).

The third highest similarity (0.9396) was measured for the comparison between Topic 19 from Twitter, which refers to landmarks in two different neighborhoods, and Topic 6, from which words did not point towards a discernible thematic profile. Apart from the outlier, we noticed that three particular place-based topics were associated with Twitter topics in six different comparisons. Topic 6, with no specific thematic profile; Topic 29 (shopping activities); and Topic 5, which is vaguely related to general services in the city.

5.3. Spatial similarity

Based on the previous identified topics, we ran the Getis-Ord $G_i^{*}$ hotspot analysis to seek the local high values of the topic distribution. For each output, we selected the cells with z-scores denoting 90% confidence or higher. Cell-based hotspot areas are better at depicting the relevant regions in regard to the original distributions of topics, which oftentimes are spread across the city. Then, for each topic pair we computed the Jaccard index based on the distribution of the selected cells as shown in Table 4. The three highest outputs are highlighted.

Table 4. Jaccard indices between selected hotspot areas of similar topic pairs from space-based and place-based sources.

Topics			Topica
Twitter	Google and OSM	Jaccard index	Twitter	Google and OSM	Jaccard index
−1 (outlier)	5	–	15	5	0.05
0	5	∼0	16	20	0.07
1	29	0.01	17	6	0.05
2	19	0.15	18	21	∼0
3	29	0	19	6	∼0
4	14	0.03	20	29	0.03
5	29	0.16	21	5	0.04
6	6	∼0	22	15	0.18
7	6	∼0	23	5	0.06
8	5	0.11	24	5	0.02
9	21	0.01	25	6	0.02
10	26	0.04	26	14	0.12
11	5	0.03	27	29	0
12	28	0.04	28	7	∼0
13	29	0.06	29	6	0.03
14	20	0.07	30	10	0.05

With their thematic profiles linked to Lisbon’s airport, the Jaccard index between the Topic 22 (Twitter) and Topic 15 (Google Places and OSM) scored the highest value (0.18). Terms in the topics include ‘departures’, ‘arrivals’ and ‘taxi’ as well as references to the name of the airport. Another instance of similar themes in the geographic space is represented by the football topic pair, which yielded the third-highest Jaccard index (0.15).

The second highest measurement was yielded by the Topic 5 and 29 pair (0.16). By itself, the topic from Twitter does not point towards a specific thematic profile, however, the place-based topic is strongly related to shopping activities in the city. Therefore, the high spatial relationship suggest that the uncertain thematic profile might also be linked to shopping, even though the topic is polluted with noise.

5.4. Content-location relationships

The core of this study lies at providing a framework to extract thematic and spatial relationships between content generated from space-based and place-based sources, ultimately enriching the discussion on content-location relationships within a given city. Since it is not feasible to discuss about all relationships in regard to comparisons between the topics’ vector and geographic space, we brought forward visualizations of topic pairs selected on the basis of their spatial and thematic similarities. In Figures 5 and 6, we display the high-value hotspot distributions from the two most similar topic pairs according to cosine similarity and Jaccard index, respectively.

Figure 5. Hotspot distribution and Jaccard index of the two most similar topic pairs (top and bottom) based on the cosine similarity.

Figure 6. Hotspot distribution and Jaccard index of the two most spatially similar topic pairs (top and bottom) based on the Jaccard index.

The football theme is represented by the topic pair with the highest similarity in the vector space as well as significant spatial overlap. Visual inspection allow us to observe their similar hotspot distribution. The output suggests the content-location correlation for this thematic profile is high. This is not the case for the second most similar topic pair, which had no spatial overlap whatsoever. The topic from the social network mainly revealed landmarks of Lisbon’s downtown, while it also contained words linked to the ‘FIL’ exhibition center. Although identified as part of the same topic, these two thematic signatures are related to distinct regions. The lack of spatial correlation indicate that despite having high cosine similarity, their thematic profiles are distinct, as its corresponding place-based topic consisted of shopping-related words.

As for the most spatially similar topic pair, the airport thematic profile evidences a high content-location correlation between Twitter and the place-based counterparts. This might suggest that when users geo-tag content related to airports, they are most likely engaging in activities afforded by the airport location. However, content-location relationships become blurry when comparing the distribution of the second-highest spatially similar topic pair. Contaminated with noise, the topic from Twitter does not indicate a clear thematic profile, yet the comparison with the corresponding place-based shopping theme shows a significant spatial correlation. We selected place-based topics based on the highest cosine similarity against each Twitter topic, yet this pair had yielded the lowest value from all topic pairs. In Figure 7 we present two final examples of topic pairs to complement our discussion.

Figure 7. Hotspot distribution and Jaccard index of two topic pairs (top and bottom).

Similar thematic profiles regarding education showed negligible spatial overlap across the city. The topic from Twitter was mainly linked to instances related to higher education and universities, whereas the place-based topic also contained words related to education in general. While we identified university campuses in both distributions, they pointed towards different areas in the city resulting in significantly low overlap. In this case, one can argue that content-location correlation is low as the spatial similarity is close to zero. However, the geo-tagged content from Twitter collectively refers to the location of the University of Lisbon. The intricate relationship between thematic and spatial similarities between geo-tagged activity and the corresponding place-based content is also exemplified in the last topic pair. Despite being linked to different neighborhood and landmarks, evidenced by a weak cosine similarity, their spatial overlap is mostly located in a historic and touristic region. Both Belém and Alfama are historic districts in Lisbon, enclosing important landmarks and attractions.

6. Discussion

The information we harvested from space-based and place-based sources of unstructured text were collectively analyzed in the form of topics. Both sources yielded thematic profiles that described locations, activities and functions of the urban landscape. The steps of our framework are able to quantitatively compare topics derived from geo-tagged social media activity with the most similar topics emerged from place-based sources (Google Places and OSM). However, elucidating content-location relationships based on thematic and spatial correlations depend on careful interpretations of the results.

Although we applied the embedding-based BERTopic without others models for comparison, the topic clusters showed that the algorithm was able to output many coherent and interpretable topics, including geo-indicative topics of interest that are related to specific activities, functions and affordances of different regions within the city. The algorithm is freely available to the public and does not require substantial text preprocessing. In addition, the algorithm yields an optimal number of topics according to cluster parameters and hence does not force instances to belong to topics, which is a better option for oftentimes noisy or incomplete data. Therefore, studies that source data from geo-tagged online activity should not only take advantage of the advances in embedding-based models, but also compare with other traditional and novel topic models

On the other hand, topics with unclear thematic profiles (such as Topic 5 and 6 from place-based sources) frequently scored high values of cosine similarity with topics retrieved from Twitter. Textual data sourced from user-generated content is noisy, unstructured and messy by nature. When adding the spatial dimension, a new layer of complexity is included and researchers must be aware of the limitations of the data themselves prior to the analysis. By developing a straightforward reproducible framework using an embedding-based topic model, researchers can test thematic content-location relationships by changing model parameters, confidence levels, thresholds, preprocessing steps as well as the resolution of the spatial unit of analysis.

We observed that the degree to which geo-tagged content from social media is connected to the corresponding place-based characteristics of the city will vary depending on thematic profiles. Similar insights were found in related literature, but differences were portrayed by place types (Hahmann et al. 2014, McKenzie and Adams 2017). Here, we represent both space-based and place-based geo-text dimensions as collective topics to be objectively evaluated against each other. Although previous works have developed methods to geo-locating social media activity, we developed an approach to extend the discussion on how discursive information in intentionally geo-tagged text might be associated with urban settings and activities (Adams and Janowicz 2021).

Football, a topic that potentially has a high disconnect between content and location, was characterized by one of the highest interpretable correlation between sources. The relation suggests that in Lisbon, geo-tagged content linked to football is connected to locations that afford football related activities. We observed the same relationship in the airport topic, indicating that geo-tagged content thematically associated with airports is mostly generated near the airport location. However, uncertain thematic profiles and different types of categories (activities, neighborhoods and places) show that choosing topic modeling to explore content also reveals that correlations between content and spatial context is intricate and open to discussion.

We were able to identify similar topic pairs coming from space-based and place-based sources using the cosine similarity metric. The following spatial similarity analysis disclosed distinct relationships from which interpretations are not necessarily straightforward. Our results can be translated through two somewhat contrasting viewpoints. One hand, dissimilarities reinforce the limitation of using geo-tagged UGC, as it only connects spatial footprints with textual data (Papadakis et al. 2020). On the other hand, similarities between sources strengthen the justifications of using UGC to infer the interaction between people and places within the urban environment (Lansley and Longley 2016, Heikinheimo et al. 2020).

Furthermore, our study supports the inquisitive discussions on the reliability and accuracy of geospatial information collected or inferred from online sources, problems that are not only a product of well-known biases (Twaroch et al. 2019), but also of the theoretical and methodological approaches behind these practices. Although our analysis was bounded to the same limitations and biases, we hope to incite other researchers to extend analytical and conceptual frameworks aimed at validating the use of geo-tagged UGC to unravel human-centered urban dimensions.

Some limitations should be pointed out. First, both sources of user-generated content are biased regarding their users’ demographic profiles and do not fully cover the whole extent of the city, which in turn affects representativeness (Zhang et al. 2018, Gao et al. 2021). In addition, aggregating geo-tagged textual data into cells can result in biases that stem from the MAUP problem (Openshaw effect), whereby thematic and spatial relationships might differ according to cell size or scale (Goodchild 2022). Lastly, results also show that interpretation of thematic and spatial relationships are often constrained to prior familiarity with and knowledge about the city in regard to specific places, activities and neighborhoods. To improve interpretability as well as insights about content-location relationships, future work should consider applying spatially explicit topic models, gathering additional data from online sources as well as implementing alternative metrics and spatial and temporal units of analysis.

7. Conclusions

Geo-tagged social network data has become an extremely popular data source in urban studies as information is used to map, explore and infer the several dimensions of human-environment interactions, including human mobility, urban perception, sentiment analysis among many other activities and opinions. However, the content-location relationships in social media activity are intricate and not always clear. In this article, we introduced a methodological framework to explore the vector-space and geographic-space similarities between thematic profiles emerged from space-based (Twitter) and place-based (Google Places and OSM) sources of geographic user-generated content.

The stages included applying a transformer-based topic modeling, retrieving cosine similarity measurements between topics, running Getis-Ord $G_i^{*}$ hotspot analysis to extract representative topic cells as well as computing Jaccard indices to calculate spatial similarities. The results showed that content-location relationship between the surrounding urban settings and the thematic content of in-situ online activity are heavily dependent on the thematic signatures. Nonetheless, the framework can easily be implemented and extended in other cities in order to explore novel insights and support discussions on the use of geo-tagged UGC in GIScience.

Author contributions

Vicente Tang: Conceptualization, Methodology, Software, Formal analysis, Investigation, Data Curation, Writing – Original Draft, Writing – Review and Editing, Visualization.

Marco Painho: Conceptualization, Resources, Writing – Original Draft, Writing – Review and Editing, Supervision, Project administration, Funding acquisition

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data and codes availability statement

The data and code that support the findings of this study are available at: https://doi.org/10.6084/m9.figshare.19307936

Additional information

Funding

The authors acknowledge the funding from the Portuguese national funding agency for science, research and technology (Fundação para a Ciência e a Tecnologia – FCT) through the CityMe project (EXPL/GES-URB/1429/2021; https://cityme.novaims.unl.pt/) and the project UIDB/04152/2020 - Centro de Investigação em Gestão de Informação (MagIC)/NOVA IMS.

Notes on contributors

Vicente Tang

Vicente Tang is a PhD candidate in Geoinformatics at the NOVA Information Management School of the NOVA University Lisbon and also a research member of the CityMe project (https://cityme.novaims.unl.pt/). His research interests include the use of user-generated content and participatory approaches within GIScience to explore places and regions in the city. Twitter: @CityMe5 @vicetang_

Marco Painho

Marco Painho is a full professor of Geographic Information Science at the NOVA Information Management School of the NoOVA University Lisbon and the scientific coordinator of the Geoinformatics and Analytics Laboratory. He holds a Master in Regional Planning from the University of Massachusetts, Amherst and a PhD in Geography from the University of California Santa Barbara. Twitter: @painho

Appendix

Table A1. List of identified topics from Twitter including the ones mentioned throughout the paper (cont.): words, translations, probabilities as well as the count of documents and instances.

Topic 0

Topic 1

Football (Topic 2)

Downtown (Topic 3)

Alfama (Topic 4)

Topic 5

Hexagons: 179/tweets: 110757

Hexagons: 105/tweets: 53074

Hexagons: 49/tweets: 41079

Hexagons: 43/tweets: 33237

Hexagons: 43/tweets: 26754

Hexagons: 179/tweets: 25049

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Vou

(I) go

0.0242

Padrão

Touristic landmark

0.0341

Estádio

Stadium

0.1462

Comércio

Business

0.0754

Duetos

Duet

0.0698

Vou

(I) go

0.0268

Vasco

Vasco da Gama mall

0.0144

Descobrimentos

Touristic landmark

0.0338

Benfica

Benfica football team

0.1183

Terreiro

Terreiro do Paço plaza

0.0611

Bar

Bar

0.0472

Escola

School

0.0213

Gama

Vasco da Gama mall

0.0143

Alfama

Alfama neighborhood

0.0249

Sport

Benfica football team

0.0995

Praça

Plaza

0.0610

Amp

Organization in Alfama

0.0434

Tou

(I) am

0.0208

Casa

House/home

0.0117

Arena

–

0.0209

Alvalade

José Alvalade stadium

0.0604

paço

Terreiro do Paço plaza

0.0596

Alfama

Alfama neighborhood

0.0420

Chiado

Chiado neighborhood

0.0167

Tou

(I) am

0.0114

Parque

Eduardo VII park

0.0204

José

José Alvalade stadium

0.0568

fil

FIL exhibition center

0.0173

Gastronomia

Gastronomy

0.0400

Brasileira

Brasileira café

0.0160

Parque

Park

0.0111

Eduardo

Eduardo VII park

0.0193

slbenfica

Benfica football team

0.0470

Museu

Musesum

0.0171

Restaurante

Restaurant

0.0375

Accessories

–

0.0139

Dormir

To sleep

0.0108

vii

Eduardo VII park

0.0178

Xxi

–

0.0414

cinema

–

0.0167

Café

Café

0.0341

Jewelry

–

0.0139

Escola

School

0.0107

Meo

Telecommunication company

0.0163

Sporting

Sporting football team

0.0374

lxfactory

LXFactory open mall

0.0164

Praça

Plaza/square

0.0299

Jewellery lover

–

0.0139

Mim

Me

0.0103

Nações

Parque das Nações parish

0.0160

Luz

Luz stadium

0.0301

Nacional

National

0.0160

Mercado

Market

0.0297

Jewellery design

–

0.0139

Acho

(I) think

0.0099

Centro

Center

0.0123

Carregabenfica

Benfica football team

0.0267

Feira

Market/fair

0.0155

Ribeira

Area in Lisbon

0.0277

Jewellery designer

–

0.0138

Técnico

IST university

0.0099

Telheiras

Telheiras neighborhood

0.0115

Campo

Field

0.0266

Internacional

International

0.0144

Fado

Fado music

0.0269

Jewellery addict

–

0.0138

Amanhã

Tomorrow

0.0098

Museu

Museum

0.0106

Bairro

Neighborhood

0.0235

Cinemateca

Cinemateca film archive

0.0133

Sobremesa

Dessert

0.0268

Paiva

Portuguese surname

0.0137

Vida

Life

0.0095

Vou

(I) go

0.0102

Alto

Tall/high

0.0232

Posted

–

0.0132

Comércio

Business

0.0260

Silver

–

0.0137

IST

IST university

0.0094

Saldanha

Saldanha neighborhood

0.0095

Slb

Benfica football team

0.0214

Livraria

Bookshop

0.0122

Música

Music

0.0252

Jewelry gram

–

0.0132

Centro

Center

0.0094

Melhor

Best/better

0.0090

Sportingcp

Sporting football Team

0.0204

Main

–

0.0111

Concerto

Concert

0.0248

Carmo

Carmo convent

0.0132

Topic 6

Topic 7

Topic 8

Topic 9

Topic 10

Topic 11

Hexagons: 37/tweets: 14304

Hexagons: 32/tweets: 16763

Hexagons: 32/tweets: 10100

Hexagons: 27/tweets: 14514

Hexagons: 26/tweets: 12839

Hexagons: 24/tweets: 12316

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Ajuda

Ajuda neighborhood

0.0322

Pombal

Marquês de Pombal (square)

0.1006

Campo

Field

0.0257

Parque

Parque das Nações parish

0.1621

Conhecimento

Pavilhão do Conhecimento (museum)

0.0361

Vou

(I) go

0.0358

Alcântara

Alcântara neighborhood

0.0179

Marquês

Marquês de Pombal (square)

0.0971

Ilha

Island

0.0237

Oceanário

Oceanarium

0.1404

Pavilhão

Pavilhão do Conhecimento (museum)

0.0279

Livro

Book

0.0193

Europe

–

0.0137

Amaz

–

0.0369

Vou

(I) go

0.0230

Nações

Parque das Nações parish

0.1376

RTP

Portuguese public broadcasting

0.0270

Mim

Me

0.0180

República

Republic

0.0132

Metro

Subway/metro

0.0258

Amores

Love (plural)

0.0224

expo

Lisbon Expo ’98

0.0599

Vou

(I) go

0.0223

Casa

House/home

0.0164

Ponte

Bridge

0.0120

Chafariz

Fountain

0.0224

Jardim

Garden

0.0168

Naçõesparque

Parque das Nações parish

0.0505

Televisão

Television

0.0195

Feira

Market

0.0161

Hotelaria

Hospitality

0.0110

Enoteca

Wine cellar

0.0217

Peninsular

–

0.0166

Tweetgram

–

0.0324

Rádio

Radio

0.0192

Chilihotgirl

–

0.0159

Turismo

Tourism

0.0109

Vou

(I) go

0.0211

Rua

Street

0.0164

Tweegram

–

0.0322

Ciência

Science

0.0179

Jogos

Games

0.0153

Abril

April (bridge)

0.0104

Edinho

Portuguese football player

0.0184

Heróis

Heros

0.0158

Webstagram

–

0.0320

Casa

House/home

0.0163

Comigo

With me

0.0132

Marquês

Marquês de Pombal (square)

0.0102

Vinho

Wine

0.0172

Campos

Fields

0.0154

Wow

–

0.0320

Pessoa

Fernando Pessoa (portuguese writer)

0.0149

Falar

(To) speak/talk

0.0128

Avenida

Avenue

0.0102

Praça

Plaza/square

0.0151

Tou

(I) am

0.0131

Congressos

Congresses

0.0275

Fernando

Fernando Pessoa (portuguese writer)

0.0143

Tou

(I) am

0.0121

Rato

Rato neighborhood

0.0098

Blog

–

0.0145

Guerra

War

0.0131

Okportugal

–

0.0246

Twitter

–

0.0136

Merda

Swear word

0.0120

Pombal

Marquês de Pombal (square)

0.0090

Prof

–

0.0140

Metro

Subway/metro

0.0122

Amarlisboa

(to) love lisbon

0.0213

Viva

Alive/cheer

0.0130

Amanha

Tomorrow

0.0115

Espaco exibicionista

Local art gallery

0.0087

Vídeo

–

0.0140

República

Republic

0.0121

Posted

–

0.0200

Damaralines

–

0.0123

CRL

Swear word

0.0114

Nações

Parque das Nações Parish

0.0086

Acabou

(it) ended/ran out/just happenEd

0.0125

Casa

House/home

0.0113

Monsanto

Monsanto park

0.0198

Tou

(I) am

0.0113

Melhor

Best/better

0.0113

Exhibition

–

0.0084

Hotel

–

0.0105

Campolide

Campolide neighborhood

0.0113

tejo

Tagus river

0.0192

Campo

Field

0.0108

Dormir

(to) sleep

0.0111

Topic 12

Topic 13

Topic 14

Topic 15

Belém and tourist attractions (Topic 16)

Topic 17

Hexagons: 23/tweets: 9592

Hexagons: 23/tweets: 16909

Hexagons: 23/tweets: 11105

Hexagons: 23/tweets: 15015

Hexagons: 23/tweets: 20123

Hexagons: 20/tweets: 7132

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Amoreiras

Amoreiras mall

0.0695

Chiado

Chiado neighborhood

0.1190

Mega

–

0.0822

Vou

(I) go

0.0294

Torre

Tower (Belém tower)

0.1179

Lusófona

Lusófona university

0.0354

Hospital

–

0.0689

Estação

Station

0.0942

Craque

ACE/expert

0.0766

Dágua

Linha d’Água restaurant

0.0230

Belém

Belém neighborhood

0.1036

ULHT

Lusófona university

0.0291

Humanas

Humanities

0.0603

Oriente

Oriente train station

0.0850

Clube

Club

0.0597

Linha

Linha d’Água restaurant

0.0197

Castelo

Saint George castle

0.0872

Laranjeiras

Laranjeiras neighborhood

0.0273

Martim

Martim Moniz square

0.0581

Armazéns

Warehouses/ Armazéns do Chiado mall

0.0849

Rios

Rivers

0.0397

Mim

Me

0.0130

Jorge

Saint George castle

0.0776

Carlotatavares

–

0.0272

Moniz

Martim Moniz square

0.0573

Gare

Oriente train Station

0.0754

Rua

Street

0.0332

Vida

Life

0.0121

Belem

Belém Neighborhood

0.0560

Humanidades

Humanities

0.0255

Ciências

Sciences

0.0521

Ferroviária

Railway

0.0654

Catedral

Cathedral

0.0315

Acho

(I) think

0.0118

Miradouro

St. Peter of Alcântara viewpoint

0.0482

Cidadão

Citizen

0.0236

Sociais

Social

0.0520

Rossio

Rossio square

0.0588

Mundial

Worldwide

0.0285

Dormir

(to) sleep

0.0118

Alcântara

St. Peter of Alcântara viewpoint

0.0469

Universidade

University

0.0225

Fcshunl

Social sciences university

0.0484

Baixachiado

Chiado neighborhood

0.0469

Esperança

Hope

0.0264

Falar

(to) speak/talk

0.0116

pedro

St. Peter of Alcântara viewpoint

0.0438

Tecnologias

Technologies

0.0208

Faculdade

faculty/university

0.0453

Baixa

Lisbon’s downtown

0.0379

Crossfit

–

0.0264

Tou

(I) am

0.0115

Tower

–

0.0432

Morgas

–

0.0195

Center

–

0.0411

Starbucks

–

0.0345

Augusta

Augusta street

0.0250

Mãe

Mother

0.0112

Cultural

–

0.0402

Vou

(I) go

0.0189

Shopping

–

0.0385

AZVD

–

0.0334

Estação

Station

0.0238

Casa

House/home

0.0111

Saldanha

Saldanha neighborhood

0.0293

Loja

Store/shop

0.0168

SAMS

private hospital

0.0287

Metro

Subway/metro

0.0333

Ferroviária

Railway

0.0238

Pessoa

Person

0.0111

CCB

Belém culture center

0.0274

Amigadop

–

0.0155

FCSH

Social sciences university

0.0284

Entrecampos

Entrecampos metro station

0.0328

Susanagateira

Local store

0.0237

Amanhã

Tomorrow

0.0102

Centro

Center

0.0269

Modernização

Modernization

0.0147

Clínico

Clinic

0.0277

Others

–

0.0225

Megacraque

(Mega) ACE/expert

0.0211

Ihugof

–

0.0099

Berardo

Berardo museum

0.0231

Administrativa

Administrative

0.0144

Universidade

University

0.0233

Pontinha

Pontinha neighborhood

0.0224

Vitória

Victory

0.0200

Ficar

(to) stay

0.0097

Mosteiro

Monastery (in Belém)

0.0214

Academica

Academic

0.0141

Table A2. List of identified topics from Twitter including the ones mentioned throughout the paper: words, translations, probabilities as well as the count of documents and instances.

Topic 18

Picoas and Saldanha neighborhoods (Topic 19)

Topic 20

Topic 21

Airport (Topic 22)

Topic 23

Hexagons: 20/tweets: 12159

Hexagons: 19/tweets: 6312

Hexagons: 17/tweets: 30270

Hexagons: 16/tweets: 6189

Hexagons: 16/tweets: 14991

Hexagons: 15/tweets: 4778

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Camões

Luis de Camões square

0.0694

Picoas

Picoas neighborhood

0.0332

Senhora

Madam/woman

0.0199

chapitô

Chapitô circus

0.0539

Aeroporto

Airport

0.3283

Iscal

ISCAL university

0.0527

Chiado

Chiado neighborhood

0.0693

Petiscar

To nibble (food)

0.0303

Acessórios

Accessories

0.0165

Adrianamars

–

0.0418

Lis

Lisbon airport

0.1976

Contabilidade

Accounting

0.0311

Luís

Luis de Camões square

0.0612

Villaret

Villaret heatre Picoas

0.0293

Isabel

Isabel Ribeiro wedding clothes

0.0153

Marcação

Appointment

0.0392

Others

–

0.0559

Administração

Administration

0.0268

Gulbenkian

Calouste Gulbenkian garden

0.0592

Gulbenkian

Gulbenkian park

0.0269

Ribeiro

Isabel Ribeiro wedding clothes

0.0146

Gel

–

0.0376

Delgado

Lisbon airport

0.0467

Noivos

Grooms (wedding)

0.0255

Jardim

Garden

0.0585

Jardim

Garden

0.0266

Homem

Man

0.0136

Blog

–

0.0373

Humberto

Lisbon airport

0.0466

Noivas

Brides

0.0232

Doces

Sweets

0.0565

Fontes

Fontes P.M avenue

0.0261

Cerimónia

Cerimony

0.0133

Strength

–

0.0330

Chegadas

Arrivals

0.0446

Acessórios

Accessories

0.0229

Calouste

Calouste gulbEnkian garden

0.0553

Entrenospetisca

To nibble (Food)

0.0258

Corroios

Corroios pariSh

0.0128

Imersaoempo

Manicure

0.0326

Airport

–

0.0432

Vestimos

(we) dress

0.0227

Sonhos

Dreams

0.0520

Saúde

Health

0.0258

Glamour

–

0.0126

Verniz

Nail polish

0.0233

Arrivals

–

0.0431

Instituto

Institute

0.0217

Fundação

Foundation

0.0491

Tecnologia

Technology

0.0250

Criança

Child

0.0126

Thestrengthclinic

Local gym

0.0232

Terminal

–

0.0423

Rapaz

Young man

0.0214

Zoológico

Zoo

0.0483

Melo

Fontes P.M avenue

0.0244

Noivos

Groom (wedding)

0.0122

Health

–

0.0191

Departures

–

0.0367

Superior

–

0.0211

Bairroaltohotel

Hotel

0.0410

Estesl

Estesl university

0.0235

Noiva

Bride

0.0121

Vou

(i) go

0.0171

Partidas

Departures

0.0367

Altis

Altis hotels

0.0208

Praça

Square

0.0338

Conversar

To chat

0.0231

Qualidade

Quality

0.0121

Vitoriapti

Pti election winning

0.0165

Comunidades

Communities

0.0332

Portela

Portela parish

0.0196

Largo

Square

0.0307

Pereira

Fontes P.M avenue

0.0229

Requinte

Refinement

0.0118

Longevity

–

0.0161

Lisboalis

Lisbon airport

0.0321

Senhora

Madam/woman

0.0192

Taberna

Tavern

0.0296

Superior

Higher

0.0220

Vida

Life

0.0118

Escola

School

0.0159

Internacional

International

0.0318

Avenidas

Avenidas Novas parish

0.0181

Botânico

Botanic

0.0262

Escola

School

0.0216

Lojas

Stores

0.0117

Areeiro

Areeiro neighborhood

0.0149

Portuguesas

Portuguese

0.0312

Especializados

Specialized

0.0180

Topic 24

Topic 25

Topic 26

Topic 27

University (Topic 28)

Topic 29

Hexagons: 15/tweets: 31609

Hexagons: 14/tweets: 22732

Hexagons: 13/tweets: 7744

Hexagons: 12/tweets: 10069

Hexagons: 11/tweets: 3580

Hexagons: 10/tweets: 2386

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

District

–

0.0459

Graça

Graça neighborhood

0.0247

Monte

Nossa Senhora do Monte viewpoint

0.1055

Mude

(you) change

0.0470

Faculdade

Faculty/university

0.1310

Estufa

Estufa Fria garden

0.0851

Cais

Cais do Sodré neighborhood

0.0229

Urban

Urban Beach nightclub

0.0227

Senhora

Nossa Senhora do Monte viewpoint

0.0858

Augusta

Augusta street Arch

0.0404

Ciências

Sciences

0.1254

Fria

Estufa Fria garden

0.0759

Sodré

Cais do Sodré neighborhood

0.0219

Beach

Urban Beach nightclub

0.0206

Miradouro

Nossa Senhora do Monte viewpoint

0.0720

Arco

Augusta street Arch

0.0370

Universidade

University

0.1218

Eduardo

Eduardo VII park

0.0748

Pensão

Pensão Amor bar

0.0212

Miradouro

Viewpoint

0.0178

Lust

Lust in Rio nightclub

0.0690

Design

–

0.0364

Colombo

Colombo mall

0.0557

vii

Eduardo VII park

0.0595

Amor

Pensão Amor bar

0.0192

Européia

European

0.0156

Micasaenlisboa

–

0.0611

Moda

Fashion

0.0318

Cinemas

–

0.0552

Eleven

Venue in Eduardo VII park

0.0564

Iscteiul

ISCTE university

0.0171

Cidade

City

0.0154

Rio

Lust in Rio nightclub

0.0605

Comércio

Commerce/comércio square

0.0301

Campus

–

0.0530

Depósitos

Portuguese bank

0.0423

Vídeo

Video

0.0156

Marina

Sea park

0.0153

Lav

–

0.0385

Marcações

appointments

0.0290

Justiça

Justice

0.0505

Caixa

Portuguese bank

0.0393

Universitário

University

0.0143

União

Union

0.0150

Bomdia

Good morning

0.0383

Rua

Street

0.0279

FCUL

University of Lisbon

0.0381

Sede

headquarters

0.0392

Acabou

(it) ended/ran out/just happened

0.0130

Vida

Life

0.0120

Mstattoos

Tattoo place

0.0335

Whatsapp

–

0.0255

Medicina

Medicine

0.0355

Park

–

0.0365

ISCTE

ISCTE university

0.0126

Photography

–

0.0107

Vivo

(I) live

0.0239

Disponíveis

Availables

0.0253

Dentária

Dental

0.0343

Parque

Parque

0.0287

Justa

Santa Justa neighborhood and llift

0.0120

Photooftheday

–

0.0107

Tattoolisboa

–

0.0237

Agende

(you) schedule

0.0240

Holmes

Local gym chain (Holme’s place)

0.0265

Roma

Roma avenue

0.0282

Instituto

Institute

0.0115

Celso

Celso barbershop

0.0101

Tattooportugal

–

0.0236

Praça

Square/plaza

0.0236

Campo

Field

0.0264

Avenida

Roma avenue

0.0175

Msg

Message

0.0110

Barbeiro

Celso barbershop

0.0098

Tattoo

–

0.0199

Museu

Musesum

0.0231

Place

Local gym chain (Holme’s place)

0.0208

Claudiaccamposh

–

0.0148

Vamuver

YouTube channel

0.0102

Melhor

best/better

0.0097

Tattoos

–

0.0199

Horário

Time/timeslot

0.0219

Filme

Film

0.0168

Frutalmeidas

Café at Roma avenue

0.0135

Carmo

Carmo convent/Carmo square

0.0099

Igers

Online photography community

0.0097

Meninos

Boys

0.0189

Triunfal

Augusta street Arch

0.0198

IMAX

IMAX cinema

0.0160

Lar

Home

0.0134

Topic 30

Hexagons: 10/tweets: 7112

Word

Translation

Prob.

Coliseu

Coliseu dos Recreios auditorium

0.1654

Recreios

Coliseu dos Recreios auditorium

0.1286

Marriott

–

0.1185

Hotel

–

0.0963

Restauradores

Restauradores square

0.0639

Hard

Hard Rock Café

Rock

Hard Rock Café

0.0477

bbc

–

0.0440

Café

–

0.0225

Praça

Square/Plaza

0.0221

Bar

–

0.0216

Coliseudelisboa

Coliseu dos Recreios auditorium

0.0154

Belém

Belém neighborhood

0.0148

MAAT

MAAT museum

0.0143

Table A3. List of identified topics from OSM and Google Places including the ones mentioned throughout the paper (cont.): words, translations, probabilities as well as the count of documents and instances.

Topic 0

Topic 1

Topic 2

Topic 3

Health (Topic 4)

Services (Topic 5)

Hexagons: 93/OSM: 1555/Google POI: 515

Hexagons: 88/OSM: 341/Google POI: 417

Hexagons: 53/OSM: 432/Google POI: 198

Hexagons: 52/OSM: 546/Google POI: 263

Hexagons: 41/OSM: 198/Google POI: 232

Hexagons: 40/OSM: 196/Google POI: 209

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Hotel

–

0.0262

Profissionais

Professionals

0.0175

Igreja

Church

0.0461

Food

–

0.0232

Good

–

0.0145

Recomendo

(I) recommend

0.0219

Hostel

–

0.0211

Excelente

Excellent

0.0174

Church

–

0.0294

Oriente

Oriente train station

0.0193

Atendimento

Service/treatment

0.0144

Serviço

Service

0.0182

Location

–

0.0208

Serviço

Service

0.0171

Santo

Saint

0.0180

Great

–

0.0182

Excelente

Excellent

0.0140

Atendimento

Customer service

0.0180

Place

–

0.0175

Atendimento

Service/treatment

0.0165

Senhora

Madam/woman

0.0157

Nice

–

0.0174

Farmácia

Pharmacy

0.0133

Ajuda

Ajuda neighborhood

0.0168

Good

–

0.0169

Lda

Company

0.0156

Capela

Chapel

0.0156

Staff

–

0.0169

Service

–

0.0129

Excelente

Excellent

0.0152

Clean

–

0.0168

Empresa

Company

0.0148

Convento

Covent

0.0153

Place

–

0.0166

Saúde

Health

0.0127

Restelo

Restelo neighborhood

0.0152

Great

–

0.0154

Melhor

Best/better

0.0147

Place

–

0.0143

Good

–

0.0160

Clínica

Clinic

0.0124

Rua

Street

0.0152

Stay

–

0.0151

Recomendo

(I) recommend

0.0143

Santa

Saint

0.0142

Amazing

–

0.0135

Centro

Center

0.0114

Adorei

(I) loved (it)

0.0138

Room

–

0.0141

Rua

Street

0.0142

Nice

–

0.0142

Friendly

–

0.0134

Hospital

–

0.0112

Melhor

Best/better

0.0124

Nice

–

0.0132

Preços

Prices

0.0127

Beautiful

–

0.0141

Restaurant

–

0.0129

Simpatia

Sympathy

0.0103

Profissional

Professional

0.0121

Apartment

–

0.0132

Profissional

Professional

0.0127

Cruz

Cross

0.0112

Best

–

0.0119

Café

–

0.0103

Avenida

Avenue

0.0116

Staff

–

0.0131

Excelentes

Excellent

0.0118

Colégio

Private school

0.0101

Service

–

0.0116

Great

–

0.0103

Lda

Company

0.0116

Rooms

–

0.0120

Qualidade

Quality

0.0107

Clara

Claire/bright

0.0099

One

–

0.0110

Lda

Company

0.0102

Profissionais

Professionals

0.0116

Friendly

–

0.0116

Caixa

Box/cashier

0.0100

Memória

Memory

0.0097

Super

–

0.0107

Appointment

–

0.0092

Bairro

Neighborhood

0.0114

Breakfast

–

0.0108

Top

–

0.0095

Good

–

0.0091

Estação

Station

0.0107

Rua

Street

0.0091

Simpatia

Sympathy

0.0109

Topic 6

Education (Topic 7)

Topic 8

Topic 9

Topic 10

Topic 11

Hexagons: 35/OSM: 225/Google POI: 241

Hexagons: 35/OSM: 140/Google POI: 96

Hexagons: 34/OSM: 170/Google POI: 251

Hexagons: 33/OSM: 244/Google POI: 212

Hexagons: 32/OSM: 366/Google POI: 204

Hexagons: 29/OSM: 126/Google POI: 157

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Escola

School

0.0155

Escola

School

0.0800

Lda

Company

0.0169

Great

–

0.0216

Hotel

–

0.0154

Service

–

0.0188

Ajuda

Ajuda neighborhood

0.0148

School

–

0.0652

Excelente

Excellent

0.0167

Good

–

0.0194

Rato

Rato neighborhood

0.0130

Shop

–

0.0149

Lda

0.0127

University

–

0.0312

Atendimento

Customer serviec

0.0165

Service

–

0.0174

embaixada

embassy

0.0128

Atendimento

Customer service

0.0142

Atendimento

Customer service

0.0124

Faculdade

faculty/university

0.0218

Recomendo

(i) recommend

0.0139

Food

–

0.0156

Place

–

0.0111

Great

–

0.0140

Service

–

0.0121

Universidade

university

0.0216

Serviço

Service

0.0134

Atendimento

Customer service

0.0154

Lda

–

0.0106

Good

–

0.0136

Loja

Store

0.0115

Azulejos

Portuguese tiles

0.0197

Simpatia

Sympathy

0.0122

Nice

–

0.0136

Staff

–

0.0103

Santogal

–

0.0136

Benfica

Benfica neighborhood

0.0113

Registo

Registration

0.0181

Avenida

Avenue

0.0112

Staff

–

0.0132

Hostel

–

0.0102

Place

–

0.0131

Good

–

0.0113

Teachers

–

0.0169

Top

–

0.0109

Marquês

Marquis

0.0124

Great

–

0.0095

Oficina

Workshop

0.0126

Excelente

Excellent

0.0112

José

José Fontana square

0.0166

Super

–

0.0109

Belém

Belém neighborhood

0.0121

Chile

Chile square

0.0095

Staff

–

0.0125

Marquês

Marquês de P. square

0.0108

Ensino

Teaching/education

0.0164

Profissionalismo

Professionalism

0.0106

Place

–

0.0117

Room

–

0.0094

Super

–

0.0123

Pombal

Marquês de P. square

0.0107

Professores

Professors

0.0154

Profissionais

Professionals

0.0105

Excelente

Excellent

0.0107

Rua

Street

0.0093

Lda

Company

0.0115

Espaço

Space

0.0107

Fontana

José Fontana square

0.0152

Great

–

0.0102

Serviço

Service

0.0105

Excelente

Excellent

0.0093

Nice

–

0.0113

Nice

–

0.0104

Faculty

–

0.0142

Empresa

Company

0.0101

Friendly

–

0.0099

Good

–

0.0092

Kia

–

0.0108

Centro

Center

0.0102

Superior

Higher (education)

0.0137

Campo

Field

0.0101

Qualidade

Quality

0.0094

Atendimento

Customer service

0.0091

Friendly

–

0.0104

Rua

Street

0.0100

Campus

–

0.0131

Rua

Street

0.0097

Espaço

Space

0.0089

Avenida

Avenue

0.0091

Helpful

–

0.0101

Topic 12

Topic 13

Topic 14

Airport (Topic 15)

Topic 16

Topic 17

Hexagons: 26/OSM: 224/Google POI: 150

Hexagons: 26/OSM: 96/Google POI: 44

Hexagons: 24/OSM: 272/Google POI: 67

Hexagons: 24/OSM: 91/Google POI: 135

Hexagons: 23/OSM: 112/Google POI: 114

Hexagons: 23/OSM: 142/Google POI: 111

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Good

–

0.0230

Monsanto

Monsanto park

0.1403

Benfica

Benfica neighborhod

0.0349

Aeroporto

Airport

0.0424

Work

–

0.0204

Food

–

0.0265

Pingo

Pingo Doce supermarket

0.0173

Parque

Monsanto park

0.0644

Sapadores

Firefighters

0.0290

Airport

–

0.0403

Nice

–

0.0171

Caselas

Caselas neighborhood

0.0222

Doce

Pingo Doce supermarket

0.0164

Interpretação

Interpretation

0.0611

Avenida

Avenue

0.0255

Flight

–

0.0168

Top

–

0.0155

Ice

–

0.0221

Sebastião

Portuguese name

0.0147

Infantil

Infantile

0.0325

Comida

Food

0.0226

Terminal

–

0.0145

Great

–

0.0144

Cream

–

0.0207

Store

–

0.0130

Acesso

Access

0.0320

Farmácia

Pharmacy

0.0216

Taxi

–

0.0138

Bar

–

0.0136

Portuguese

–

0.0150

Place

–

0.0124

Alvito

Alvito neighborhood

0.0276

Simpatia

Sympathy

0.0179

Euros

–

0.0136

Company

–

0.0135

Great

–

0.0150

Expo

Lisbon Expo ’98

0.0123

Centro

Center

0.0269

Grão

Grain

0.0178

Worst

–

0.0135

Edificio

Building

0.0128

Campolide

Campolide neighborhood

0.0147

Nice

–

0.0121

João

Portuguese name

0.0268

Atendimento

Customer service

0.0178

Olivais

Olivais neighborhood

0.0131

Staff

–

0.0128

Typical

–

0.0146

Service

–

0.0112

Água

Water

0.0251

Laranjeiras

Laranjeiras neighborhood

0.0172

Ever

–

0.0131

Rua

Street

0.0119

Lda

Company

0.0143

Fonte

fountain

0.0109

Tribunal

Court

0.0244

Prazeres

Prazeres cemitery

0.0171

Queiroz

Portuguese surname

0.0129

Studio

–

0.0112

Comida

Food

0.0126

Well

–

0.0107

Oliveiras

Olive trees

0.0244

Good

–

0.0170

Waiting

–

0.0126

People

–

0.0112

Excelente

Excellent

0.0126

Sushi

–

0.0106

Campolide

Campolide neighborhood

0.0238

Produtos

Products

0.0157

Car

–

0.0124

Job

–

0.0112

Amazing

–

0.0126

One

–

0.0104

Estacionamento

Car park

0.0237

Tiffosi

Portuguese clothing store

0.0155

Excelente

Excellent

0.0122

Gestão

Management

0.0111

Restelo

Restelo neighborhood

0.0125

Etnologia

Ethnology

0.0102

Gil

Portuguese name

0.0231

Qualidade

Quality

0.0154

DPD

DPD parcel delivery

0.0121

Place

–

0.0111

Good

–

0.0125

Food

–

0.0102

Running

–

0.0228

Restaurante

Restaurant

0.0152

Minutes

–

0.0120

Figura

Figure

0.0110

Rua

Street

0.0124

Table A4. List of identified topics from OSM and Google Places including the ones mentioned throughout the paper: words, translations, probabilities as well as the count of documents and instances.

Shopping mall (Topic 18)

Football (Topic 19)

Alfama (Topic 20)

Topic 21

Topic 22

Topic 23

Hexagons: 20/OSM: 107/Google POI: 156

Hexagons: 19/OSM: 55/Google POI: 77

Hexagons: 18/OSM: 202/Google POI: 125

Hexagons: 18/OSM: 44/Google POI: 34

Hexagons: 18/OSM: 70/Google POI: 116

Hexagons: 16/OSM: 78/Google POI: 77

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Atendimento

Customer service

0.0215

Stadium

–

0.0523

Alfama

Alfama neighborhood

0.0393

miradouro

Viewpoint

0.0925

Boat

–

0.0448

Penha

Penha de França neighborhood

0.0369

Colombo

Colombo mall

0.0161

Club

–

0.0501

Fado

Fado music

0.0279

Claros

Montes Claros garden

0.0829

Tours

–

0.0314

França

Penha de França neighborhood

0.0363

Serviço

–

0.0147

Gym

–

0.0308

Casa

House

0.0177

Montes

Montes Claros garden

0.0829

Sailing

–

0.0314

Car

–

0.0216

Good

–

0.0144

Football

–

0.0293

Amazing

–

0.0157

View

–

0.0709

Experience

–

0.0288

Ctt

Portuguese postal service

0.0213

Lda

Company

0.0139

Sport

Benfica football team

0.0284

Palace

–

0.0153

Moinho

Moinho park

0.0510

River

–

0.0287

Police

–

0.0205

Loja

Store

0.0131

Benfica

Benfica football team

0.0279

Apartment

–

0.0147

Nice

–

0.0485

Amazing

–

0.0286

Portuguesa

Portuguese

0.0176

Empresa

Company

0.0130

Estádio

Stadium

0.0265

Place

–

0.0145

Place

–

0.0480

Tour

–

0.0245

Rotunda

Roundabout

0.0171

Centro

Center

0.0127

Great

–

0.0235

Food

–

0.0139

Parque

Park

0.0471

Crew

–

0.0236

Matos

Júlio de Matos hospital

0.0155

Excelente

Excellent

0.0119

Best

–

0.0207

Like

–

0.0136

Park

–

0.0435

Great

–

0.0227

Psp

Portuguese civil police

0.0152

Really

–

0.0111

Good

–

0.0192

Great

–

0.0130

Vista

View

0.0430

View

–

0.0224

Pay

–

0.0140

Service

–

0.0099

Staff

–

0.0176

Location

–

0.0124

Nature

–

0.0406

Nice

–

0.0221

Charneca

Charneca neighborhood

0.0139

Profissionalismo

Professionalism

0.0096

Luz

Luz stadium

0.0175

Beautiful

–

0.0123

Kids

–

0.0377

Tejo

Tagus river

0.0209

Rent

–

0.0134

Preço

Price

0.0095

Place

–

0.0169

Chafariz

Fountain

0.0123

Picnic

–

0.0372

Trip

–

0.0204

Hospital

Júlio de Matos hospital

0.0129

Telheiras

Telheiras neighborhood

0.0094

Olivais

Olivais neighborhood

0.0162

Clean

–

0.0122

Olhão

Portuguese town

0.0367

Recommend

–

0.0176

Júlio

Júlio de Matos hospital

0.0128

Equipa

Team/staff

0.0093

Courts

–

0.0161

Everything

–

0.0121

Mocho

Moinho park

0.0355

Docas

Docks

0.0162

Lda

Company

0.0126

Topic 24

Topic 25

Topic 26

Topic 27

Topic 28

Topic 29

Hexagons: 15/OSM: 117/Google POI: 55

Hexagons: 15/OSM: 185/Google POI: 91

Hexagons: 14/OSM: 50/Google POI: 84

Hexagons: 13/OSM: 114/Google POI: 112

Hexagons: 12/OSM: 42/Google POI: 63

Hexagons: 12/OSM: 47/Google POI: 76

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Museu

museum

0.0348

Campo

Field

0.0406

Radio

–

0.0284

Lumiar

0.0217

Cuf

CUF hospital

0.0284

Portas

Portas de Benfica building

0.0289

Carnide

Carnide neighborhood

0.0312

Good

–

0.0228

Renascença

Portuguese radio station

0.0173

Alameda

Alameda square

0.0210

Carnide

Carnide neighborhood

0.0253

Atendimento

customer service

0.0286

Teatro

Theater

0.0286

mouraria

Mouraria neighborhood

0.0226

Great

–

0.0172

Microsoft

–

0.0209

Excelente

Excellent

0.0199

Spacio

Spacio mall

0.0279

Arte

Art

0.0246

Great

–

0.0207

Soluções

Solutions

0.0168

Clinic

–

0.0189

Lumiar

Lumiar neighborhood

0.0196

Boavista

Boavista neighborhood

0.0275

Nacional

National

0.0229

Place

–

0.0176

Best

–

0.0163

Clínica

Clinic

0.0184

Hospital

–

0.0183

Benfica

Benfica neighborhood

0.0231

Theater

–

0.0202

Nice

–

0.0171

Lda

Company

0.0161

Atendimento

Customer service

0.0148

School

–

0.0173

Shopping

–

0.0226

Cultural

–

0.0188

Best

–

0.0169

Alvalade

Alvalade neighborhood

0.0159

Nice

–

0.0140

Pavilhão

Pavilion

0.0167

Produtos

Products

0.0210

Roque

Saint Roch church

0.0185

Belém

Belém neighborood

0.0164

Empresa

Company

0.0152

Emel

Lisbon’s mobility

0.0126

Desportivo

Sports

0.0167

Qualidade

Quality

0.0186

Museum

–

0.0185

Moniz

Martim Moniz square

0.0157

Encarnação

Encarnação neighborhood

0.0148

Excelente

Excellent

0.0125

Pay

–

0.0167

Roupa

Clothing

0.0184

Piano

–

0.0181

Loureiro

Loureiro neighborhood

0.0151

Rádio

Radio

0.0144

Farmácia

Pharmacy

0.0119

Lispolis

LISPOLIS technology hub

0.0165

Costa

Coast

0.0182

Venue

–

0.0176

Chão

Ground/floor

0.0146

Equipa

Team/staff

0.0142

Henriques

Portuguese surname

0.0118

Dresses

–

0.0161

Excelente

Excellent

0.0176

Antiga

Old/ancient

0.0169

Food

–

0.0144

Professional

–

0.0140

Medicina

Medicine

0.0116

Macau

–

0.0157

Loja

Store

0.0173

Estufa

Greenhouse

0.0168

Mamede

São Memede neighborhood

0.0138

Instruments

–

0.0135

Centro

Center

0.0111

Escola

–

0.0155

Serviço

Service

0.0155

Lourenço

Portuguese name

0.0156

Martim

Martim Moniz square

0.0128

Rfm

Portuguese radio station

0.0135

Best

–

0.0109

Staff

–

0.0152

José

Portuguese name

0.0153

Locanda

–

0.0153

Hostel

–

0.0125

Mascarenhas

Portuguese surname

0.0135

Clínicas

Clinics

0.0109

Portuguese

–

0.0150

Good

–

0.0151

Sports (Topic 30)

Topic 31

Topic 32

Topic 33

Topic 34

Hexagons: 12/OSM: 23/Google POI: 26

Hexagons: 11/OSM: 27/Google POI: 15

Hexagons: 11/OSM: 142/Google POI: 92

Hexagons: 10/OSM: 32/Google POI: 64

Hexagons: 10/OSM: 77/Google POI: 60

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Word

Translation

Prob.

Futebol

Football

0.0554

Parque

Park

0.1208

Pizza

–

0.0409

Nice

–

0.0168

Bike

–

0.0254

Campo

Field

0.0516

Keil

Keil do Amaral park

0.1008

Good

–

0.0186

Staff

–

0.0163

Climbing

–

0.0243

Musgueira

Musgueira sports complex

0.0451

Amaral

Keil do Amaral park

0.0968

Chiado

Chiado neighborhood

0.0176

Service

–

0.0161

Tours

–

0.0237

Bandeiras

Flags

0.0451

Park

–

0.0909

Food

–

0.0171

Diamond

–

0.0159

Cais

Cais do Sodré neighborhood

0.0235

Desportivo

Sports

0.0449

Dog

–

0.0779

Staff

–

0.0162

Friendly

–

0.0159

Tour

–

0.0220

Ténis

Tennis

0.0407

Canino

Canine

0.0707

Browns

–

0.0155

Good

–

0.0156

Great

–

0.0210

Park

–

0.0387

Alameda

Keil do Amaral park

0.0517

Restelo

Restelo neighborhood

0.0154

Aranha

Spider

0.0153

Beer

–

0.0199

Universitário

University

0.0368

Picnic

–

0.0510

Place

–

0.0154

Patudos

Bigeye fish

0.0153

Excellent

–

0.0195

Tennis

–

0.0346

Dogs

–

0.0504

Best

–

0.0148

Happy

–

0.0153

Bar

–

0.0193

Clube

Club

0.0327

Pedra

Rock/stone

0.0454

Dominos

–

0.0146

Helpful

–

0.0147

Good

–

0.0192

Ferreira

Portuguese surname

0.0311

Infantil

Infantile

0.0403

Great

–

0.0141

Place

–

0.0144

Place

–

0.0173

Condições

Conditions

0.0311

Beautiful

–

0.0373

Hotel

–

0.0129

Apply

–

0.0141

Service

–

0.0167

Desportiva

Sports

0.0306

Garden

–

0.0367

Loja

Store

0.0124

Work

–

0.0139

Vespa

Wasp

0.0165

Court

–

0.0302

Animais

Animals

0.0363

Tour

–

0.0120

Leitão

Pork

0.0138

Sodré

Cais do Sodré neighborhood

0.0160

Amaral

Portuguese surname

0.0283

Size

–

0.0348

Nice

–

0.0118

Crossfit

–

0.0138

Pesca

Fishing

0.0154