In 2017 Lucas LaRochelle created Queering The Map as part of a class project. This is a "community generated counter-mapping platform for digitally archiving LGBTQ2IA+ experience in relation to physical space", whose purpose is provide a safe space for queer people around the globe to share experiences anonimously, and link them to a physical space.
The information of the page arises this question:
How does race, gender, sexuality, citizenship, ability, and class affect the ways in which we relate to, move through, and create space?
The purpose of this project is to give some insights in the process of answering this question, respecting the anonimity of the community involved, through the analysis of the discourse of the shared experiences and its relationship with the physical space defined by each country.
- How is the web used?
- Where is it mostly used?
- How are the communities we can find within a country?
- How is the sentiment of the comments found in the web?
- How is it by country?
- How many characters do people use and what's the emotionality of the comments found in the web?
- How are they by country?
- How do they relate to the countries' characteristics?
- What are the country variables that affect the most to the metrics of the comments found in the web?
- folium: Map
- folium.plugins: HeatMap
- geopandas
- json
- langid
- matplotlib.pyplot
- msgpack
- nltk
- nltk.corpus: stopwords
- nltk.sentiment.vader: SentimentIntensityAnalyzer
- numpy
- pandas
- pymongo: MongoClient
- requests
- scipy.stats: f_oneway
- seaborn
- sklearn: preprocessing
- sklearn.cluster: KMeans
- sklearn.linear_model: LinearRegression
- spacy
- statsmodels.formula.api
- wordcloud: WordCloud
- base64
- matplotlib.pyplot
- nltk
- nltk.corpus: stopwords
- nltk.sentiment.vader: SentimentIntensityAnalyzer
- numpy
- pandas
- PIL: Image
- random
- streamlit
- streamlit.components.v1
- wordcloud: WordCloud
- Description:
- Web
- Discourse
- Analysis of the relationship between countries' characteristics and sentiment and emotionality.
- Streamlit page creation.
-
H0: The characteristics of the context in each country don't interfere with the nature of the discurse shared on the web.
-
H1: These characteristics affect the discurse shared on the web.
After a quick bibliographic research about country determinants for stigma and discrimination related to sexual orientation and gender identity1, 18 different variables were collected2-3:
- Region:
- East Asia & Pacific
- Europe & Central Asia
- Latin America & Caribbean
- Middle East & North Africa
- North America
- South Asia
- Sub-Saharan Africa
- Political context:
- Economic context:
- Income group:
- High income
- Upper middle income
- Lower middle income
- Low income
- GDP8:
- Measure as GDP (current US$)
- Income group:
- Education context:
- Health context:
- LGBT+ context3:
- LGBT+ rights index
- Same-sex sexual acts illegal
- Censorship of LGBT issues
- Gender marker change
- Hate crimes based on sexual orientation or gender identity are an aggravating circumstance
The values considered were the mean between 2017 and present time, considering that the webpage started in 2017 and there is no data available about the dates of the comments. In case of missing values, these were imputed calculating the mean value for each variable missing, grouping by region and income group, as these were the ones that seemed more associated with the continuous variables (no statistical test was done to check this hypothesis).
The nature of the discurse was defined by the sentiment determined with the nltk Python library, and the emotionality.
Emotionality is a metric created to face the issue of false neutral sentiment. It takes into consideration the sentiment absolute value but also the character length in the comments. It represents the emotional intensity of the comments, without reflecting if they are positive or negative.
The relationship of each of the variables and the discourse was evaluated on their own, and as a whole, using linear regression.
These are the countries where we find more comments in the web:
The USA contains 46.81% of all the comments.
| Country | Count | Percentage |
|---|---|---|
| USA | 69976 | 46.81 |
| GBR | 13175 | 8.81 |
| CAN | 12690 | 8.49 |
| AUS | 7163 | 4.79 |
| DEU | 3697 | 2.47 |
| FRA | 3502 | 2.34 |
| ITA | 1706 | 1.14 |
| POL | 1603 | 1.07 |
| RUS | 1339 | 0.90 |
| BRA | 1293 | 0.86 |
The language of the comments was detected using the library langid from Python.
As more than 95% of the comments were in English according to this library, and there were issues translating, those in a different language weren't translated.
| Language | Count | Percentage |
|---|---|---|
| en | 142688 | 95.45 |
| fr | 1696 | 1.13 |
| es | 1303 | 0.87 |
| de | 427 | 0.29 |
| pt | 317 | 0.21 |
| it | 295 | 0.20 |
| nl | 274 | 0.18 |
| et | 272 | 0.18 |
| da | 217 | 0.15 |
| ru | 202 | 0.14 |
To predict the geographic clusters for each country, a sample with the top ten countries in web use helped calculate the number of clusters in each country. Using the elbow method, a maximum of 5 to 6 clusters was considered to predict the clusters for every country, finally varying from one to six.
The final prediction of the use for geographic clusters is available on the streamlit page.
It was done only considering those comment in English. Using the library nltk.corpus, and the stop words present by default in it, the most repeated English words in the world were "first", "love" and "time".
| Top words | Count |
|---|---|
| first | 104962 |
| love | 56184 |
| time | 44910 |
| came | 31703 |
| never | 28278 |
| still | 28254 |
| like | 26964 |
| here. | 25196 |
| girl | 25024 |
| one | 24172 |
To create the wordclouds, no stop words were considered as the visual results were more evident including them in the graphs.
In the streamlit page, it is possible to explore the wordclouds for every country.
As you can see in the graph, there were a lot of neutral values (16.84%) for the sentiment using the nltk library in Python.
If we ignore those 16.84% of neutral values, this was the distribution of sentiment in the discourse shown in the web. It is clear that most of the comments show a positive sentiment.
There were negative ouliers below the sentiment value of -0.5576. This would have been the distriution without neutral and outlier values.
This was the distribution of the number of characters per comnent in the web. There were some few comments with a lot of characters:
There weree 5.64% of outliers. This would have been the distribution without condidering the outliers with more than 413 characters.
The distribution of emotionality was similar to the distribution of length but more homogenous, showing 6.31% of outliers
This would have been the distribution without the outliers with an emotionality value higer than 0.1235.
The folowing table shows the coefficients for those variables that showed a statistically significant relationship with the discourse sentiment or emotionality, and the p-values for all the variables included.
| Sentiment | Emotionality | ||||
|---|---|---|---|---|---|
| Quantitative variables | Coefficients if significant | p-value | Coefficients if significant | p-value | |
| Political stability | - | 0.534625 | -0.000025 | 0.001841 | |
| Rule of Law | -0.000269 | 0.000056 | - | 0.249311 | |
| Proportion of female seats | -0.000395 | 0.000973 | 0.000126 | 1.520583e-18 | |
| Voice and Accountability | -0.000182 | 0.002112 | - | 0.605112 | |
| GDP | -1.253206e-14 | 0.000007 | 1.319243e-15 | 0.000083 | |
| Children out of school | 0.001325 | 0.000294 | - | 0.069918 | |
| % of education expenditure | -0.004844 | 2.319145e-07 | 0.000654 | 6.108558e-09 | |
| Adult literacy rate | -0.000491 | 0.001606875 | -0.000047 | 0.01114636 | |
| Antiretroviral therapy coverage | -0.000426 | 0.000211 | - | 0.326746 | |
| % of health expenditure | -0.004398 | 4.661136e-09 | 0.000276 | 0.002159 | |
| UHC coverage index | -0.000589 | 0.000021 | - | 0.6962483 | |
| LGBT+ rights index | -0.001712 | 0.000009 | - | 0.486615 | |
| Legality of same-sex sexual acts | 0.042458 | 0.000094 | 0.002900 | 0.026142 | |
| Hate crimes protection | - | 0.575047 | - | 0.134443 |
| Categorical variables | Coefficients if significant | p-value | | Coefficients if significant | p-value Region, East Asia & Pacific (ref) Europe & Central Asia | - | 0.053812 | | - | 0.877080 Latin America & Caribbean | - | 0.102969 | | - | 0.791358 Middle East & North Africa | - | 0.578289 | | 0.004308 | 0.000247 North America | - | 0.769441 | | -0.001541 | 0.004599 South Asia | - | 0.723594 | | - | 0.168310 Sub-Saharan Africa | - | 0.411084 | | - | 0.063408 Income, High income (ref) Upper middle income | 0.014205 | 0.007249 | | 0.002037 | 0.000967 Lower middle income | - | 0.574642 | | 0.003837 | 0.000035 Low income | 0.074843 | 0.024217 | | 0.008464 | 0.028941 Censorship, Imprisonment as punishment (ref) No censorship | -0.029614 | 0.02316450 | | - | 0.05776440 Varies by region | - | 0.05982512 | | -0.003578 | 0.01821738 Ambiguous | - | 0.1345074 | | - | 0.07393498 Other punishment | - | 0.4718123 | | - | 0.4946909 Fine as punishment | - | 0.7742703 | | - | 0.8349811 State-enforced | - | 0.6813948 | | - | 0..708714 Gender marker change, Legal, surgery not required (ref) Legal, but requires surgery | - | 0.106977 | | 0.002478 | 0.000528 Varies by region | - | 0.939524 | | - | 0.160484 Ambiguous | 0.020352 | 0.022415 | | - | 0.133209 Illegal | - | 0.229530 | | 0.002533 | 0.012426
The characteristics of the countries barely explained the variability of the discourse: when including all the countries variables, the model had a R-squared of 0.002 for sentiment and of 0.005 for emotionality.
The folowing table shows the coefficients for those variables that showed a statistically significant relationship with the discourse sentiment or emotionality when including all the variables in the model, and their p-values.
| Sentiment | Emotionality | ||||
|---|---|---|---|---|---|
| Variables | Coefficients if significant | p-value | Coefficients if significant | p-value | |
| Political stability | 0.0005298575 | 0.04707291 | 0.00006340095 | 0.04612302 | |
| Rule of Law | - | 0.4667247 | - | 0.2486797 | |
| Proportion of female seats | - | 0.4565885 | 0.0001015801 | 0.01061505 | |
| Voice and Accountability | - | 0.7111574 | - | 0.2153887 | |
| GDP | - | 0.2471276 | 1.510002e-15 | 0.002967066 | |
| Children out of school | - | 0.3371182 | - | 0.2652239 | |
| % of education expenditure | - | 0.1730333 | - | 0.4211090 | |
| Adult literacy rate | - | 0.6207066 | - | 0.6681972 | |
| Antiretroviral therapy coverage | -0.0005085612 | 0.01743667 | -6.549984e-05 | 0.01016903 | |
| % of health expenditure | - | 0.4085597 | -5.020974e-04 | 0.02951934 | |
| UHC coverage index | -0.001239511 | 0.01076728 | - | 0.6426139 | |
| LGBT+ rights index | - | 0.9783647 | -6.603307e-04 | 0.005983406 | |
| Legality of same-sex sexual acts | 0.08272276 | 0.0005092681 | - | 0.9414430 | |
| Hate crimes protection | 0.03770694 | 0.0008464976 | 0.004804704 | 0.0003581757 | |
| Region, East Asia & Pacific (ref) | |||||
| Europe & Central Asia | - | 0.5277016 | -0.006949421 | 9.481720e-07 | |
| Latin America & Caribbean | -0.03118771 | 0.04168091 | 0.005412410 | 0.003010116 | |
| Middle East & North Africa | - | 0.09291297 | -0.008608279 | 5.342520e-05 | |
| North America | - | 0.1117380 | - | 0.4267092 | |
| South Asia | - | 0.1851692 | - | 0.9386940 | |
| Sub-Saharan Africa | 0.08012712 | 0.0009698598 | - | 0.05382876 | |
| Income, High income (ref) | |||||
| Low income | -0.1366794 | 0.0009293425 | -0.01696737 | 0.0005604679 | |
| Lower middle income | - | 0.4612828 | 0.005934258 | 2.281194e-04 | |
| Upper middle income | - | 0.3007255 | - | 0.1544082 | |
| Censorship, Imprisonment as punishment (ref) | |||||
| No censorship | 0.07907812 | 0.0001211594 | 0.006998407 | 0.004299755 | |
| Varies by region | 0.1174667 | 5.692929e-07 | 0.007870485 | 0.004911829 | |
| Ambiguous | - | 0.8170347 | - | 0.1155930 | |
| Other punishment | -0.09580888 | 0.03052081 | -0.01140211 | 0.03070329 | |
| Fine as punishment | 0.08313862 | 0.002126322 | - | 0.2521477 | |
| State-enforced | 0.07800163 | 0.0001786075 | 0.005853332 | 0.01825340 | |
| Gender marker change, Illegal (ref) | |||||
| Legal, surgery not required | -0.03134414 | 0.01960927 | -0.01086634 | 1.118489e-11 | |
| Legal, but requires surgery | - | 0.4242546 | -0.007699718 | 6.881194e-07 | |
| Varies by region | -0.04647625 | 0.01220829 | -0.01050085 | 2.009555e-06 | |
| Ambiguous | - | 0.8860066 | - | 0.3327387 |
A more positive sentiment was related to a higher political stability, better legal consideration of same-sex sexual acts and the recognition of hate crimes. The sentiment was significantly better in Sub-Saharan Africa in respect to East Asia & Pacific, and in countries were censorship of LGBTQ issues is done through imprisonment or there are other punishments than imprisonment or fines.
A more negative sentiment was related to a better ARV therapy and UHC coverage. It was worse in Latin America & Caribbean in respect to East Asia & Pacific, in low income countries in respect to high income conuntries, if there is censorship of LGBTQ issues done through imprisonment or other punishments than imprisonment or fines, and if gender chance is legal or varies in respect to it being illegal. Some of these results were very unexpected.
A more intense emoitionality was related to a higher political stability, higher proportion of female seats in national parliament, higher GDP and better protection against hate crimes. The emotionality was significantly more intense in Latin America & Caribbean in respect to East Asia & Pacific, in lower middle income countries in respect to high income ocuntries, and in countries with no censorship, variable or state-enforced in respect to imprisonment as punishment.
A less intense emotionality was related to a better ARV therapy coverage, higher health expenditure and higher LGBT+ rights index. It is also less intense in Europe & Central Asia and Middle East & North Africa in respect to East Asia & Pacific, in low income countries in respect to high income countries, if there is censorship punish with other than fines or imprisonment, and if gender change is legal or varies in respect to it being illegal.
The most repeated words were love and first, it seems that this web is mainly used to share first experiences or coming out stories.
The nltk Python library showed limitations to provide a reliable sentiment. To evaluate intensity the length of the commments can be used, or a related metric to include the weight of the sentiment detected.
Countries with less web use usually had “worst” country indicators, but these didn't necessarily translate in a number emotionality or lower sentiment.