As part of the Forschungspraktikum/Projektpraktikum, Online political polarization Research Lab, Visual Crawler Group, from University of Koblenz-Landau
Our task is to create an expert Twitter Crawler that monitors tweets to find posts that have images with matching faces of certain predefined politicians and then to grab replies to that tweet in order to find out how much hate that politician reveive within the Twitter. These results can then result in an index to show how popular or infamous a politician is on socail media.
Our base data set consists of 4 World Powers ( Mr Donald Trump, Mr Valdimir Putin, Mrs. Angela Merkel and Mr Bouris Johnson). We have 3 images of each and each tweet that includes images will be checked against all the photos.
We use a mixture of different techniques to take advantage of Tweeter API. On average, 10 percent of the collected tweets have images. Out of these, usually 2.1 percent are original tweets (from the news channels amd not re-tweets) and have detectable images of our politicians dataset. For these tweets we use a second api called TWARC to just grab related replies. This process is very slow and requires more access keys and process power to work for larger scales. Last but not least, all the replies will be checked to see any cases hate speech is occuring. We have found out that there is at least 5.3 percent of hate speech involved in the politicians' tweets which we call observable hate since it is the amount a machine learning algorithm can detect.
In order to be able to use the Visual Crawler one has to first apply for at least 1 Twitter Stream API key pair. After this, through the following steps Visual Crawler can be deployed and used:
Parameters in parameter.yml file have to be proberly set. Here is the description for each value:
# parameters
consumer_key: 'Twitter consumer key'
consumer_secret_key: 'Twitter secrert'
access_token: 'Access token provided by Twitter'
access_token_key: 'Access key provided by Twitter'
pin_code: 'Twarc pin code'
crawl_directory: './data/crawl'
process_directory: './data/processed'
dataset_directory: './dataset'
image_directory: './data/images'
log_directory: './data/log'
class_directory: './data/classes'
memory_file: 'history'
lock_file: 'locker'
face_file: 'facer'
label_file: 'labeler'
parsed_file: 'dataset.csv'
reply_file: 'replies.csv'
class_name: 'Proper class name that represents class directory '
dataset_file: 'labeled_tweets.csv'
max_power: 'Max number of wighted tokens for labeling'
refresh_lock: 'Distance between lock update default is 3'
t_handles: 'array [] of all news handels'In order to use Visual Crawler, one has to provide a base detection class. A class consists of a csv file that has name of the politicans or people that should be tracked and detected by the script. Each row requires at least 1 image representing a clear photo of the person's face. Based on our experience, 3 different photos provide you with a good performance and accuraccy.
The structure of the index.csv file shall be as mentioned here:
name token tags
<name of person> <image prefix> <tags about related topics>
<name of person> <image prefix> <tags about related topics>
<name of person> <image prefix> <tags about related topics>Running streamer will collect and save temp json files of tweets that have been sent in between the last 7 days period. Crawler_streamer.py only collects the raw data and process each time you run it.
To extract Tweets and download images of each Tweet, we can run Crawler_extract.py. This script does the first hand analysis on data and makes a folder with all avaible phots alongs with data.csv file which is all of valid Tweets. This script has the ability to detect and ignore already parsed json files so removing stream files manually is not needed.
In this part, we run the face recognizer on all of the valid Tweets of data.csv. Crawler_detect.py is the code that is able to find faces of politicians using Python Face Recognition libraries. One image can include face of 1 or more politicians, if so vote column of that Tweet will be set to number of matched politicans and matches will be list of all names.
Crawler_replies_twarc.py collects the replies to Tweets that had at least 1 matching face on their original Tweets. This code uses TWARC library that needs pre-configuration. It is worth mentioning that the code is very slow in finding replies and it might take days to make generate useful information. All the replies will be saved in a new csv file called replies.csv.
More information about Twarc: https://github.com/DocNow/twarc
To label replies, you can run labeling_nlp.py. This code uses a mixture of 2 semi-supervised dictionary approaches to find cases of hate speech . This results in a new column called automatic on each reply row. This can of course be against a human proof reader to check the validity of votes. In our sample of 1000 Tweets, we have reached 86 % agreement rate.
More information about implementation, achievable results and challenges are avaible on our report document sheet via the link below, however due to data protection rules we are not allowed to share and distrubte our datasets.
Report Doc: https://drive.google.com/file/d/1Fmvtc28Yz5C5yAevarQ9CpJqWgFjs2KM/view?usp=sharing
-
Amir Reza Javaher Dashti, javaher@uni-koblenz.de
-
Marina Rukavitsyna, mrukavitsyna@uni-koblenz.de
-
Rahini Chandrasekaran, rahinic@uni-koblenz.de
-
Marina Ernst, ernstmargo@uni-koblenz.de
This is the Visual Crawler Research Lab Project held by University of Koblenz-Landau in Summer Semester of 2019. The is held by Dr. Oul Han and Frau Ipek Baris as supervisors. In case, you have questions you can stay in touch with the team directly via email:
-
Dr. Oul Han, han@uni-koblenz.de
-
Ipek Baris, ibaris@uni-koblenz.de
-
Institute for Web Science and Technoloogies (WEST), University of Koblenz-Landau
All code is licensed under MIT License .