3alama / capstone-52

Twitter Dialect Datasets and Classifiers (EG + GULF Arabic Corpus)

Note: This repo (capstone-52) is used primarily to conduct topic modeling and classification on an Arabic corpus containing both Egyptian (EG) and Gulf (GULF) dialects. For relevant GULF and EG Arabic twitter streams and users, please refer to capstone-35 and capstone-34 repos respectively.

Twitter Dialect Datasets and Classifiers (Arabic)

A project to harvest corpora for Egyptian Arabic and Gulf Arabic from Twitter, conduct descriptive analyses of the resulting corpora, and show that a simple classifier can predict dialect quite effectively.

Getting Started

Clone this repository to your local harddrive: git clone https://github.com/telsahy/capstone-52.git

Prerequisites

Install dependencies from the included requirements.txt file by running either of the following commands:

• !pip install -r requirements.txt
• $ pip install -r requirements.txt

Harvesting Twitter Data and Required Infrastructure

Streaming:

• Create list of dialect specific keyword search terms to use for twitter streamers.
• Create Docker file containing tweepy authentication tokens + other modules added to the jupyter scipy docker image to make the code generalized enough to work with different instances.
• Stream prefiltered keywords list for each class (EG, and GULF). Requires a crone job in order to:
  • Collect 1-username, 2-tweet, 3-location.
  • Decode Arabic Unicode characters.
  • Store data as jsonl or json on AWS instance.
  • Automatically restart tweet streams in case of common errors.

Storing:

• Store raw data into Mongo collection (e.g: raw_gulf, with documents being raw_stream and raw_timelines).
• Raw data remains stored on AWS instance.

Infrastructure:

• Two t2.micros with unique oauth to stream two dialects to decrease chances of dialects mixing.
• One t2.large for modeling and more computationally expensive tasks.

Munging/Cleaning/Storing the Data

Instructions on working with resulting datasets using pandas DataFrames are provided within the related Jupyter Notebooks.

Cleaning data:

• Using regex to filter out emojis, links, http, excluded Arabic unicode in many cases. An easier way to clean the data is to import tweet-preprocessor, the twitter preprocessing package provided in the requirements.txt file.
• Check for duplicates before converting document formats.
• Pickle cleaned data into a seperate folder (e.g: gulf_twitter_pickled).

Storing data in MongoDB:

• Storing should be taking place at each stage of the process.
• Build up corpus, store in Mongo collection as two documents for each class, EG and Gulf.
• Store combined documents under a new collection on Mongo.
• Store cleaned data into Mongo collection (e.g: cleaned_gulf, with documents being cleaned_stream and cleaned_timelines).

Basic EDA/Visualization

• Inspect keyword documents for excessive advertisement and remove duplicates.
• Inspect geographic origins of keyword documents to determine the document's utility to the overall collection.
• Identify users who contribute most to the keyword stream and add them to the timelines stage

EDA, Tokenization, and SVD

• Perform EDA, tokenization and SVD on collected data:
  • Check for term co-ocurrences in EG and Gulf documents and add to stopwords list.
  • Subtract co-occurances of terms between dialects from the data before tokenization?
  • Identify dialectically different keywords and include in the twitter streaming pipeline.
  • Identify users with the richest dialectal tweets and add them to timeline streams.
  • Confirm geographic origin of tweets and make term substitutions in stop word list as needed.
  • Continue rinsing and repeating until terms appear mostly in either one or the other documents.
• Repeat process for user timelines using Twitter REST API
• Optional: Stanford Arabic Parser (with built-in ATB) to lametize and seg the data. Use Stanford Arabic Word Segmenter concurrently with Parser, before or after?
• Use the three techniques below and explore best results:
  • Tfidf, SVD, latent semantic analysis
  • Okapi best match, SVD, latent semantic analysis
  • Kullback-Leibler Divergence Model, SVD.

Train/Test Estimators on Collected Data (Classes: EG & GULF)

Classifiers:

• Naive Bayes
• Multinomial LR classifiers
• Logistic Regression

Results:

• Perform plotting, confusion matrix, classification report, roc curve, etc.
• Optional: Clustering estimators, DBSCAN, KMeans, Spectral Clustering

Deep Learning - Text Classification with RNN

• Word2Vec
• Word embeddings using Keras or Gensim

Authors

• Tamir ElSahy

Acknowledgments

• Full acknowledgments available in the file titled Building Datasets for Dialect Classifiers using Twitter.pdf contained within this repo.