search

elainespak / glassdoor_aspect_based_sentiment_analysis

ABSA with Glassdoor company reviews
8 stars 0 forks source link

readme

Exploring Employee Experience & Discovering Company Characteristics with Glassdoor Reviews

Anonymous company rating and review websites, such as Glassdoor, are a platform for employees to anonymously write about their work experience. Glassdoor data is extremely rich, both in quantity and dimensions.

A user can anonymously write about her experience in the 'summary,' 'pros,' 'cons,' and 'advice' categories, and rate her experience in the following 9 aspects: Overall, Culture & Value, Career Opportunities, Senior Management, Work/Life Balance, Compensation and Benefits, Approval of CEO, Recommendation to Friend, Business Outlook

Data Collection

I created a custom webscraper and collected Glassdoor review data of every company that belonged to S & P 500 at least once from year 2008 to 2018. Please refer to the 'webscrape' folder.

Text Preprocessing

  1. Select only English reviews
    • Pre-trained FastText model (considered a SOTA for language identification as of July, 2020) (for Windows, best way to download is to get the appropriate wheel file at this website and run pip install ~wheelname~.whl)
    • Among 'summary,' 'pros,' 'cons,' and 'advice,' select the longest text. Check if the longest text is predicted as English or not. If yes, keep the review. If not, discard it.
  2. Remove stopwords and punctuations
  3. Make every letter lower case
  4. Create bigrams and trigrams
  5. Apply stemming

Result: 1,401,126 sentences from 741 companies

Aspect Extraction

  1. Apply Chi-Square statistics from Latent Aspect Rating Analysis (LARA)
    • Select relevant aspect: culture and value; career opportunities; senior leadership; work-life balance; compensation and benefits; approval of CEO; recommendation to friend; business outlook
    • Seed words:
      • culture and value: ['cultur', 'valu']
      • career opportunities: ['career', 'opportun']
      • senior leadership: ['senior', 'leadership', 'manag']
      • work-life balance: ['life_bal', 'life', 'balanc']
      • compensation and benefits: ['compens', 'benefit']
      • business outlook: ['busi', 'outlook', 'futur']
    • Final words:
      • culture and value: ['cultur', 'valu', 'compani', 'core', 'corpor', 'strong', 'add', 'collabor', 'larg', 'big', 'within', 'move', 'innov', 'stabl', 'size', 'chang', 'organ', 'forward', 'global', 'promot', 'year', 'constant', 'polit', 'world', 'huge', 'structur', 'last', 'ago', 'past', 'layoff', '5', 'organiz']
      • career opportunities: ['career', 'opportun', 'advanc', 'growth', 'path', 'learn', 'grow', 'lot', 'room', 'limit', 'develop', 'new', 'plenti', 'progress', 'train', 'technolog', 'skill', 'potenti', 'slow', 'curv', 'program', 'provid', 'latest', 'invest', 'profession', 'intern', 'softwar', 'tool', 'hire', 'process', 'cross', 'talent']
      • senior leadership: ['senior', 'leadership', 'manag', 'upper', 'store', 'level', 'middl', 'micro', 'team', 'member', 'poor', 'district', 'lack', 'assist', 'entri', 'commun', 'execut', 'mid', 'director', 'leader', 'lead', 'account', 'incompet', 'support', 'direct', 'decis', 'lower', 'staff', 'advic', 'style', 'sr', 'listen', 'top']
      • work-life balance: ['life_bal', 'life', 'balanc', 'work', 'famili', 'worklif', 'home', 'person', 'flexibl', 'hour', 'schedul', 'hard', 'environ', 'day', 'weekend', 'week', 'long', 'time', 'shift', 'holiday', 'fun', 'full', 'part', 'night', 'enjoy', 'vacat', 'get', 'done', 'paid', 'sick', 'easi', 'peopl', 'rid']
      • compensation and benefits: ['compens', 'benefit', 'great', 'good', 'pay', 'health', 'packag', '401k', 'place', 'decent', 'insur', 'match', 'salari', 'low', 'competit', 'discount', 'averag', 'need', 'start', 'dental', 'medic', 'compar', 'overal', 'mani', 'increas', 'bonu', 'wage', 'make', 'perk', 'sometim', 'like', 'ok']
      • business outlook: ['busi', 'outlook', 'futur', 'unit', 'model', 'run', 'analyst', 'uncertain', 'line', 'bottom', 'front', 'financi', 'oper', 'custom', 'strategi', 'servic', 'rude', 'focu', 'sale', 'deal', 'repres', 'rep', 'associ', 'goal', 'sell', 'product', 'specialist', 'solut', 'consult', 'focus', 'client', 'floor', 'unrealist']
  2. Apply Attention-Based Aspect Extraction (ABAE)
    • Original code published and generously shared by the authors here. Check the link for specific steps.
    • Note for Windows users: The code must run with Python 2.7; For training, run the following to avoid errors: set THEANO_FLAGS=device=cuda,floatX=float32 & python train.py --emb ../preprocessed_data/$domain/w2v_embedding --domain $domain --aspect-size $k -o output_dir -v $vocab_size; Similarly, for evaluation, run the following: set THEANO_FLAGS=device=cuda,floatX=float32 & python evaluation.py --emb ../sample_data/$domain/w2v_embedding --domain $domain --aspect-size $k -v $vocab_size
    • For our purposes, we reached the best result when k=12 and vocab_size=12000
    • Final words on Pros (truncated to the first 10 words):
      • ['benefit', '401k_matching', 'insurance', 'eap', 'benfits', 'tuition_reimbursement', 'medical_dental_vision', 'benifits', 'espp', 'profit_sharing']
      • ['employee', 'leadership', 'management', 'input', 'sincere', 'constructive', 'evident', 'transparency', 'subordinate', 'strongly']
      • ['hour', 'schedule', 'time', 'scheduled', 'weekend', 'unpaid', 'appointment', 'day', 'home', 'weekday']
      • ['pay', 'wage', 'salary', 'payout', 'compensation', 'rate', 'paying', 'payouts', 'commision', 'hourly']
      • ['breakroom', 'catering', 'cooky', 'freebie', 'salon', 'merch', 'themed', 'wardrobing', 'swag', 'clothing']
      • ['midtown', 'location', 'office', 'centrally_located', 'london', 'hq', 'conveniently_located', 'located', 'tampa', 'boston']
      • ['quit', 'told', 'didnt', 'anyway', 'wont', 'sad', 'eventually', 'remember', 'somewhere', 'saying']
      • ['people', 'colleague', 'coworkers', 'teammate', 'atmosphere', 'environment', 'ppl', 'collegues', 'enviroment']
      • ['research', 'analysis', 'analytics', 'instrumentation', 'analytic', 'implementation', 'database', 'deployment', 'computing', 'instrument']
      • ['company', 'compnay', 'conglomerate', 'financial_institution', 'corporation', 'telco', 'comapny', 'merck', 'symantec', 'firm']
      • ['opportunity', 'possibility', 'opportunites', 'opportunties', 'opps', 'avenue', 'oportunities', 'oppurtunities', 'oppertunities', 'oppurtunity']
      • ['ruin', 'stunning', 'stepped', 'transformed', 'ruined', 'unprofessional', 'reflection', 'appeared', 'mantra', 'heaven']
    • Final words on Cons (truncated to the first 10 words):
      • ['reorgs', 'headcount_reduction', 'orgs', 'reorganization', 'restructurings', 'reorg', 'restructures', 'restructuring', 'rifs', 'reorganisation']
      • ['company', 'corporation', 'qualcomm', 'firm', 'eastman', 'hilton', 'raytheon', '3m', 'institution', 'compnay']
      • ['business', 'integrate', 'creation', 'executing', 'operational', 'architecture', 'product', 'implementation', 'delivering', 'execute']
      • ['hour', 'weekday', 'week', 'work', 'day', 'noon', 'workday', 'shift', 'peak_season', 'monday_friday']
      • ['really', 'think', 'honestly', 'laugh', 'love', 'gonna', 'want', 'know', 'guess', 'smile']
      • ['management', 'managment', 'mgmt', 'mgt', 'mangement', 'managemnt', 'leadership', 'managament', 'managerment', 'managent']
      • ['pay', 'salary', 'wage', '401k_matching', 'compensation', 'bonus', 'payout', 'profit_sharing', 'substantially', 'payouts']
      • ['promotion', 'opportunity', 'skillset', 'advancement', 'mentoring', 'upward', 'candidate', 'position', 'role', 'lateral']
      • ['po', 'register', 'supposed', 'submitted', 'backup', 'receipt', 'confirmation', 'verify', 'ordered', 'promo']
      • ['stressfull', 'demanding', 'strenuous', 'exhausting', 'physically_exhausting', 'unfulfilling', 'fast_paced', 'unnecessarily', 'hectic', 'stressful']
      • ['employee', 'passenger', 'costumer', 'customer', 'associate', 'employes', 'baristas', 'guest', 'coworkers', 'employess']
      • ['resembles', 'shabby', 'describes', 'transformed', 'egypt', 'disgraceful', 'dot_com', 'infested', 'refers', 'ultra_conservative'] 
        @InProceedings{he-EtAl:2017:Long2,
author    = {He, Ruidan  and  Lee, Wee Sun  and  Ng, Hwee Tou  and  Dahlmeier, Daniel},
title     = {An Unsupervised Neural Attention Model for Aspect Extraction},
booktitle = {Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
month     = {July},
year      = {2017},
address   = {Vancouver, Canada},
publisher = {Association for Computational Linguistics}
}