Since the launch of its public beta in 2008, Glassdoor has become the gold standard for company satisfaction ratings. They also have a classifieds section comparable to LinkedIn, Indeed, and others.

Glassdoor provides examples of job descriptions with company ratings in context. This makes it easy to compare what good and bad companies have in common when it comes to word usage.

Gathering the data

First, I downloaded more than 8,000 job postings from some of America’s most common jobs:

Retail Salesperson, Cashier, Office Clerk, Registered Nurse, Waiter, Customer Service, Janitor, Counter Attendant, Material Mover, Stock Clerk, Secretary, Bookkeeping Clerk, Accounting Clerk, Auditing Clerk, General Manager, Truck Driver, Elementary School Teacher

I formatted each listing like this:

Next, I grouped all listings into one job_descriptions.txt file, for training a word segmentation model. I lower-cased the descriptions to boost performance.

Using code from Word Segmentation in Python, we can split the job postings into a collection of subwords:

if not os.path.isfile('{0}.model'.format(model_path)):
    spm.SentencePieceTrainer.Train('--input=job_descriptions.txt --model_prefix=job_descriptions --vocab_size=6000 --split_by_whitespace=False')
sp = spm.SentencePieceProcessor()
sp.load('job_descriptions.model'.format(model_path))
# To tokenize, use: sp.EncodeAsPieces(job_description)

Looking at the data

The top 25% of posts are from companies with 3.8⭐ or more. The bottom 25% are from companies with 3.1⭐ or less.

Next, I used the word segmentation model to create a frequency-normalized bag of words model. Then I compared the word frequencies of well-rated companies to poorly-rated companies.

Phrases to use

The following words were used at least 10 times in both good and bad listings, and occur at least 2x more frequently in good listings as bad listings:

Phrases to avoid

The following words were used at least 10 times in both good and bad listings, and occur at least 2x less frequently in good listings as bad listings:

Putting it in context

provides professional nursing care for the comfort and well-being of patients. prepares equipment and assists physician during examinations and treatments. administers prescribed medications, changes dressings, cleans wounds, and monitors patient vital signs. observes and maintains records on patient care, condition, reaction, and progress. requirements graduate of an accredited school of nursing current louisiana state rn license current american heart association basic life support must be able to communicate effectively and professionally with patients, families, and medical staff both verbally and written one (1) year of gi nursing experience preferred

matches highly skilled professionals with accounting finance jobs at the best companies on a temporary and temporary-to-hire basis. our mission is to provide you with a rewarding finance or accounting job that is well matched to your professional skills helping you to advance in your career. our experience, combined with the resources of our worldwide network of offices, makes us a great resource for your career. we offer excellent opportunities to find temporary accounting and finance jobs for all experience levels. from accounting clerks and bookkeepers to accounts payable and staff accountants, we can provide you unparalleled access to exciting career opportunities. but don’t take our word for it. our company has appeared on fortune® magazine’s list of “world‘s most admired companies” since 1998, and 9 out of 10 of our customers would recommend our service to a colleague. contact your local accountemps office at 888-490-3195 or visit our website to apply for this job now or find out more about other job opportunities. all applicants applying for u.s. job openings must be authorized to work in the united states. all applicants applying for canadian job openings must be authorized to work in canada. © 2019 an equal opportunity employer m/f/disability/veterans by clicking ‘apply now‘ you are agreeing to robert half terms of use.

Takeaways

Avoid clichés

Avoid phrases like “don’t take our word for it”, and “each and every”. Clichés are not only inauthentic, but also a waste of valuable space.

Avoid stating the obvious

The phrase “described here”, and “find out more” is negative. “Connect with us” and “facebook” are positive.

Don’t try to oversell

Companies with low ratings tend to use phrases like “exciting career”, “our mission is”, “everything we do”, and “award winning”.

Nobody wants to read a list of rules

Phrases like “required by”, “all applicants”, “is expected” don’t do very well. Don’t focus on what the employee can do for you.

Do focus on what employees can do for others. Phrases like “ensure quality”, “customer focus” are good. Phrases like “highly desired” make employees feel wanted.

Employees like choices

Phrases like “or part time”, “other duties”, “also includes”, and “negotiable” are positive.

Employees like benefits

Phrases like “weekly pay”, “plus benefits”, and “rewards program” do well. Not all benefits are good; “sign on bonus” is a negative phrase.

Certain industries are just more enjoyable than others

The registered nurse jobs tended to do very well. Others did not do so well. Attractive industries tend to be the most efficient. As an employee they’re the hardest jobs to get, and as a business owner, they’re the hardest to compete in.

The same job can sound more appealing depending how it’s phrased. “Prepare food”, and “foodservice” is a positive phrase. “Food preparation” is a negative phrase.

Bonus: applying LDA topic modeling

We can apply latent dirichlet allocation (LDA) topic modeling to job descriptions to find similar jobs. This is a natural extension of bag of words, and other naive frequency models. Documents have a distribution of topics, and topics are a distribution of words.

Here is the code:

tokenized_descriptions = [sp.EncodeAsPieces(doc) for job_description in job_descriptions]
dictionary = gensim.corpora.Dictionary(tokenized_descriptions)
dictionary.filter_extremes(no_below=4, no_above=0.3)
corpus = [dictionary.doc2bow(tokenized_description) for tokenized_description in tokenized_descriptions]
if not os.path.isfile(model_path):
    print('Training topic model...')
    lda_model = gensim.models.ldamulticore.LdaMulticore(corpus=corpus,num_topics=20,id2word=dictionary,chunksize=100,workers=7,passes=50,eval_every=1,per_word_topics=True)
    lda_model.save(model_path)
else:
    lda_model = gensim.models.ldamodel.LdaModel.load(model_path)
    lda_model.print_topics()

Example job description:

the primary responsibilities of the janitor a involve the upkeep of the property in order to enhance and maintain its curb appeal. the janitor a will also assist the rest of the staff, as directed. duties and responsibilities: conduct all business in accordance with the habitat company’s policies and procedures, state and federal laws; e.g. osha, ada, fair housing etc. perform activities to maintain the continuity of essential environmental services in order to guarantee resident comfort and security. maintains awareness of proper safety precautions at all times. complete assigned work orders according to company guidelines. complete final work of all units turned over for leasing. performs electrical repairs on items such as appliances, fixtures, switches, outlets, circuits, etc. performs plumbing work, such as clearing stoppages, replacing fittings, etc. replaces tile, carpet, screens, garbage disposals, fixtures, appliances, blinds, locks, etc. paints interior and exterior. performs scheduled maintenance on all equipment based on the manufacturer’s recommendations and operating manuals. has knowledge regarding water and gas cutoffs, all units and fixtures cutoffs and sewer clean outs. assists in unit move in and move out inspections. removes and transfers heavy appliances and equipment from storage area to apartment as necessary. full time maintenance experience required concern for work-safety programs. ability to participate in a standby emergency schedule for evening, weekend, and holiday coverage. strong plumbing skills required successful completion of the seiu local 1 two-year apprenticeship program or ability to complete upon hire

The LDA model correctly identified the topic:

67.0% topic 17
0.006” warehouse” + 0.006” cleaning” + 0.006” equipment” + 0.004” items” + 0.004” clean” + 0.004” inventory” + 0.004” perform” + 0.004” must be able to” + 0.004” orders” + 0.004” safety” + 0.003” duties” + 0.003” stock” + 0.003” areas” + 0.003” area” + 0.003” responsible for” + 0.003” shift” + 0.003” proper” + 0.003” maintenance” + 0.003” maintain” + 0.003” product” + 0.003” up to” + 0.003” as needed” + 0.003” materials” + 0.003” in a” + 0.003” ensure” + 0.003” standards” + 0.003” daily” + 0.003” lift” + 0.003” as assigned” + 0.002” supplies” + 0.002” lbs” + 0.002” any” + 0.002” physical” + 0.002” products” + 0.002” order” + 0.002“ers” + 0.002” service” + 0.002” pounds” + 0.002” food” + 0.002” merchandise” + 0.002” re” + 0.002” procedures” + 0.002” trash” + 0.002” shipping” + 0.002” etc” + 0.002” floors” + 0.002” furniture” + 0.002” working” + 0.002” clean and” + 0.002” must be” + 0.002” safe” + 0.002“up” + 0.002” other duties” + 0.002” assigned” + 0.002” such as” + 0.002“w” + 0.002” using” + 0.002” and other” + 0.002“g” + 0.002” forklift” + 0.002” school diploma” + 0.002” complete” + 0.002” hours” + 0.002” other” + 0.002” policies” + 0.002” receiving” + 0.002” general” + 0.002“in” + 0.002” cleanliness” + 0.002” performs” + 0.002” position” + 0.002” hand” + 0.002” 50″ + 0.002” floor” + 0.002” &” + 0.002” to the” + 0.002” operate” + 0.002” department” + 0.002” /” + 0.002” tasks” + 0.002” attendant” + 0.002” job description” + 0.002” maintains” + 0.002” when” + 0.002” up” + 0.002” accurately” + 0.002” manner” + 0.002” follow” + 0.002” supervisor” + 0.002” or equivalent” + 0.002” lifting” + 0.002” move” + 0.002” must” + 0.002” and equipment” + 0.002” may” + 0.001” communicate” + 0.001” check” + 0.001” quality” + 0.001” use” + 0.001” storage”

Future improvements

This analysis was done on 25 MB worth of job description text. Ten to one hundred times more data would have been better. I also did not do any data balancing/controls. The purpose of this post was to show the process, and some of the fundamentals of Applicant Tracking Systems (ATS).