I'm trying to find job titles in job ads, but the same title can be written lots of different ways. An "RN" is the same as a "Registered Nurse", and broadly the same role as "Registered nurses". As a preprocessing step to job title discovery I need to normalise the text.

The process I use is simple:

1. rewrite terms containing of, e.g. "Director of Sales" to "Sales Director"
2. Expand puntuation with whitespace; e.g. "Receptionist/Administrator" to "Receptionist / Administrator"
3. Singularize each word; e.g. "Cleaners" to "Cleaner"
4. Expand known acronyms; e.g. "DBA" becomes "Database Administrator"
5. Lowercase the text; e.g. "iOS Developer" becomes "ios developer"
6. Replace common variants and misspellings; e.g. "Adviser" becomes "Advisor"
7. Replace muliple whitespace with a single space

In code this looks like:

def normalise_text(text, acronyms=None, variants=None):
text = rewrite_of(text)
text = expand_punctuation(text)
text = singularize(text)
text = expand_acronym(text, acronyms)
text = text.lower()
text = expand_acronym(text, variants)
text = compress_whitespace(text)
return text

This works reasonably well, but the order really matters and it's a little bit fragile. We need to expand_punctuation before we singularize, because singularize tokenizes on spaces and so "Receptionists/Administrators" would be singularized to "Receptionists/Administrator", but "Receptionists / Administrators" would correctly transform to "Receptionist / Administrator". We need to singularize before we expand_acronym, so that for example RGNs can be transformed to RGN before expanding to Registered General Nurse. We need to lowercase before expanding variants, because they apply for any casing.

There are still some cases where this goes wrong, like ENGINEERS will not be singularised correctly unless we insert a second round of singularisation after lower casing. But it does a pretty reasonable job.

We could be a lot more aggressive with out normalisation, in particular we could use stemming instead of making words singular, drop stop words and remove all punctuation. However this would potentially lose some useful linguistic information, and I would rather gradually remove these as needed (by examining output data) rather than doing it all up front.

The rest of this article goes over each piece, except for rewrite_of and singularize which are covered in their own articles

Expanding Punctuation

This is a simple process of putting extra space around each punctuation mark. This helps downstream processes that rely on processing separate tokens work. One caveat is we need to be careful with punctuation that can be part of an acronym (like A&E). Perhaps it would be safer to do this just at the boundaries of words; but in practice this works well enough on the Adzuna dataset.

EXP_PUNC_RE = re.compile('([/()\'":,])+')
def expand_punctuation(text):
return EXP_PUNC_RE.sub(r' \1 ', text)

Expanding Acronyms

There are lots of common acronyms that need to be expanded to be matched. This is a simple process of substituting at word boundaries.

def expand_acronym(title, acronyms):
for source, target in acronyms.items():
title = re.sub(fr'\b{source}\b', target, title)
return title

Here's the list of acronyms I used:

acronyms = {
'PA': 'Personal Assistant',
'RGN': 'Registered General Nurse',
'RMN': 'Registered Mental Health Nurse',
'NQT': 'Newly Qualified Teacher',
'CEO': 'Chief Executive Officer',
'MD': 'Managing Director', # Medical doctor doesn't occur often here
'EA': 'Executive Assistant',
'GP': 'General Practitioner',
'ODP': 'Operating Department Practitioner',
'A&E': 'Accident and Emergency',
}

Normalising variants

Spelling variants is a very similar problem to acronyms, picking a common target way of writing something. I reuse the expand_acronyms function on a list of variants; I do this after lowercasing to get common forms.

variants = {
}
WHITESPACE_RE = re.compile(r'\s+')
return WHITESPACE_RE.sub(r' ', text)