Client Resume Matcher

Connecting Employees to Employers

A client company is looking for employees. Employees are looking for jobs. Employees send in a resume and the company sends in a job description. Extract exact which resume matches the job description the best. YOUTUBE VIDEO

Learning more about how it works

All my reasearch and powerpoint descriptions are in this directory https://github.com/Kunal2341/ClientResumeMatching/tree/master/DocumentsEXPLAINING

Architecture Diagram

Example architecture of the process of the program. Lays out the details for database connections and fucntions deployment.

Naive Baise Classifier

The following Jupyter Notebook gives a step-by-step description on Naive Baise Classifier https://github.com/Kunal2341/ClientResumeMatching/blob/master/Jnotebooks/Naive%20Bayes%20Classifier.ipynb

External APIs

Uses GCP cloud platform and their vision, NLP, and entity extraction APIs which support the process of the program

• Vision API
  • Uses OCR to convert PDFs to readable files and gives location information about the text groups
• NLP API
  • Extract Basic entities from document
• Entity Extraction API
  • Using custom built AI to extract entities for common designs of resumes
• Document AI
  • To get the data in a tabular form

GCP Entity Extraction Personal Model

The data is still being generated but the model will be able to extract entities from the resume. Check https://github.com/Kunal2341/ClientResumeMatching/blob/master/Jnotebooks/ALL%20CODE%20.ipynb to learn more.

How to run it

Must install requirments.txt in order to have all the different libaries Open Command Prompt - Navigate to where you installed requirments.txt and run It takes **~ 2 minutes** to run

pip install -r requirements.txt

The file is displayed below

numpy==1.18.5
pandas==1.2.0
google-cloud-documentai==0.3.0
google-cloud==0.34.0
google-auth==1.23.0
google-auth-oauthlib==0.4.2
google-cloud-storage==1.35.1
google-cloud-vision==2.0.0
imutils==0.5.4
opencv-python==4.5.1.48
oauthlib==3.1.0
pdf2image==1.14.0
Pillow==8.1.0
scipy==1.6.0
spacy==3.0.1
Shapely==1.7.1
termcolor==1.1.0
webcolors==1.11.1
pyresparser==1.0.6
en-core-web-lg @ https://github.com/explosion/spacy-models/releases/download/en_core_web_lg-3.0.0/en_core_web_lg-3.0.0-py3-none-any.whl

File Managment

Before Running

└── ResumeFiles
    └── Document_201.pdf
    └── Document_202.pdf
    └── Document_203.pdf

After Running

└── ResumeFiles
    └── Document_201
	    └── Document_201.pdf
	    └── Document_201_1.jpg
	    └── Document_201_1.xlsx
	    └── Document_402_1FINALIMAGERUN.png
	    └── Document_402_11.jpg
	    └── Document_402_14.xlsx
	    └── Document_402_102-10_09
    └── Document_202.pdf
    └── Document_203.pdf

Pdf2image

So for my project I really struggled in converting a pdf to a image so here is a step by step instuction\

1. Install pdf2image
   1. pip install pdf2image
2. Install poppler
   1. This is very important
   2. https://poppler.freedesktop.org/ 3. Click on "https://poppler.freedesktop.org/poppler-0.86.1.tar.xz" 4. This will install a compressed file of poppler so then you use 7zip (must install seperatly) 5. Once you extract it, move the file to your specific location 6. Now you know the poppler location path
3. Make sure you have these libaries imported. Use pip install if not working 4. import PIL 5. from PIL import Image 6. from pdf2image import convert_from_path 7. import os
4. Code Make sure you change your poppler path

def convert_pdf_2_image(uploaded_image_path, uploaded_image,img_size):
    project_dir = os.getcwd()
    os.chdir(uploaded_image_path)
    file_name = str(uploaded_image).replace('.pdf','')
    output_file = file_name+'.jpg'
    pages = convert_from_path(uploaded_image, 200,poppler_path='/Users/kunal/Documents/VdartWorking/Poppler/poppler-0.68.0_x86/poppler-0.68.0/bin/')
    for page in pages:
        page.save(output_file, 'JPEG')
        break
    os.chdir(project_dir)
    img = Image.open(output_file)
    img = img.resize(img_size, PIL.Image.ANTIALIAS)
    img.save(output_file)
    return output_file

os.chdir("/Users/kunal/Documents/")
convert_pdf_2_image("/Users/kunal/Documents/", "invite.pdf", (200,200))

Final Output

Check ExampleResult.txt for an example printed result The Final Output will be the Document_201_1.xlsx file. Inside the file you get data about

• Character_Data
  • Each character a-z and A-Z data
    • Number of those Characters in Document
    • Mode of the Height of Character
    • Mode of the Width of Character
    • Median of the area taken up by the character
    • Mean of the area taken up by the character
    • Mode of the area taken up by the character
    • Largest Area taken up by a character
    • Z-score of the distribution of the area of that character
    • IQR of the distribition of the area of that character
    • Approximate stroke width of all characters
• HorizontalLines_Data
  • Vertical-Position
  • Starting Point
  • Ending Point
  • Length of line
• VerticalLines_Data
  • Horzontal-Position
  • Starting Point
  • Ending Point
  • Length of line
• Faces_Data
  • Top Left X
  • Top Left Y
  • Top Right X
  • Top Right Y
  • Bottom Left X
  • Bottom Left Y
  • Bottom Right X
  • Bottom Right Y
  • Matrix
• Entities_Detected_Data
  • name
  • email
  • mobile_number
  • skills
  • college_name
  • degree
  • designation
  • experience
  • company_names
  • no_of_pages
  • total_experience
• BoldWordsDetected_Data
  • Matrix
  • Score
  • Word
  • Which way it achieves threshold
• Tabular_Data
  • Header
  • Row
  • Value

Character Data

The first 2 rows out of all 56 is shown below. This data can be vital to understand and calculate to understand if a entire word is bold. Check lines 62-330 for more information

Char	NumberOfChars	ModeHeight	ModeWidth	MedianArea	MeanArea	ModeArea	MaxOutlierNum	NumOutlierZScore	NumOutlierIQR	AvgStrokeWidth
a	159	31	10	46.1538	47.7806	46.1538	60.1099	1	17	2.8397
b	21	31	12	35.4067	36.207	35.406	52.1931	0	0	2.8397

Bold Words

Part	Threshold Name	Definition	Number
1	THRESHOLDSYMBOLMEANTESTWORD	Splits words into symbols and run tests for each symbol using the findInfo() and their weights. If the average of all those weights are greater than this number then it is counted	0
1	THRESHOLDPARAGROUPSYMWORD	Uses same method as the avg of symbols but then add another layer of averge of words	3
1	capitalWeight	How much extra weight for capital words	0.6
2	THRESHOLDFORWORD	threshold for the weights of the entire word summary (out of 100) VAR = "weights" FOR ALL THE WEIGHTS 0 AND 1 are True and False for each out of the 4 tests --> mean, median, mode, and outlier	10
2	THRESHOLDFORPARA	Threshold for each word in the paragraph indiviually. Para can have different thresholds as a single word. Also counts in the threshold of "TOTALSUMFINALPARATHRESHOLD"	0
2	CUTOFWORDSTHATARE3ORUNDER	Simple cut off words that are under 3 letters. Watch out for arconmys and for words that are incorectly detected EX: = "&"	T
2	MAXIMUMLENGTHOFWORDSPARA	The maximum number of words that should be in a paragraph for it to be even counted as a possiblity for bold	10
2	TOTALSUMFINALPARATHRESHOLD	threshold for average of the total weights for each word in the paragraph (out of 100) - same as word (just average)	0
2	THRESHOLDFORANYTHINGTHATISNOTALETTER	There are 52 detected possible letters in the english dictionary. If it is a punctionaltion or special character. It defaults to this number for all of the 4 tests (Mean,median,mode,outlier). EX: "." or "?"	30

Faces

An example Image is the following once the face has been highlighted on Document The area is the calculated using a Polygon Matrix ((114, 157), (399, 157), (399, 442), (114, 442)) Each is a point on a polygon as shown in the image

Entities Data

An Example Result is as the following

Key	Data
Name	Serhii Kalinin
Email	kalininsergg92@gmail.com
Mobile Number	437-231-5807
Skills	['Algorithms','Reports','Quality management','Usability','Quality assurance','Automation','Sql','Windows','Linux','Segmentation','Programming', 'Agile','Workflow','Python','Java','Recruiting','System','Javascript'...]
College Name	None
Degree	Bachelor of veterinary medicine
Designation(Race)	['Ukrainian - native']
Experience	['LLC Materialise Ukraine','Apr 2019 — March 2020','Test-Engineer', 'Image recognition algorithms testing;','Manual QA Engineer','April 2018 — April 2019','Start IT Training Center','Febr 2018 — April 2018','QA Trainee', 'Espresso Window', 'Sept 2015 — Febr 2018', 'Small business co-founder and co-owner', ...]
Page Count	5
total_experience	4.67

Horizontal and Veritcal Lines

This can be used to detect page breaks and other major sectioning in the resume

Horizontal Lines

The first 3 lines of an example result of Horizontal lines is the following

Line #	YPos	Starting	Ending	Distance
1	146	1319	1654	335
2	206	1439	1654	215
3	207	1439	1654	215

Vertical Lines

Were not detected in this document but the format will be the same as Horizontal lines except it will be HPos instead of YPos.

Tabular Data

This is from the DocumentAI API from GCP. An example is the following:

Page number: 1
Table 0: 
Header Row 0: Finance Consultant, Consulting Engagements (3 Firms), 2018-2020

Row 0: • Interim o o o Director Of Finance & Interim Controller, Aptitude Health (Engaged by RGP), 2019-2020
Issue & Scope: Director Of Finance for $17 Million multi-entity organization resigned with short notice
Tools & Analysis: Intacct & Excel
Findings & Results: Fulfilled month-end & year-end responsibilities with minimal training & guidance,
enabling organization to continue financial operations without disruption; Performed analysis that yielded

Row 1: 95% decrease in quarterly commissions calculations time; Improved financial processes

Row 2: • IT Finance Consultant, McKesson (Engaged by Strive Consulting), 2019

Entities Document AI

This is using the Document AI API which returnns a list of major values in the Document with their score. An example is shown below:

Entity	Percent Accurate	[Number of Times detected, Start Index, End Index]
kalininsergg92@gmail.com	0.966053	[[1, 59, 91]]
437-231-58-07	0.939058	[[1, 92, 112]]
05/27/1992	0.486468	[[1, 113, 137]]

This can be used to understand what exactly the document is doing

Entites Form Key Value Document AI

This can be much more usefull than the Entities Form because it gives a Key to the Value in a dictonary format

Key	Value	Percent Accurate	Page #
E-mail:	kalininsergg92@gmail.com\n	0.966053	1
Phone:	437-231-58-07\n	0.939058	1
DATE OF BIRTH\n	05/27/1992\n	0.486468	1

Conclusion

This project and give a very conclusive result for any resume. I need to uptade the code so that it will adapt to resumes for multiple pages.

Made by Kunal Aneja through Vdart Digital. kunal.aneja101@gmail.com