![Example
• Title = “Junior Software Engineer”
• Domain keywords Set = [“Software”]
• Functional keywords Set = [“Engineer”]
• Attribute Keywords set = [“Junior”]
Title = “Junior Software Engineer”
Map Domain, Functional, Attribute keyword sets of one title to those of the
other title](https://image.slidesharecdn.com/titlematching-181114162612/85/Similarity-computation-exploiting-the-semantic-and-syntactic-inherent-structure-among-job-titles-8-320.jpg)
![Methods
• Objective: Any job title can be split into the attribute, functional and core descriptor/domain words.
• Input:
• Job Title (T)
• Output:
• 3 sets , Attribute words set (SA), functional words set (SF) and core descriptor/domain words set (SD)
• Resources/ Existing techniques used:
• Acronym dictionary (DictA ), Spell checker technique (TechS ), Classifier model (Mclass)
• Algorithm:
• Step 1: SWord = split the title T into separate words
• Step 2: for each word in Sword
• Step 2.1: word = resolve acronyms of word using DictA
• Step 2.2: word = resolve the spelling mistake using TechS
• Step 2.3: classify word using Mclass as either a Attribute (A) word or a functional word (F) or a core descriptor/domain word (D)
• Step2.4: Append word to the corresponding set (SA , SF , SD ) depending upon it’s class label (A, F, D)
• Feature vector used in Classifier model (Mclass):
• [POS (part of speech) of the word, position of the word in job title (T) (first word/last word/in between
word), POS of the root word for each word, word ends with “er”/”or”/”ar” or not]](https://image.slidesharecdn.com/titlematching-181114162612/85/Similarity-computation-exploiting-the-semantic-and-syntactic-inherent-structure-among-job-titles-9-320.jpg)
Similarity computation exploiting the semantic and syntactic inherent structure among job titles
- 1. Similarity Computation Exploiting the Semantic and Syntactic Inherent Structure Among Job Titles Authors: Sarthak Ahuja1, Joydeep Mondal1, Sudhanhsu Shekhar Singh1 and David Glenn George2 1 IBM Research Lab, India 2 IBM Talent Management Solutions, Portsmouth, UK Presenter: Joydeep
- 2. What exactly it is trying to solve?
- 3. List of Available Job Titles • System Engineer • Software Developer • Senior Software Engineer • Junior Network Engineer • Junior Software Tester Query Job Title • Junior Software Engineer No Other information (job descriptions or other details except TITLE) is available corresponding to these jobs Similarity Computation Similarity ComputationSimilarity Computation Similarity Computation Similarity Computation Best Match
- 4. Business Application (Where & Why it is needed?)
- 5. • IBM Watson Recruitment (IWR) : https://www.ibm.com/talent- management/hr-solutions/recruiting-software Mapping requisition jobs to the available job taxonomy without using computation intensive and time consuming sate of the art document similarity methods by narrow down the search space
- 6. How the problem has been solved?
- 7. Job Title Matching Split Title keywords into Three categories (Domain, Functional, Attribute) Map each category of one job title to those of the other title
- 8. Example • Title = “Junior Software Engineer” • Domain keywords Set = [“Software”] • Functional keywords Set = [“Engineer”] • Attribute Keywords set = [“Junior”] Title = “Junior Software Engineer” Map Domain, Functional, Attribute keyword sets of one title to those of the other title
- 9. Methods • Objective: Any job title can be split into the attribute, functional and core descriptor/domain words. • Input: • Job Title (T) • Output: • 3 sets , Attribute words set (SA), functional words set (SF) and core descriptor/domain words set (SD) • Resources/ Existing techniques used: • Acronym dictionary (DictA ), Spell checker technique (TechS ), Classifier model (Mclass) • Algorithm: • Step 1: SWord = split the title T into separate words • Step 2: for each word in Sword • Step 2.1: word = resolve acronyms of word using DictA • Step 2.2: word = resolve the spelling mistake using TechS • Step 2.3: classify word using Mclass as either a Attribute (A) word or a functional word (F) or a core descriptor/domain word (D) • Step2.4: Append word to the corresponding set (SA , SF , SD ) depending upon it’s class label (A, F, D) • Feature vector used in Classifier model (Mclass): • [POS (part of speech) of the word, position of the word in job title (T) (first word/last word/in between word), POS of the root word for each word, word ends with “er”/”or”/”ar” or not]
- 10. • Why we used these features? • POS (part of speech) of the word : We found most of the attribute-words are adjectives, e.g. Senior, Junior etc., most of the functional-words are noun, e.g. developer, tester, teacher and most of the core descriptor/domain words are also noun, e.g. Software, Network etc. • position of the word in job title (T) (first word/last word/in between word) : We found that attribute-words are generally the first or last words of the title e.g.: Senior software developer, Network administrator junior etc. Most of the functional-words appear as in- between or last word of the title e.g.: Senior software developer, Network administrator junior etc. We also found that most of the core descriptor/domain words appears as in-between or first word in a title e.g.: Senior software developer, Network administrator junior etc. • POS of the root word for each word : Our analysis showed that POS of the root word corresponding to the functional-words are verb, e.g. : Senior software developer : root word for developer = “develop” which is a verb. We used https://www.vocabulary.com/dictionary/ open source online dictionary to get the root words. • word ends with “er”/”or”/”ar” or not: We also found that most of the functional words end with either of these three substrings “er”/”or”/”ar”, e.g. : teacher, developer, engineer etc.
- 11. I’m the Best! Functional classifier o/p -> input of Attribute Classifier Functional Classifier o/p + Attribute Classifier o/p -> input of Domain Classifier
- 12. Methods Objective: mapping three category-set of words (Attribute, Functional and core descriptor/domain) corresponding to the two titles among themselves using classical imbalanced assignment problem. Then the mapping scores are combined based on weighted or hierarchical scoring scheme to generate job title similarity. • Input: • Job Title1 (T1), Job Titl2 (T2) • Output: • Similarity score (s) between T1 and T2 • Resources/ Existing techniques used: • Wordnet Dictionary API (W), Hungarian method to solve imbalanced assignment problem (TH) • Algorithm: • Step 1: extract (SA1 , SF1 , SD1 ) from T1 and (SA2 , SF2 , SD2 ) from T2 by previous method • Step 2: Get the mappings as MA(SA1 : SA2 ), MF(SF1 : SF2 ) and MD(SD1 : SD2 ) by TH • Step 3: calculate the mapping similarity score simA , simF and simD for MA , MF and MD respectively. • Step 4: S = simD (1+ simF (1 + simA ))/ (IndicatorD + IndicatorF + IndicatorA ) // importance order : D, F and A respectively. • We used Wordnet Dictionary API (W) to calculate semantic similarity between two words. We built a semantic similarity score matrix for each pair of sets (SA1 : SA2 ), (SF1 : SF2 ) and (SD1 : SD2 ) and provide this matrix to TH as input. We also use the same matrix to calculate simA , simF and simD for MA , MF and MD.
- 14. System Architecture Diagram + Example
- 15. Results
- 16. Core Novelty 1 . Any job title can be split into three categories the attribute, functional and core descriptor/domain words. 2. Job title similarity calculation involves mapping of these three categories of words corresponding to the two titles among themselves using classical imbalanced assignment problem. Then the mapping scores can be combined based on weighted or hierarchical scoring scheme to generate job title similarity. 16