Intelligent search system for resume and labor law

Ontology for organizing job applicants and job recruitments

Ontology is the essential paradigm for organizing information that fulfils the requirements of a recruiting system (Chandrasekaran, Josephson & Benjamins, 1999). In the semantic web model, the meaning of intended information, such as data or knowledge, is addressed using ontology (Calaresu & Shiri, 2015). Ontology also goes beyond the semantic web’s most basic idea by granting capabilities to conventional reasoning, which frequently rests on the declaration of inference rules. In the resume-searching system, ontology is used to represent the data that was taken from applicants’ CVs and recruiters’ JDs.

Definition 3.1: The Job-Onto ontology structure represents job applicants and job recruitments as a triple: ${\displaystyle \left(\mathbb{CV},\mathbb{JD},\mathbb{R}e\right),}$ where

• CV denotes the set of applicants’ curriculum vitae,

• JD denotes the set of job descriptions, and

• ℝe denotes the set of relationships between job applications and job descriptions.

Standard ontology languages, such as OWL, can be used to build ontology documents.

Structure of a curriculum vitae

Each curriculum vitae c of an applicant, c ∈ CV has a structure as follows: ${\displaystyle c=\left(Profile,History\text{\_}work,Edu,Courses,Skills,Position,Others\right),}$ where

• Profile: stores the applicant’s personal information.

• History_Work: These were the applicant’s positions at earlier firms.

• Edu: applicant’s educational background.

• Courses: It comprises certain additional training courses completed by the candidate. These courses provide more information regarding the candidate’s comprehension and self-study.

• Skills: the information about the candidate’s skills. There are two kinds of skills:

  • Domain skills: they store the major skills of the applicant in IT. They are divided into two categories: programming environment skills and general programming abilities.

  • Soft skills: they are some of the applicant’s soft skills, such as teamwork, communication skills, leadership, and logical and critical thinking.

• Position: the applicant’s application position.

• Others: other applicant information.

Structure of a job description

Each job description d ∈ JD has the following structure: ${\displaystyle d=\left(Infor,Descript,Req,Skills\right)}$ where

• Infor: general information about the recruited position.

• Descript (Descript = {s₁, s₂, …, s_m}): set of sentences that describe working at the recruited position. Each sentence s_j (1 ≥ j ≥ m) is a set [text, keywords ], where text is a sentence of text to describe the job, and keywords are a set of keywords extracted from text.

• Req: detailed requirements for a candidate.

• Skills: the skills for the recruited position. The categories of talents are organized similarly to a resume. The system may automatically enter this component from the recruiter or derive it from data in the Descript and Req components.

Ontology for representing legal knowledge domain about labor

When users seek suitable jobs, they search for several legal documents regarding labor and work to understand more about labor regulations. This section presents an ontology representing the law documents about labor and work to search for this legal domain. The proposed ontology is an integrated structure of the knowledge model of relations, Rela-model (Do, Nguyen & Selamat, 2018; Nguyen et al., 2021), and the conceptual graph representing the linking of key phrases in law documents (Filtz, 2017).

Vietnamese labor law documents include the Law of the National Assembly and sub-law documents. The Law of the National Assembly is a general rule of conduct, commonly binding and repeatedly applied to agencies, organizations, and individuals nationwide in the labor field. Sublaw documents are legal files that describe the details of the law established by the National Assembly. The built legal ontology has to represent the meaning of those documents and the links between their articles.

Definition 3.2: Ontology representing a legal document, called Legal-Onto, is built based on the Rela-model. It includes the following components: ${\displaystyle \mathbb{K}=\left(\mathbb{C},\mathbb{R},\mathbb{RULES}\right),}$ where

• ℂ is the set of concepts, but each concept in ℂ has improved its internal structure to organize its law information;

• ℝ is the set of relations, and those relations are between concepts, key phrases, and the database storing the legal document’s content and

• RULES is the set of inference rules on the domain of labor laws.

ℂ—the set of concepts

The law contains general principles based on widely accepted concepts. In the real world, in addition to the content, the structure of each concept in a legal document has to present relationships between itself and the articles in the corresponding document.

Definition 3.3: Given a law document d as Legal-Onto, each concept c ∈ ℂ consists of five elements: ${\displaystyle c=\left(Name,Meaning,Rel\text{-}Inner,Attrs,Phrases\right)}$ where

• Name: The legal concept’s name.

• Meaning: The meaning or the content.

• Rel-Inner: a list of the document’s articles connected to the relevant concept.

• Attrs: List of elements (or other concepts) that support the corresponding concept in document d, if one is necessary.

• Phrases: the list of keywords describing the meaning in each document article d.

Example 3.1: The structure of the concepts in the National Certificate of Vocational Skills (Government of Vietnam, 2015) and unemployment insurance in Law on Employment 2013 are organized as in Table 1.

ℝ—the set of relations

The set ℝ is a set of relations. Each relation in ℝ is one of three kinds: ${\displaystyle R=R_{con}\cup R_{key}\cup R_{data}}$ where,

• R_con is a set of relations between concepts in ℂ. These relations are “is-a” (hierarchical relation), “has-a”, “a-part-of”, and several other relations between concepts. ${\displaystyle R_{con}=\left\{r|r\subseteq \mathbb{C}\times \mathbb{C}\right\}.}$ Several properties of each relation r ∈ R_con are considered symmetric and transitive.

• R_key is a set of connections between important terms in the legal document. Many relationships between a concept and key phrases, which are distinctive words that help define a concept’s meaning, are also included in this set. ${\displaystyle R_{key}=\left\{r|r\subseteq Key\times Key\right\}\cup \left\{r^{\prime }|r^{\prime }\subseteq Key\times \mathbb{C}.Phrases\right\}}$

• R_data is a set of relations between concepts and key phrases linked to a database of the law document.

RULES—the set of rules

The regulations within the RULES embody limitations and the derivation of connections among fundamental words and concepts. Employing deductive rules during the construction of ontology data can streamline the responsibilities of a knowledge engineer. The RULES-set infers both direct and indirect links between keywords or concepts, subsequently employed to establish semantic similarity (Nguyen et al., 2022b). A rule rul ∈ RULES conforms to a deductive pattern grounded in information related to pivotal words and concepts, characterized as follows (Nguyen et al., 2022b): ${\displaystyle rul:\left\{h_1,h_2,\dots ,h_m\right\}\to \left\{t_1,t_2,\dots ,t_n\right\}}$ where {h₁, h₂, …, h_m} are hypothesis facts and {t₁, t₂, …, t_n} are goal facts of rule rul. There are four kinds of facts as shown in Table 2.

Example 3.2: Several rules in the knowledge domain:

r₁: if [ Ω is symmetric] and [ k₁Θk₂], then [ k₂Θk₁] (Ω is a relation, Ω ∈ ℝ)

r₂: if [ θ is transitive] and [ k₁θk₂] and [ k₂θk₃], then [ k₁θk₃] (θ is a relation, θ ∈ ℝ)

Keyphrase graphs of law documents

Knowledge of a law document is a keyphrase graph, which is organized using the tube (ℂ, ℝ, RULES) as the structure of the Rela model. While getting legal content, several key phrases in the query have been linked to the information through their semantics. An improved conceptual graph is used in this study to structure the semantics of key phrases.

The structure of the Rela model organizes the knowledge of a law document. However, in practice, when retrieving the content of the law, several key phrases in the query sentence have been connected to the knowledge through their semantics. In this study, the semantics of key phrases are organized by an improved conceptual graph.

Definition 3.4: The structure of the key phrases graph of law document d is a tube: ${\displaystyle \mathbb{KG}=\left(Key,Rel,w_k,w_e\right)}$ where

• Key = {k|k is a key phrase of document d}.

• Rel = {e = (k₁, k₂) ∈ Key × Key|k₁ and k₂ are key phrases that appear in the same law document article}.

• w_k: Key → R ×R is the weighted map to compute the similarity binary vector for each key phrase in Key (ℝ is the set of real numbers). The vector (tf(p, d), idf(p, d)) computes the key phrase similarity measure, where tf(p, d) is the term expressing the frequency of a keyphrase p in a document d, and idf(p, d) is the inverse article frequency showing the specificity of the document d’s keyphrase p. The following is how the formulas for (tf(p, d), idf(p, d)) are calculated (Huy et al., 2021): (1)${\displaystyle tf\left(p,d\right)=c+\left(1-c\right)\frac{n_{p,d}}{\max \left\{n_{p^{\prime },d}|p^{\prime }\in Key\right\}}}$ where n_p,d is the number of occurrences of the key phrase p in document d, c ∈ [0, 1] is a parameter that represents the minimum value for each key phrase. (2)${\displaystyle idf\left(p,d\right)=\log \left[\frac{card\left[\bigcup _{ar\in d}key\left(ar\right)\right]}{1+card\left(A_p\right)}\right]}$ where, A_p: = {ar ∈ Arctics(d)|p ∈ key(ar)} Artics(d) is the set of articles of the law document d, key(ar) is the set of key phrases of article ar in document d.

• w_e: Key × Key →R is the weighted map to compute the similarity binary vector for each key phrase in Key (R is the set of real numbers). (3)${\displaystyle w_e\left(p,p^{\prime }\right)=\frac{card\left(B_{p,p^{\prime }}\right)}{card\left(Arctics\left(d\right)\right)}}$ where, ${\displaystyle B_{p,p^{\prime }}:=\left\{ar\in Arctics\left(d\right)|p,p^{\prime }\in key\left(ar\right),r_{ar}\left(p,p^{\prime }\right)\right\}}$ key(ar) is the set of key phrases in article ar, r_ar(p, p′) is the relation between p and p′ in article ar.

Ontology and keyphrase graph integration

Legal-Onto ontology is used to represent the information within a legal document. The utilization of a knowledge graph is efficient in establishing connections among various components in a legal document. Consequently, achieving a comprehensive model for depicting the content of a legal document needs the combination of both Legal-Onto and the knowledge graph. The combination model is described as follows: ${\displaystyle \mathbb{K}=\left(\mathbb{C},\mathbb{R},\mathbb{RULES}\right)⨁\left(Key,Rel,w_k,w_e\right).}$

The symbol ⨁ means the ontology (ℂ, ℝ, RULES) as Rela-model is integrated with the conceptual graph (Key, Rel, w_k, w_e). The integration is worked through relations between the key terms of concepts in ℂ and key phrases in Key.

The key phrase graph and ontology (ℂ, ℝ, RULES) as the Rela-model are integrated through relations between key phrases in Key and key terms in ℂ.Phrases. These relations are represented in R_Key.

Definition 3.5: Given an ontology (ℂ, ℝ, RULES) as a Rela-model and a key phrase graph (Key, Rel, w_k, w_e). Integrating them works based on relations in the following set: (4)${\displaystyle Integ\text{\_}rel=\left\{r^{\prime }\in \mathbb{R}|r^{\prime }\subseteq Key\times \mathbb{C}.Phrases\right\}.}$

Thus, Integ_rel ⊂ R_Key

From an input query, the key phrase graph helps to quickly extract the semantics in legal documents. Matching the meaning of that query with the knowledge of legal documents works based on relations in Integ_rel. Through that, the process is able to retrieve suitable contents from corresponding law documents.

The knowledge base, represented by the integration of ontology Legal-Onto and knowledge graph, can be applied to build an intelligent recruitment system with the law search system. The section presents the utilization of designing a search system for job applicants in information technology (IT) and Vietnamese labor laws.

Architecture of IT recruitment system

The CSO classifier is an unsupervised technique used by the IT recruiting system to categorize documents based on ontology-based content (Salatino et al., 2019). The syntactic module and the semantic module make up this technique. The Job-Onto ontology represents the information content of CVs and JDs via the CSO classifier. In addition, this classifier also improves the processing of the labor law document to extract key phrases from this document. They are material to build the Legal-Onto ontology for representing labor laws. The workflow of the IT recruitment system combining the searches for labor laws is shown in Fig. 1.

The functions of several modules in the CSO classifier are as follows:

• Syntactic module: This module associates textual n-gram chunks with concepts. The input text is eliminated the stop words before extracting chunks of unigrams, bigrams, and trigrams are collected. The Levenshtein distance similarity (Arockiya Jerson & Preethi, 2023; Government of Vietnam, 2015) with the topic labels in the ontologies is then calculated for each n-gram. The minimum similarity level has been manually set at 0.94. We can distinguish a wide range of distinctions between ideas and ontologies.

• Semantic module: The semantic module was created to identify subjects in the text that are semantically connected. The text makes no mention of such matters. To determine how semantically similar the terms in the text are to those in the ontologies, they use Word2Vec’s word embeddings.

This study uses the Word2Vec model (Salatino & Osborne, 2019; Gao et al., 2022) to construct word embeddings for preprocessing collected data. First, the model is trained for text collection from the CVs and JDs in the IT domain. The Job Posting dataset is collected from Dice.com (https://www.dice.com/) and StackOverflow (https://stackoverflow.com/). ESCO (https://ec.europa.eu/esco/portal) is used for an alternative name and skill description. Finally, 50 thousand sentences containing extracted skills are preprocessed from the Job Post in the Skill2Vec (https://github.com/duyet/skill2vec-dataset) dataset. After the training process, the classifier can extract skills for an input JD or CV. Second, in addition to IT skills, several key phrases in the IT domain and labor have been deduced from the training results. From the labor law document, several key phrases in the labor law are also extracted. All of the produced key phrases are material for building the ontology Legal-Onto to represent the corresponding law document. They are also used to process the query on the meaning of the law document represented by ontology Legal-Onto.

Creating ontologies

The CSO classifier’s documentation is used to create the domain-specific IT ontology from the job postings manually.

• Building ontology Job-Onto in the IT domain: There are 10,000 job listings for each branch of IT, such as frontend developer. Each job post has all of the keywords. After removing all stop words and unrelated terms, a set of skills extracted from all keywords is built by mixing them with Stackshare skills. The domain-specific ontology is constructed using the Protégé program once all of the abilities had been explored. The domain-specific IT ontology was manually generated from the job listings based on its documentation to input them to the CSO classifier.

• Building ontology Legal-Onto for Labor Laws: The legal ontology about labor is created from the collected labor law documents. This ontology includes knowledge components about concepts, relations, and inference rules for this domain. In addition, several common questions and answers for querying labor knowledge are collected from Nguyen (2021). Based on those, the list of key phrases in labor law is extracted and represented as a conceptual graph connecting to the articles of the documents. The created ontology is used to organize the semantics of the corresponding document and tends to search for the labor knowledge of legal documents.

Figure 2 displays the process of creating ontologies for this search system:

• Step 1: Define resources of the knowledge domain; then collect data/knowledge from them.

• Step 2: Extract key phrases from crawled data

• Step 3: Build the content of ontology

  • Step 3.1: Determining the structure and knowledge model. The data conforms to the RDF schema serving the Rela-model.

  • Step 3.2: Convert the collected data according to the specified structure. Construct and initialize classes; the classes are extracted keywords.

  • Step 3.3: Forming the attributes of each class. Each class will have its own characteristics to distinguish classes and represent the semantics of that concept more effectively. Create hierarchical or peer-to-peer relationships between classes by creating hierarchical or peer-to-peer interactions.

Resume search system in information technology

Architecture of the searching system

The resume search system can rank CVs for an input JD and recommend JDs for an input CV. This method implements automatic CV ranking using a set of criteria related to general technical skills. Those skills are extracted from a CV, domain technical skills, domain skills that the JD is looking for, and soft skills. The operation of the resume searching system is described in Fig. 3.

There are several modules in this work:

• Resume parser module: This module takes the uploaded CV and parses it into text for skill extraction.

• Extracting skills module: This module extracts a list of abilities from the JD and CV by categorizing them into four categories: general technical skills, education, domain skills, and soft skills. CSO-Classifier, an unsupervised technique for autonomously categorizing papers according to the Job-Onto filter (Salatino et al., 2019), is used. The extracted skills are represented by a skill graph.

• Matching module: This module uses candidate or JD selection criteria to calculate the relevance score of each criterion for the applied post. The grading function is built using the extracted skills module’s determined similarity of skill graphs. The matching process is presented in the next section.

Moreover, when searching for suitable resumes, the system permits the recruiter to add the weights assigned to the various needed abilities, and the system displays a list of candidate resumes that have been ranked in accordance with their overall ratings.

The method for matching CV and JD

Data collection: There are several datasets available on the internet, but selecting one that fits our needs is exceedingly tough. First, we gather a substantial number of job postings from a variety of employment websites, including Dice.com, Indeed, and others. We extracted the job description from each job posting and cleaned the raw data before feeding it into our word2vec model, detailed later in this section.

We gather and prepare a large dictionary of skills, as recommended in Gugnani, Kasireddy & Ponnalagu (2018), to perform decent skill extraction. This talent dictionary is created by scouring many websites for information on skill terms and field terminology (ESCO, StackShare; https://www.stackshare.io). We are able to create a dictionary containing over 30,000 skill phrases after cleaning and analyzing these terms.

Data processing: extracting text from a CV. First, we assume that parsing a CV from a PDF or Word document into raw text would be straightforward; however, we are mistaken. Because a CV is an unstructured document, building a parser for it is extremely difficult. We test PyPDF2, PDFMiner, and PDFtoTree, as well as other existing tools for parsing PDF formats into text. One frequent blunder made by all the preceding programs is that the extracted text is occasionally placed in the wrong location. Optical character recognition (OCR) is therefore used in the suggested technique. We attempt to parse a CV into text using this method, which includes the steps below:

• Creating pictures from the CV; each image represents a single page of the CV.

• Using OpenCV (https://opencv.org/), create a threshold image.

• Using tesseract (https://github.com/tesseract-ocr/tesseract) OCR to parse text from a threshold picture.

This method has various drawbacks; for example, it takes longer to extract text straight from the source file, and the quality of the retrieved data is poor if the scanned pages are blurry. However, after testing approximately 1000 CVs, this technique yields positive findings that can be used in practice.

Pre-processing data for the Word2Vec model

The gathered data has a variety of structures due to Because the data is gathered from many sources. Therefore, the pre-processing will be somewhat altered. The following stages will often be used to process data:

• Eliminate any HTML tags you find.

• Rearrange sentences into paragraphs.

• Eliminate special characters, but keep a few common in programming languages, such as “+”, “#”, and “C#”.

• Eliminate stop words, which frequently exist in English literature but don’t provide skill phrases any meaning (e.g., a, the, and, or, it, if).

• Lemmatize each sentence.

After separating sentences and word processing, we have a dataset of 1.8 million sentences suitable for training the Word2Vec model.

Extracting skills from CV and JD

The CSO classifier must extract the top 10 related terms for all tokens and compare them to CSO subjects while studying a text, which is a laborious process. A cached model for each domain is used to solve this problem. All tokens in the model’s vocabulary are directly linked to skill terms in the ontology by the cached model, which serves as a dictionary. This item makes it easier to rapidly find all skill phrases implied by a certain token.

Next, a dictionary is created by placing the cached models in it. This extraction module’s user must first select a domain before being able to extract skill phrases from the text, after which a CSO classifier for that domain is generated and utilized for extraction. The technique of obtaining skill terms is briefly depicted in Fig. 4.

Skill graph generation

For feature matching, the extracted skill list is used to generate the skill graph. The domain ontology determines the edges between nodes drawn using the NetworkX graph library. The tree-like created network contains a root that is the domain ontology’s name and relevant skills nodes. The following figures show an example of skill graphs generated from a CV (Fig. 5) and a JD (Fig. 6).

Compatibility between CV and JD information

The graph edit distance (GED) is a similarity measurement between two graphs that we use to match features. In particular, the graph created by a CV will be updated to match the graph from JD using insert/update/delete activities. The graphs become more distinct as the number of transformation activities increases. Gmatch4py, a package specialized for graph matching, is utilized in this stage. It is calculated using a graph edit distance approach that combines Hausdorff matching and the greedy assignment (Fischer, Riesen & Bunke, 2017). The graph structures are saved as objects in the NetworkX graph.

The score of each criterion between CV and JD is obtained after the matching process. There are numerous sorting techniques to rank a set of CVs for each JD. Depending on the scenario and the demands of their firm now, recruiters may have their ideas about the characteristics they wish to emphasize. As a result, our resume search engine for job seekers provides such sorting by allowing recruiters to change the parameters as needed.

Figure 7 presents the recommendation processing of CVs and JD. When a candidate inputs his or her CV, the system parses this CV to text and extracts skills from that document. After that, the system creates a skill graph for the candidate and matches it with stored JDs in the system. In contrast, when a recruiter posts a JD, the system also extracts the required skills from the JD and creates a skill graph for the JD. Then, it matches this with stored CVs in the system.

Search system on labor laws

Let 𝕂 represent a knowledge domain of the ontology Legal-Onto model and the legal document d. The search system processes a search query to extract key phrases when one is entered to get knowledge from 𝕂. These key phrases are matched with the appropriate content in the knowledge base based on their semantics. The matching is carried out by examining the relationships between key phrases that lead to corresponding concepts and the structure of the knowledge model’s constituent parts. The knowledge base’s inference rules aid in the matching process by enabling the deduction of additional relations pertinent to the query. The retrieval outcomes for the entered query are then shown. Several of the most common issues with finding legal documents are as follows: (1) Problem 4.1: The first issue is classifying the query that was entered. This problem takes the important keys of the query from the query supplied as Vietnamese text. Those phrases are used to determine the meaning of the query and classify it. (2) Problem 4.2: The second issue is finding relevant articles in the document and looking for concepts in the text based on keyphrase matches. The set of keywords obtained from the query is used to give a method for assessing how closely the meaning of the keywords matches that of the data in the knowledge base. This technique determines the knowledge material needed for the entered question.

Problem 4.1 has been solved in other studies, such as in Do, Nguyen & Selamat (2018) and Nguyen et al. (2022a). This section only presents the solution of Problem 4.2. Following the extraction of the utterances of keywords and intents, the system will match those words and intents to the knowledge content to define and compare texts using the Legal-Onto ontology as defined in Definition 3.2. Based on the approach in Do, Nguyen & Selamat (2018) and Pham et al. (2020), the matching method for the search engine can be created, but it makes several improvements, as shown in Fig. 8.

The technique also extracts the query’s primary key phrases after categorizing the input query. In Problem 4.2, these key phrases are used to find the appropriate information in the knowledge base. The keyword dictionary is created using the search engine’s knowledge base and input from professionals in the field of law, including attorneys, senior legal staff, and senior law lecturers.

Example 4.1: Several key phrases in the dictionary of the Vietnamese Labor Code 2019:

• Several individuals in the dictionary: representative organization, labor relation, forced labor, labor discrimination, sexual harassment, employment contract, indefinite-term employment contract, fixed-term employment contract, job- or position-based salary, form of salary payment, due date for payment of salary, allowances and other additional payments, regimes for promotion, regimes for pay rise, working hours, rest periods, protective equipment, social insurance, health insurance, unemployment insurance, basic training, advanced training, occupational skill.

• The synonyms of the key phrases in the dictionary: “what is” is equivalent to “define”, and “how to use” is equivalent to “usage”.

The extracted keywords from the speech are compared with the dictionary to provide a collection of keywords. Keywords are used by the search engine to extract legal knowledge from stored knowledge. The system also proposes several relevant pieces of information from the search results based on the relationships between the acquired results and the knowledge base’s inference rules. Finally, for each type of intent, all results are returned in the form of a document.

Algorithm 4.1: Finding matched knowledge of query q

In Algorithm 4.1, a collection of knowledge is returned by the knowledge content search based on the meaning of an inputted query. The machine derives the meaning of this question from the terms it has retrieved. This search is carried out by finding query terms with comparable meanings and utilizing stored information.

Experiment on the resume search system

The CV and JD sets are gathered from Indeed and TopCV. The ontology Job-Onto serves as their representation. ReactJS and Flask are used to implement the resume-search system for IT job hopefuls. The number of collected CVs and JDs in our system is shown in Table 3.

Based on matching scores for domain skills, general skills, and soft skills for each posted JD, the algorithm calculates a correlation score for each candidate’s CV. For instance, Fig. 9 and Table 4 display a list of ranked CVs for the JD of front-end engineers, along with overall scores. Table 2 is the score of the ten highest-score candidates matching the JD about the frontend developer role in Fig. 5B. These scores are computed based on the scores of domain skills, general skills, and soft skills. Those skills have weights from 1 to 3 (less important, important, and very important): “Domain” is 3, “General” is 3, and “Soft” is 1.

In practice, many methods can be used for matching CVs and JDs (Tejaswini et al., 2022; Kostis et al., 2022). However, Moreover, the three baseline methods, including Jaro–Winkler distance (Boudjedar et al., 2021), cosine similarity (Kamacı, 2022), and Levenshtein distance (Arockiya Jerson & Preethi, 2023), are usually applied in practical systems because their effectiveness and profit. Cosine similarity is a metric used to measure how similar two vectors are. Levenshtein distance is a straightforward method for measuring string similarity. When used to match a CV to a JD, these measures can be employed to assess the similarity between the skills and qualifications mentioned in the CV and those specified in the JD. Jaro–Winkler distance is a string similarity metric that measures the similarity between two strings. It is particularly effective for comparing short strings and can be used to compare individual skills extracted from JDs and CVs.

In this experiment, each inputted JD can be extracted for the content of general skills, domain skills, and soft skills. After that, those extracted kills are computed matching scores with CVs by cosine similarity, original Levenshtein distance, and the proposed method. Given a JD, there are two stages for comparing the results of the CV filter of each method for that JD:

• At first, establish lists of emerging CVs that were produced by each method.

• After that, there are two experts in IT recruitment building their lists of CVs for that JD. They will be asked to choose the Top 8, Top 5, and Top 3 of emerging CVs in their list. If there are differences between the two lists, another expert, who is the head of the Human Resources (HR) Department, will be asked to give the final decision.

Finally, the results of each method will be compared with the lists established by the experts. Table 5 compares the matching proportions of each method with the list built by experts in the Top 8, Top 5, and Top 3 for JDs.

The results of matching proportions of each experimented method with the list built by experts in cases.

The experiment results indicate that the proposed method exhibits better-matching proportions than other methods in most cases. Particularly in real-world recruitment scenarios, recruiters typically shortlist only three to five potential candidates for a given job opening. Moreover, the proposed method demonstrates its superior performance when the Top 3 or Top 5 candidates are being evaluated.

Experiment on the search system of labor laws

There are 17 chapters and 220 articles in Vietnam’s Labor Code 2019 (Vietnam National Assembly, 2019), while the Law on Employment 2013 has seven chapters and 62 articles (Vietnam National Assembly, 2013). They are general rules that apply to all labor-related activities. To further structure the legal knowledge base from Decree on Detailing Unemployment Insurance 2015 (Government of Vietnam, 2015), which contains the comprehensive unemployment insurance regulations, is also gathered. First, a database that follows the format Chapter–Section–Articles–Paragraph–Point is used to organize the legal documents. The substance and meaning of several terms in the labor law documents gathered from Government of Vietnam (2020) are then represented using the ontology Legal-Onto.

A labor law search system aims to retrieve labor law knowledge for user needs. This system is able to provide relevant information taken from legal documents when users request definitions for terms in labor legislation or any administrative processes which are linked to their work. In reality, the law search system concentrates on five categories of questions to suit the needs of individuals looking for IT jobs:

• Kind 1: Employment support policy

• Kind 2: Labor market information

• Kind 3: Issuing a national vocational certificate

• Kind 4: Organization and activities of employment services

• Kind 5: Unemployment insurance

First, the user enters a query into the search system. The system classified input questions into definitions, procedures, and/or related knowledge. After that, it returns responses to the users. When receiving results, the user will evaluate their usefulness for himself or herself. In addition, the results are also saved, and they will be examined by a lawyer and a law instructor in labor resources later.

Example 5.1 The input query q1 = “What is unemployment insurance?” The system will extract keywords from query q: “What is,” “unemployment insurance”. The word “what is” is used to classify the query into kinds, declaring the meaning of a concept. The keyword “unemployment insurance” helps to find the concept. Then, from that point, the system returns the result for query q1:

“According to Paragraph 3, Article 3, Law of Employment 2013:

3. Unemployment insurance is a regime to meet a part of an employee’s income when he or she loses a job, to support workers in vocational training, job maintenance, and job search on the basis of contributions to the Unemployment Insurance Fund”.

Table 6 shows the test results on the content about definition queries, asking about an administrative procedure, and related knowledge. The average accuracy of the search is more than 60%. The search can answer queries about definitions correctly. In addition, by using the ontology Legal-Onto to represent relations between knowledge in the law, the system can extract well-related knowledge. However, the system is not good with questions about administrative procedures. To complete an administrative procedure, there are many decisions from the government and instruction decrees from related ministries; thus, the identification of documents for the procedure is not accurate.

Figure 10 visualizes the results of 209 queries based on five kinds of queries. The system is not good when processing queries about labor market information because the data of this market are not collected fully. The system is most effective with queries about unemployment insurance (more than 70% accuracy). This is also an important requirement for candidates who are using the system to seek jobs.

Vietnamese Law (https://thuvienphapluat.vn/) is a library of legal documents in Vietnam. This program includes almost all documents in many majors. It is one of the largest law e-libraries in Vietnam. It can search for stored documents and their update over time.

Better Work is a special program that is held by the International Labor Organization (ILO) and the International Finance Corporation (IFC). It released an application of Vietnam Labor Law Guide (Vietnam LLG; https://play.google.com/store/apps/details?id=org.betterwork.llg&pcampaignid=web_share) to strengthen compliance with labor standards and enhance the capabilities of businesses in the global apparel industry (Fig. 11). This application can run on the Android and iOS operating systems. It covers the content of Vietnamese Labor Laws, provisions on occupational safety and health, and minimum wages. It also includes frequently asked questions (FAQs) and quiz questions.

Because Vietnamese Law and Vietnam LLG only support the search for legal documents, Table 7 compares them and our system when finding legal documents about the abilities to store documents and search for the laws.

Comparing resume and labor law search systems

Currently, many systems support the search for resumes and legal documents. For example, Gambaru (https://gambaru.io/en/search-matching), TopCV (https://www.topcv.vn/) and Indeed (https://vn.indeed.com/?r=us) are useful systems for job seeking in the IT domain, and Vietnamese Law (https://thuvienphapluat.vn/) and the National Database (http://vbpl.vn/botuphap/Pages/Home.aspx) about Law Documents support finding legal documents. However, those systems have not yet completely combined these functions to become convenient for users. Those functions are separate; they are not connected to work, and they also have several limitations to meet practical requirements.

Gambaru is a recruitment platform with many experienced IT sources. This system also supports talent searching for jobs that are suitable for their skills. It helps to place the high calibers for the goal of the job through AI technology. TopCV is a website that helps users create their CVs and seek suitable jobs with candidate majors. It also helps entrepreneurs search for candidates to recruit. Table 8 compares Gambaru, TopCV, and our system as a combined system of resume search and support for labor law search.