AI and Machine Learning Glossary for Recruiters

Recruiting vendors use the words “AI,” “machine learning,” and “NLP” interchangeably in sales decks. They are not interchangeable. Each term describes a distinct capability with distinct requirements, failure modes, and costs. Misreading them leads to buying the wrong tools, setting the wrong expectations, and building resume parsing automation pipelines on foundations that won’t hold.

This glossary defines every term a recruiter or HR leader is likely to encounter when evaluating or deploying AI-assisted hiring technology. Definitions are organized by concept cluster — not alphabetically — so the relationships between terms are clear from the first read.

Core AI Concepts

These three terms form the foundation. Every other term in this glossary builds on them.

Artificial Intelligence (AI)

Artificial intelligence is the broad discipline of building computer systems capable of performing tasks that would otherwise require human judgment — pattern recognition, decision-making, language comprehension, and adaptation to new inputs. In recruiting, “AI” is often used as a catch-all label. In practice, most recruiting AI is a specific combination of machine learning models and NLP pipelines, not general-purpose intelligence. McKinsey Global Institute research identifies AI-driven automation as one of the highest-impact levers for knowledge worker productivity, but that productivity gain only materializes when the right AI type is matched to the right problem.

Machine Learning (ML)

Machine learning is a subset of AI in which a system learns from data rather than following hard-coded rules. An ML model is trained on a labeled dataset — thousands of examples where the correct answer is already known — and then deployed to make predictions on new, unseen data. In recruiting, ML powers candidate ranking, flight risk prediction, and match-score generation. The critical constraint: ML models require substantial, high-quality historical data to train on. An ATS with 18 months of hiring history in one job family cannot produce reliable ML-based scores for a new role category. The model has nothing to learn from.

Natural Language Processing (NLP)

Natural language processing is the subset of AI focused specifically on enabling machines to read, interpret, and generate human language. NLP is the engine behind resume parsing. When a parser extracts “7 years of project management experience” from a free-form resume paragraph and writes it into a structured database field, that extraction is NLP at work. For a deeper technical walkthrough of how NLP applies specifically to candidate screening, see our guide on NLP in resume parsing.

How AI Systems Are Built

Understanding how AI tools are constructed explains why they fail — and how to evaluate vendor claims.

Algorithm

An algorithm is a defined sequence of instructions a computer executes to produce an output. In ML, the algorithm defines how the model learns from training data — which patterns it prioritizes, how it weights conflicting signals, and what it does with uncertainty. Different algorithms are appropriate for different tasks: a ranking algorithm optimizes ordering; a classification algorithm assigns categories (qualified / not qualified); a regression algorithm predicts a continuous value (e.g., estimated time-to-fill). Recruiters evaluating AI tools should ask which algorithm type underpins each claim the vendor makes.

Training Data

Training data is the labeled historical dataset an ML model learns from. Labels are the “correct answers” — for a resume parser, a labeled dataset might contain thousands of resumes with each relevant field already correctly extracted by humans. The model learns to replicate those extractions. Training data quality sets the ceiling on model performance. A parser trained exclusively on resumes from one industry or formatted in one style will systematically underperform on resumes outside that distribution. Before purchasing a parsing tool, ask: what industries and resume formats is this model trained on, and how recently was it retrained?

Model

A model is the trained artifact produced by running an ML algorithm on training data. Once trained, the model is deployed to make predictions on new inputs. Models are not static: they degrade as the world changes (this is called model drift), and they require periodic retraining on fresh data to maintain accuracy. In the context of resume parsing, a model trained in 2021 may not recognize skills and role titles that emerged in 2023 and 2024, causing those fields to be missed or misclassified.

Rule-Based System

A rule-based system uses explicitly programmed logic rather than learned patterns. A rule-based resume parser might be coded to look for a date range followed by a company name on the same line and extract that as an experience record. Rule-based systems are highly predictable and interpretable — you know exactly why an extraction succeeded or failed. They break when resumes deviate from the expected format. Many “AI” tools on the market are rule-based systems with an ML layer added only for edge cases. Neither approach is universally superior; the right choice depends on your resume volume, format diversity, and tolerance for unpredictable failure modes. See our breakdown of the three types of resume parsing technology for a full comparison.

Language and Text Processing Terms

These terms describe how AI systems interpret the actual text in resumes and job descriptions.

Parsing

Parsing is the process of analyzing a text document and extracting structured data from it. Resume parsing converts the unstructured text of a resume into defined data fields — name, contact information, work history entries, skills, education records — that can be stored in a database and used by downstream automation. Parsing accuracy is field-level: a parser can achieve 99% accuracy on email addresses and 87% accuracy on years-of-management-experience simultaneously. Evaluate accuracy by field, not as a single aggregate number.

Named Entity Recognition (NER)

Named entity recognition is an NLP technique that identifies and classifies specific entities in text — people’s names, company names, job titles, locations, dates, and skills. NER is one of the foundational techniques in resume parsing. When a parser correctly identifies “Google” as an employer and “Senior Product Manager” as a job title in the same sentence, NER is performing that classification. The accuracy of NER directly determines the accuracy of ATS field population.

Tokenization

Tokenization is the process of breaking text into individual units — typically words or subword fragments — so that an NLP model can process them. “Led a cross-functional team of 12 engineers” becomes individual tokens that the model processes in sequence to understand the full meaning. Tokenization is a preprocessing step invisible to end users but essential to every NLP-based parsing system.

Semantic Search

Semantic search retrieves results based on meaning rather than exact string matching. A keyword search for “Python” returns only resumes containing the word “Python.” A semantic search for “Python” also returns resumes mentioning Django, data pipeline development, or scripting — because the model understands those concepts are semantically related to Python proficiency. For resume database searching, semantic search dramatically reduces missed candidates who would qualify for a role but used different vocabulary to describe relevant experience. This is explored in depth in our guide to semantic search in hiring.

Keyword Extraction

Keyword extraction is the automated identification of the most significant terms in a document. In resume parsing, keyword extraction identifies skills, tools, certifications, and role titles. It is faster and simpler than full semantic analysis but produces lower-recall results — it finds what is explicitly stated, not what is implied by context. Many legacy ATS systems rely on keyword extraction; next-generation parsers combine it with NER and semantic understanding.

Sentiment Analysis

Sentiment analysis uses NLP to determine the emotional tone of text — positive, negative, or neutral. In recruiting, sentiment analysis is applied to candidate feedback surveys, exit interview transcripts, and occasionally cover letters. It is less central to resume parsing than to candidate experience analytics. Gartner research identifies sentiment analysis of candidate feedback as an emerging capability in enterprise HR analytics platforms.

Data and Infrastructure Terms

AI tools are only as good as the data infrastructure they sit on.

Structured Data

Structured data is information stored in a defined format with consistent field types — a candidate’s years of experience stored as an integer in a database column. Structured data is directly queryable, filterable, and usable by automation workflows. The entire purpose of resume parsing is to convert unstructured resume text into structured candidate records.

Unstructured Data

Unstructured data is free-form content with no predefined schema — a resume paragraph, a cover letter, a LinkedIn summary. The vast majority of candidate information arrives as unstructured data. According to research from the International Journal of Information Management, the majority of enterprise data is unstructured, and organizations that fail to extract structured signals from it operate with significant information disadvantages. In hiring, this means relying on recruiter memory and manual review rather than searchable, comparable data.

Data Pipeline

A data pipeline is the series of automated steps that moves data from source to destination — ingesting a resume file, extracting fields via NLP, validating outputs, and writing records to the ATS. The pipeline is the infrastructure layer. AI is a component within it. A broken pipeline produces corrupt data regardless of how sophisticated the AI model is. Building a reliable pipeline is the prerequisite for AI to add value — a principle that runs throughout our parent pillar on resume parsing automation.

API (Application Programming Interface)

An API is a defined interface that allows two software systems to communicate and exchange data. In recruiting automation, APIs connect your resume parser to your ATS, your ATS to your HRIS, and your HRIS to payroll. The robustness of an AI parsing tool’s API determines how cleanly it integrates into your existing tech stack. Poorly documented or unstable APIs are one of the most common reasons AI tool deployments fail after initial setup.

Confidence Score

A confidence score is a probability value assigned by an AI model to each extraction or prediction, indicating how certain the model is that the output is correct. A parser might extract a phone number with 99% confidence and a job title with 73% confidence from the same resume. High-performing parsing workflows define confidence thresholds: records with scores above the threshold route automatically; records below route to a human review queue. Parseur’s research on manual data entry documents error rates that compound quickly across high-volume records — confidence-gated routing is the mechanism that keeps those errors from propagating into your ATS.

Fairness and Compliance Terms

AI in hiring carries legal and ethical obligations that are not optional.

Algorithmic Bias

Algorithmic bias occurs when an ML model systematically produces outputs that favor or disfavor groups based on characteristics like gender, race, or age — not because the model was explicitly programmed to discriminate, but because it learned those patterns from historical training data that reflected past human bias. In hiring, a model trained on historical “successful hire” data from an organization that historically hired a non-diverse workforce will learn to replicate those patterns. The model doesn’t know it’s discriminating. It’s doing exactly what it was trained to do. For a deeper look at detection and mitigation, see our guide on bias and fairness in resume data extraction.

Bias Audit

A bias audit is a systematic evaluation of an AI model’s outputs across demographic groups to detect disparate impact. In resume parsing and candidate scoring, a bias audit compares pass-through rates, match scores, and ranking distributions across gender, age, and racial categories to identify statistically significant gaps that cannot be explained by qualification differences. SHRM guidance on fair hiring practices increasingly references bias audits as a component of responsible AI tool deployment.

Explainability

Explainability (also called interpretability) refers to the degree to which a human can understand why an AI model produced a specific output. A rule-based system is fully explainable — every decision traces back to a specific rule. A deep learning model is often a black box — it produces a confidence score with no auditable rationale. For hiring decisions, explainability matters: if a candidate challenges a rejection, you need to be able to articulate why the system ranked them as it did. Harvard Business Review coverage of AI in HR consistently emphasizes explainability as a requirement, not a nice-to-have.

Disparate Impact

Disparate impact is a legal doctrine holding that employment practices — including AI-assisted screening — that disproportionately exclude protected classes can constitute illegal discrimination even without discriminatory intent. A resume parser that systematically filters out candidates who attended HBCUs or that penalizes résumé gaps (which disproportionately affect women and caregivers) may create disparate impact liability. Awareness of this term is essential before deploying any automated screening tool at scale.

Predictive and Analytical Terms

These terms describe AI capabilities that go beyond parsing to inform strategic talent decisions.

Predictive Analytics

Predictive analytics uses historical data and statistical models to forecast future outcomes. In talent acquisition, predictive models estimate which candidates are most likely to accept an offer, succeed in a role, or leave within the first year. Microsoft Work Trend Index research consistently documents the growing organizational appetite for predictive HR capabilities — but the quality of predictions depends entirely on the quality and volume of historical data used to build the models. Explore the practical application in our guide to predictive analytics for talent acquisition.

Match Score

A match score is an AI-generated numerical rating indicating how well a candidate’s profile aligns with a job description. Match scores are outputs of ML models that weigh multiple factors — skills overlap, experience level, role title proximity, and sometimes behavioral signals. They are decision-support tools, not decisions. A 91% match score does not mean a candidate will succeed; it means the model found high alignment on the dimensions it was trained to weight. Recruiters who treat match scores as binary pass/fail gates rather than ranked signals miss candidates and compound model bias.

Candidate Ranking

Candidate ranking is the AI-generated ordering of a candidate pool by predicted suitability for a role. Ranking is more useful than binary pass/fail filtering because it preserves the full pipeline for human review while surfacing the highest-signal candidates first. It also makes bias auditing easier: demographic analysis of rank distributions reveals whether the model is systematically burying candidates from specific groups.

Talent Intelligence

Talent intelligence is the application of data analytics to inform workforce planning, sourcing strategy, and competitive hiring decisions. It incorporates both internal data (ATS records, HRIS tenure and performance data) and external signals (market compensation benchmarks, skill supply-and-demand data). Talent intelligence is an emerging category distinct from basic AI parsing — it operates at the strategic level above individual requisition decisions. Asana’s Anatomy of Work research documents the volume of low-value work that consumes recruiter time; talent intelligence platforms aim to reclaim that time by surfacing strategic signals automatically.

Related Terms Quick Reference

Shorter definitions for terms that appear frequently in vendor documentation.

• ATS (Applicant Tracking System): The database and workflow system that stores candidate records and tracks hiring pipeline progression. Resume parsing is the ingestion layer that populates ATS records.
• HRIS (Human Resources Information System): The broader HR data system that stores employee records post-hire. Clean ATS-to-HRIS data transfer prevents the kind of transcription errors that create downstream payroll and compliance problems.
• OCR (Optical Character Recognition): Technology that converts images or scanned PDFs into machine-readable text. OCR is a prerequisite for parsing scanned resume documents — without it, the NLP layer has no text to process.
• Vector Embedding: A mathematical representation of text as a point in high-dimensional space, used by semantic search and advanced NLP models to measure conceptual similarity between documents. Two job descriptions that use different words but describe the same role will have similar vector embeddings.
• Large Language Model (LLM): A category of ML model trained on massive text corpora capable of generating, summarizing, and reasoning about natural language. LLMs are increasingly being embedded in enterprise parsing and screening tools, but they introduce new risks around hallucination (generating plausible but incorrect output) that require mitigation in high-stakes hiring contexts.
• Hallucination: An AI output that is fluent and confident but factually incorrect. LLM-based tools can hallucinate candidate qualifications, invent dates, or generate plausible-sounding skill extractions that are not present in the source document. Confidence scoring and human review workflows are the primary defenses.
• Automation Workflow: A predefined sequence of actions triggered by a specific event — a resume received, a field extracted, a threshold exceeded. Automation workflows are the operational layer that connects parsed data to recruiting actions: routing to the right recruiter, triggering an acknowledgment email, or flagging a record for review.

Common Misconceptions

These misunderstandings appear consistently in recruiter conversations about AI tools.

“AI means the system learns automatically from my data”

Not by default. Most deployed AI tools do not automatically retrain on your organization’s data. The model you purchase was trained on the vendor’s dataset and is deployed as a static artifact. Continuous learning from your data, if it exists at all, requires explicit configuration and data-sharing agreements. Ask vendors specifically whether their model retrains on client data and, if so, whether your data is used to train models that serve other clients.

“Higher accuracy percentage means the tool is better”

Aggregate accuracy numbers conceal field-level variation. A parser claiming 95% accuracy may achieve that by performing flawlessly on simple fields (name, email, phone) while underperforming significantly on the fields you care most about — years of experience, management scope, or technical skills. Always request field-level accuracy breakdowns and benchmark on a sample of your own resume corpus before committing. See our guide on benchmarking resume parsing accuracy for the evaluation methodology.

“AI removes bias from hiring”

AI does not remove bias — it encodes bias at scale. An ML model trained on historical hiring decisions inherits every bias embedded in those decisions and applies them consistently across thousands of candidates simultaneously. The advantage of AI over human review is not the absence of bias but the auditability of bias. You can measure algorithmic bias statistically; you cannot measure individual recruiter bias the same way. That auditability is the starting point for correction — not a guarantee that correction will happen automatically.

“If the parser failed, AI doesn’t work”

Parsing failures are almost always infrastructure failures, not AI failures. Corrupt file formats, inconsistent schema mapping, missing ATS field definitions, and broken API connections cause the majority of parsing errors that get attributed to AI limitations. Diagnose at the pipeline level before concluding the AI model is the problem. Our guide on auditing resume parsing accuracy provides a structured diagnostic framework.

Putting It Together

The practical value of this glossary is not terminology for its own sake — it’s the ability to ask vendors the right questions, set the right internal expectations, and build automation infrastructure in the right sequence. AI adds value at the judgment points where deterministic rules break down. The data pipeline, field mapping, and routing logic must exist first.

For the full framework on sequencing your automation build, return to the resume parsing automation pillar. For metrics to track once your pipeline is live, see our guide on metrics for tracking parsing automation ROI. For the specific features that separate high-performing parsers from marketing-driven noise, see our breakdown of essential features of AI resume parsers.