## Overview LinkedIn's skills extraction initiative represents a comprehensive production machine learning system designed to power their Skills Graph, a foundational technology that underpins their vision of a "skills-first economy." The company faces the challenge of extracting and normalizing skills mentioned across diverse content types—member profiles (particularly in Summary and Experience sections where skills aren't explicitly tagged), job postings (especially those sourced externally lacking structured skill lists), LinkedIn Learning course descriptions, resumes, and feed posts. The goal is to create a comprehensive, consistent, and accurate skill repository that enables better matching and relevance across jobs, courses, and recommendations. The business context is significant: with over 41,000 skills in their taxonomy and approximately 200 global profile edits per second, LinkedIn needs to process this content at scale while maintaining strict latency requirements. The system must handle both explicit skill mentions ("expected skills for this job includes programming in Java") and indirect references ("you are expected to know how to apply different techniques to extract information from data and communicate insights through meaningful visualizations"), making this a challenging natural language understanding problem in production. ## Technical Architecture and Model Stack LinkedIn built a multi-stage AI model workflow that addresses the nuanced challenges of skill extraction and mapping. The architecture is designed to handle the reality that skills appear differently across content types and that contextual positioning matters—a skill mentioned in the "qualifications" section of a job posting carries different weight than one appearing in a "company description" section. The pipeline begins with **skill segmentation**, where raw unstructured input is parsed into well-formed structures. For job postings, this means identifying sections like "company description," "responsibilities," "benefits," and "qualifications." For resumes, it involves delineating skills sections from experience descriptions. This structured understanding allows downstream models to better interpret the importance and relevance of extracted skills based on their location within the document. The **skill tagging** phase employs a hybrid approach that balances speed and semantic understanding. The first component is a trie-based tagger that encodes skill names from LinkedIn's skills taxonomy into a trie data structure and performs token-based lookup on raw text. This approach scales exceptionally well with large volumes of text and runs extremely fast, making it suitable for high-throughput scenarios. However, it has the limitation of being dependent on the skills taxonomy capturing every possible expression of a skill in real-world text. To complement this, LinkedIn developed a semantic tagging approach using a two-tower model architecture based on Multilingual BERT. This model generates contextual embeddings for both source text and skill names, with the two-tower structure designed to decouple the generation of sentence and skill embeddings while keeping them comparable through a similarity function. This semantic approach can infer that phrases like "experience with design of iOS application" map to "Mobile Development" even without exact text matches, addressing the limitation of the trie-based approach. Following skill tagging, a **skill expansion** phase leverages the Skills Graph itself to query for relevant skills within the same skill group or those sharing structural relationships such as parent skills, children skills, and sibling skills. This expansion increases the probability of capturing all relevant skills for a given piece of content. ## Multitask Learning and Domain-Specific Scoring The most sophisticated component is the **multitask cross-domain skill scoring** model, which identifies and scores each content piece and skill candidate pair. This architecture is split into shared and domain-specific modules, reflecting LinkedIn's understanding that while some aspects of skill extraction are universal, each vertical (job postings, member profiles, feeds, etc.) has unique characteristics. The shared module contains two key encoders. The Contextual Text Encoder, built on Transformer architecture, incorporates all available textual information for each content-skill pair—this might include the specific phrase mentioning the skill, surrounding sentences or paragraphs, job titles, or a member's most recent job. Transformers were chosen for their proven superiority on language understanding tasks and inherent capability to capture contextual information through their attention mechanisms. The Contextual Entity Encoder complements the text encoder by utilizing pre-calculated embeddings for skills, titles, industries, geographic locations, and other entities to provide entity-level context. Manual features such as co-occurrence rates between entities are also incorporated, blending learned representations with engineered features based on domain knowledge. The domain-specific module features multiple dedicated model towers, one for each vertical. While these towers are developed independently, they all share the same text and entity-based contextual information from the shared module. This design assumes that entities and text affect skill extraction similarly across domains, but allows each vertical to include their own specific information sources and maintain flexibility for nuanced differences in skill understanding. This architectural decision represents a pragmatic approach to building production systems—achieving reusability and shared learning where appropriate while preserving the ability to specialize. Beyond simple extraction, LinkedIn's system identifies multiple types of content-skill relationships through multitask learning. They define "required" relationships (skills explicitly mentioned as requirements), "core" relationships (skills essential to fulfill the job's basic functionality regardless of whether they're stated), and general "mention/valid" relationships. A skill importance score aggregates predictions from these multiple relationship types, providing richer signals than a binary present/absent classification. The multitask learning framework allows the system to learn these relationships simultaneously rather than training separate models for each, improving efficiency and enabling knowledge transfer across related tasks. ## Production Serving and Model Optimization Serving these models at LinkedIn's scale presents significant operational challenges. The system must handle nearline inference for profile updates (approximately 200 edits per second globally) with each message processed in under 100 milliseconds. Simultaneously, it needs to support offline batch processing for full data reprocessing and various online serving scenarios for search, recommendations, and other downstream systems. The original 12-layer BERT model, while powerful, is computationally demanding with a large parameter count that makes meeting these latency requirements on CPU-based serving infrastructure (Samza-BEAM) extremely challenging. Rather than compromise on model quality or infrastructure, LinkedIn employed **knowledge distillation** to compress the model. This technique transfers knowledge from a larger teacher network to a smaller student network, training the student to replicate the teacher's behavior. For online serving, knowledge distillation reduced the model size by 80% without compromising performance, meeting the existing CPU serving constraints. For full data reprocessing, the team collaborated with infrastructure teams to develop Spark offline scoring capabilities. They also devised a hybrid solution that uses offline resources for batch reprocessing and nearline processors for nearline traffic, optimizing cost-to-serve while maintaining service level agreements. This hybrid approach represents practical LLMOps thinking—recognizing that different use cases (nearline vs. batch) have different latency and throughput requirements and can be served by appropriately matched infrastructure. ## Feedback Loops and Continuous Improvement LinkedIn built multiple product-driven feedback loops directly into their applications to enable continuous model improvement, demonstrating mature MLOps practices. For **recruiter skill feedback**, when recruiters manually post jobs on LinkedIn, the AI model suggests a list of skills after they fill in the posting content. Recruiters can edit this list based on whether they believe a skill is important, providing high-quality human feedback on model predictions. For **job seeker skill feedback**, when job seekers view postings, they see how many skills overlap between their profile and the job, with higher overlap indicating higher application success probability. Seekers can review the top 10 skills used for matching calculations and provide feedback if certain skills seem irrelevant to the job. This captures skill-job relationships from the job seeker perspective, providing a different signal than recruiter feedback. **Member profile skill feedback** leverages LinkedIn Skill Assessments, which are adaptive assessments designed by LinkedIn Learning experts to validate skills across domains. Members who pass assessments with 70th percentile or higher scores receive "verified skill" badges visible to recruiters. This assessment data provides ground truth signals about members' actual skill proficiency, helping ensure that extracted skills are accurate and enabling further model improvements. These feedback mechanisms are integrated directly into the product experience rather than being separate evaluation tools, making data collection natural and continuous. This approach exemplifies production-oriented thinking where model improvement is baked into the product itself rather than being a separate offline process. ## Downstream Applications and Business Impact The skill extraction capabilities enable several critical applications across LinkedIn. **Career relevant skills** identification uses the extracted member-skill graph with heterogeneous edges to understand members more deeply. By collecting contextual skill data and job application data, the system identifies the most important and relevant skills for a member's career, enabling better job recommendations and candidate suggestions to recruiters. **Skill proficiency estimation** builds on extraction by inferring members' expertise levels in their listed skills through a multitask learning framework with uncertainty weighting that incorporates signals from multiple contexts. This enriches the Skills Graph with additional dimensions beyond presence/absence of skills. For **job important skills**, the system not only extracts skills but identifies which are most important to each role by capturing content-skill relationships from multiple perspectives ("required," "core," and "mention/valid" relationships) rather than relying solely on explicit mentions. The multitask model learning these relationships simultaneously achieved significant business impact, which LinkedIn quantified through A/B testing. Measured improvements include job recommendation gains of 0.14% in member job applicants and offsite apply clickers, and 0.46% in predicted confirmed hires. Job search saw 0.15% increase in job sessions, 0.76% increase in PPC revenue, and 0.23% increase in engagements. Job-member skills matching showed 0.87% increase in qualified applications, 0.40% increase in qualified application rate, 0.24% increase in predicted confirmed hires, and 0.48% increase in applicants and apply click counts. While these percentages may appear modest, at LinkedIn's scale they represent substantial business value. ## Forward-Looking Directions and LLM Integration LinkedIn indicates ongoing investment in skill understanding capabilities, specifically mentioning plans to leverage large language models. One direction involves using LLMs to generate rich descriptions for every skill in their Skills Graph, potentially improving the semantic understanding of skill relationships and similarities. They also plan to fine-tune LLMs to improve skill extraction model performance and generate high-quality proxy labels at scale, addressing the perennial challenge in ML of obtaining sufficient high-quality training data. Another strategic direction involves moving toward embedding-based skill representations rather than exact skill text or ID matching. This would enable more semantically relevant matches in downstream models, allowing the system to understand that "machine learning" and "ML model development" are closely related even if the exact terms differ. This embedding-first approach aligns with modern trends in semantic search and retrieval systems. ## Critical Assessment LinkedIn's case study demonstrates mature production ML practices including multi-stage pipelines, hybrid approaches balancing different techniques (token-based and semantic), model compression for deployment constraints, and integrated feedback loops. The quantified business impact through rigorous A/B testing lends credibility to their claims, though the specific percentage improvements should be contextualized within LinkedIn's massive scale. The architecture shows thoughtful engineering decisions, such as the shared/domain-specific multitask model structure that balances reusability with specialization, and the hybrid serving infrastructure that matches computational resources to latency requirements. The knowledge distillation achieving 80% model size reduction without performance loss is notable, though the case study doesn't detail what "without compromising performance" means quantitatively or what metrics were used to validate equivalence. The feedback loop integration is particularly strong, collecting signals naturally through product interactions rather than requiring separate annotation efforts. However, the case study focuses primarily on technical implementation and measured output metrics (clicks, applications, hires) rather than discussing data quality challenges, model monitoring, failure modes, or how they handle adversarial cases or changing skill distributions over time. The mention of future LLM integration suggests this 2023 system predates their heavier adoption of large language models, positioning this case study as representing a "pre-LLM" or early-LLM era approach that combines classical NLP techniques (trie-based matching), transformer models (BERT), and traditional ML engineering. The forward-looking section acknowledging LLM fine-tuning and embedding-first approaches indicates recognition that the field is evolving beyond the techniques described. Overall, this represents a sophisticated production system that demonstrates strong MLOps maturity while operating at significant scale, though readers should recognize it describes LinkedIn's specific context and resources rather than necessarily being replicable at all scales.