Glossary

Candidate Scoring

Candidate scoring is the process of assigning a numerical weight to each candidate keyphrase based on features like frequency, position, and semantic similarity to rank their relevance.

Get in touch Learn more

Developer reviewing semantic search engine results on laptop, relevance scores visible, technical search demo.

KEYPHRASE EXTRACTION

What is Candidate Scoring?

Candidate scoring is the algorithmic process of assigning a numerical weight to each potential keyphrase to quantify its relevance and salience relative to a source document.

Candidate scoring is the computational step in keyphrase extraction where a numerical relevance weight is assigned to every phrase in the candidate pool. This score is derived from a combination of statistical features—such as term frequency, inverse document frequency, and positional bias—and semantic features like embedding similarity to the document centroid. The goal is to transform a raw list of n-grams into a ranked list where the highest-scoring items represent the document's core topics.

Scoring functions range from simple unsupervised heuristics like TF-IDF averaging to complex neural models that compute the cosine similarity between a candidate's BERT embedding and the document embedding. Advanced implementations incorporate phraseness and informativeness as distinct scoring dimensions, then fuse them via weighted linear combination or Maximal Marginal Relevance (MMR) to penalize redundancy among top-ranked candidates.

CANDIDATE SCORING

Core Scoring Features

The quantitative engine that assigns a relevance weight to each candidate keyphrase, transforming raw n-grams into a ranked list based on statistical, positional, and semantic signals.

Term Frequency-Inverse Document Frequency (TF-IDF)

A foundational statistical measure that evaluates how important a word is to a document in a collection. Term Frequency (TF) counts how often a term appears, while Inverse Document Frequency (IDF) penalizes terms that appear across many documents.

Mechanism: TF-IDF = TF * log(N/DF)
Strength: Excellent at filtering out common stop words and corpus-specific noise.
Limitation: A pure bag-of-words approach that ignores word order and semantics.

Positional Weighting

A heuristic scoring boost based on the location of a candidate phrase within a document's structure. The assumption is that authors place critical terms in prominent positions.

Title & Abstract: Phrases appearing here receive the highest multipliers.
First Paragraph: Often contains the core thesis statement.
Section Headers: Terms in H1/H2 tags are weighted more heavily than body text.

Phraseness vs. Informativeness

A dual-component scoring model that separates linguistic quality from topical relevance.

Phraseness: Measures how grammatically well-formed a candidate is as a multi-word unit, independent of the document's topic. It penalizes fragments like "the quick" while boosting "quick brown fox".
Informativeness: Quantifies how well the phrase captures domain-specific content. It rewards terms with high TF-IDF or strong alignment with the document's central theme.

Semantic Similarity Scoring

A neural approach that ranks candidates by computing the cosine similarity between their dense vector embeddings and the document's overall embedding.

KeyBERT: Embeds both the document and candidate phrases using a transformer model, then selects the top-K most similar phrases.
EmbedRank: Extends this by using sentence embeddings and applying Maximal Marginal Relevance (MMR) for diversity.
Advantage: Captures synonyms and paraphrases that statistical methods miss.

Graph-Based Centrality

Algorithms that represent text as a network where nodes are candidate words or phrases and edges represent co-occurrence within a sliding window. Centrality scores determine salience.

TextRank: Builds a co-occurrence graph and runs the PageRank algorithm. Words with high eigenvector centrality are selected as keyphrases.
SingleRank: A variant that adds weighted edges based on co-occurrence distance.
Benefit: Unsupervised and domain-independent, requiring no training data.

Ensemble & Fusion Scoring

Combines the ranked outputs of multiple extraction algorithms to produce a more robust final ordering. This mitigates the individual biases of any single method.

Reciprocal Rank Fusion (RRF): A simple, effective method that sums the reciprocal of each candidate's rank across multiple lists: score = Σ 1/(k + rank).
Weighted Voting: Assigns confidence weights to each base algorithm based on historical performance on a validation set.
Use Case: Ideal when combining a fast statistical method like YAKE with a semantic method like KeyBERT.

CANDIDATE SCORING INSIGHTS

Frequently Asked Questions

Explore the core mechanisms behind how keyphrase extraction systems assign numerical weights to candidate phrases, distinguishing signal from noise in unstructured text.

Candidate Scoring is the algorithmic process of assigning a numerical weight to each potential keyphrase (candidate) to quantify its relevance to the source document. Unlike simple frequency counting, modern scoring functions synthesize multiple features—including term frequency, inverse document frequency, phraseness, and informativeness—to rank candidates. The goal is to elevate phrases that are both linguistically well-formed and topically central while suppressing generic or noisy n-grams. This step directly follows phrase candidate generation and precedes final selection or re-ranking.

Enabling Efficiency, Speed & Accuracy

Intelligent Analysis, Decision & Execution

We build AI systems for teams that need search across company data, workflow automation across tools, or AI features inside products and internal software.

Talk to Us

Search across company data

Give teams answers from docs, tickets, runbooks, and product data with sources and permissions.

Useful when people spend too long searching or get different answers from different systems.

Enterprise searchRAGPermissions

Automate internal workflows

Use AI to route work, draft outputs, trigger actions, and keep approvals and logs in place.

Useful when repetitive work moves across multiple tools and teams.

AI agentsWorkflow automationGovernance

Add AI to products and internal tools

Build assistants, guided actions, or decision support into the software your team or customers already use.

Useful when AI needs to be part of the product, not a separate tool.

AI integrationDecision supportModel routing

CORE CONCEPTS

Related Terms

Understanding candidate scoring requires familiarity with the statistical and semantic features that define phrase importance. These related terms form the backbone of modern keyphrase extraction pipelines.

TF-IDF

A foundational statistical measure evaluating a term's importance by balancing its frequency in a document against its prevalence across a corpus. Term Frequency (TF) counts how often a word appears, while Inverse Document Frequency (IDF) penalizes common words. In candidate scoring, TF-IDF serves as a baseline feature for informativeness, often combined with positional and semantic signals.

1972

Originated

Phraseness

A scoring component that measures how linguistically well-formed a candidate sequence is, independent of its topical relevance. It evaluates whether an n-gram constitutes a valid phrase rather than an arbitrary word sequence. Features include POS tagging patterns, stopword boundaries, and syntactic chunking. High phraseness scores filter out grammatically incoherent candidates before semantic scoring.

Informativeness

A scoring component measuring how well a candidate captures the core topical content of a document. It quantifies domain specificity and semantic relevance. Key signals include:

TF-IDF or TF-ICF weights
Semantic similarity to the document embedding
Entity salience for named entities Informativeness distinguishes topical terms from generic language.

Semantic Similarity

A metric quantifying the conceptual relatedness between two text spans using distributional semantics or knowledge graph path lengths. In candidate scoring, the cosine similarity between a phrase's embedding and the document's centroid embedding is a powerful feature. Models like Sentence-BERT and KeyBERT rely on this to rank phrases by their alignment with the document's overall meaning.

Maximal Marginal Relevance (MMR)

A re-ranking algorithm that balances a phrase's relevance against its redundancy with already selected keyphrases. The formula combines a relevance score with a penalty for similarity to higher-ranked outputs. This prevents the final keyphrase set from being dominated by near-duplicate terms, ensuring diverse and comprehensive coverage of the document's topics.

Ensemble Scoring

A technique that combines the ranked outputs of multiple extraction algorithms using fusion methods to improve robustness. Common approaches include Reciprocal Rank Fusion (RRF) and weighted score averaging. By aggregating statistical methods like TF-IDF with embedding-based methods like KeyBERT, ensemble scoring mitigates the weaknesses of any single approach.

About the author

Prasad Kumkar

CEO & MD, Inference Systems

Prasad Kumkar is the CEO & MD of Inference Systems and writes about AI systems architecture, LLM infrastructure, model serving, evaluation, and production deployment. Over 5+ years, he has worked across computer vision models, L5 autonomous vehicle systems, and LLM research, with a focus on taking complex AI ideas into real-world engineering systems.

His work and writing cover AI systems, large language models, AI agents, multimodal systems, autonomous systems, inference optimization, RAG, evaluation, and production AI engineering.

Limited slotsGet a Free AI Consultation

How We Work

Custom AI workflows for your Business

One-fit-all AI don't work for modern businesses. At Inferensys, we aim to understand your business & custom requirements; which we use to define most efficient agentic workflows, the data, and the tools for your business.