Ensemble Scoring

The core principle is combining heterogeneous extraction methods to capture different signal types. An ensemble typically includes:

• Statistical models (TF-IDF, YAKE) for corpus-level frequency signals.
• Graph-based models (TextRank) for structural centrality.
• Embedding models (KeyBERT, EmbedRank) for semantic similarity. This diversity ensures that a phrase missed by one method due to its inductive bias is captured by another.

A robust, hyperparameter-free data fusion method that combines ranked lists without requiring relevance scores. The formula is: RRF_score(d) = Σ 1 / (k + rank_i(d)) Where k is a constant (typically 60) mitigating the impact of high rankings. RRF consistently outperforms Condorcet and CombSUM methods because it does not assume score distributions are comparable across different algorithms.

Before fusion, candidate phrases from different algorithms must be normalized to a common representation:

• Stemming/Lemmatization: Reducing words to their base form (e.g., 'running' → 'run').
• Case folding: Converting all text to lowercase.
• Stop word trimming: Removing leading/trailing stop words from n-grams. This prevents the same conceptual phrase ('Machine Learning' vs 'machine learning') from being treated as distinct entities in the fused list.

Advanced ensembles assign confidence weights to each extractor based on domain performance. For example:

• In scientific literature, TF-IDF may receive a higher weight for domain-specific terminology.
• In news articles, TextRank might be prioritized for its topic centrality. Weights are typically learned via a held-out validation set or set dynamically based on document length and genre classification.

Post-fusion, the combined list often contains semantically duplicate phrases (e.g., 'neural network' and 'deep neural network'). Maximal Marginal Relevance (MMR) is applied as a re-ranking step:

• It iteratively selects phrases that maximize relevance to the document while minimizing similarity to already selected phrases.
• This ensures the final keyphrase set provides diverse topical coverage rather than repeating the same concept.

The final fused list is pruned using a dynamic threshold to determine the optimal number of keyphrases. Common strategies include:

• Top-K selection: Taking the K highest-scoring phrases, where K is proportional to document length.
• Score drop-off: Identifying the largest gap between consecutive scores in the ranked list.
• Mean + Std Dev: Selecting all phrases scoring above one standard deviation from the mean fused score.

The process of assigning a numerical weight to each candidate phrase before fusion. Scores often combine phraseness (linguistic quality) and informativeness (topical relevance).

• Statistical Features: TF-IDF, YAKE's term frequency, and position of first occurrence.
• Semantic Features: Cosine similarity between phrase embedding and document embedding via EmbedRank.
• Graph Features: Centrality measures like TextRank PageRank scores.
• Ensemble methods normalize these heterogeneous scores before applying fusion logic.

A re-ranking algorithm often used as a post-processing step after ensemble fusion. It balances a phrase's relevance to the document against its redundancy with already selected keyphrases.

• Mechanism: Iteratively selects candidates that maximize λ * Sim1(D, Q) - (1-λ) * max Sim2(D_i, D_j).
• Purpose: Prevents the final keyphrase set from being dominated by near-duplicate n-grams.
• Integration: Applied after Reciprocal Rank Fusion to diversify the top-K output list.

The primary metric for evaluating ensemble scoring pipelines. It computes the harmonic mean of precision and recall for the top-K predicted keyphrases against a gold-standard set.

• Precision@K: |Predicted ∩ Relevant| / K.
• Recall@K: |Predicted ∩ Relevant| / |Relevant|.
• Benchmarking: Datasets like KP20k and Inspec use F1@5, F1@10, and F1@15 to compare ensemble methods against single-algorithm baselines.
• Mean Reciprocal Rank (MRR) is a complementary metric focusing on the rank of the first correct hit.

The initial filtering step that produces a set of potential n-grams for downstream scoring and fusion. The quality of this set directly limits ensemble performance.

• POS Tagging: Uses regular expressions like (JJ)*(NN)+ to extract noun phrases.
• Stopword Delimitation: RAKE splits text on stopwords and punctuation to form candidate sequences.
• N-gram Blocking: Generates all n-grams up to a maximum length (e.g., trigrams) and filters by frequency.
• Ensemble scoring cannot recover from a candidate set that misses the gold-standard keyphrase entirely.

A family of algorithms that represent text units as vertices in a co-occurrence graph and use structural centrality to determine salience. These often serve as inputs to an ensemble.

• TextRank: Builds a word graph with edges weighted by co-occurrence within a window; applies PageRank to score vertices.
• SingleRank: Extends TextRank by adding edge weights based on co-occurrence frequency.
• PositionRank: Incorporates word position information into the PageRank bias.
• Ensemble methods fuse these graph-based scores with embedding-based scores for robustness.