This systematic review by Daria Guseva and Olga Mitrofanova provides a comprehensive roadmap of automatic keyphrase extraction (KE), tracing its evolution from simple statistical counts to modern generative Large Language Models (LLMs). It uniquely bridges the gap between text and speech modalities, highlighting how acoustic cues like pitch and pauses can enhance the identification of salient information in spoken discourse.
TL;DR
The paper synthesizes decades of research into a unified framework for automatic keyphrase extraction, moving from traditional statistical methods to generative LLM-based approaches. Its core contribution is the detailed analysis of "prosodic specificity"—how acoustic features in speech can be used alongside text to identify key information more accurately.
What Problem Does This Paper Address?
As the volume of textual and audiovisual data grows exponentially, manual indexing and summarization have become impossible. While automatic keyphrase extraction (KE) is a mature field for written text, it remains a "complex, underexplored challenge" for spoken language.
Spoken language introduces "messy" variables absent in written text: disfluencies (ums and ahs), varying audio quality, and errors from Automatic Speech Recognition (ASR) systems. The authors address the need for a unified approach that handles both modalities and navigates the shift from simply extracting words found in a text to generating relevant phrases that might not be explicitly present.
Methods
The authors categorize KE methods into four main buckets, while introducing a fifth "new wave" paradigm:
- Statistical: Methods like TF-IDF or Chi-squared that rely on word frequency. These are fast but often ignore context.
- Hybrid (Linguostatistical): Algorithms like RAKE (Rapid Automatic Keyword Extraction) and YAKE!, which combine frequency with positional cues and part-of-speech tagging.
- Machine Learning-Based: Supervised classifiers (like KEA) and modern transformer-based models (like KeyBERT) that use semantic embeddings to find phrases closest to the document's meaning.
- Structural (Graph-Based): Methods like TextRank that treat a document as a network of words, where "important" words are those with the most connections.
- Generative (The LLM Shift): Using models like GPT-4 to predict keyphrases. This allows for "absent" keyphrases—terms that perfectly describe a topic but never actually appear in the source text.
For speech, the paper compares transcript-based methods (running text KE on ASR output) with acoustic-based methods that analyze raw audio features like pitch, energy, and duration.
Datasets
The review cites several foundational and modern benchmarks:
- English: Inspec (scientific abstracts), DUC2001 (news), and KP20k (a massive corpus of 20,000 computer science articles).
- Web-Scale: OpenKP, containing 100,000 web documents with expert annotations.
- Russian-Specific: The authors highlight the NeuroRNC group’s work on Russian news and research papers, noting that morphologically rich languages like Russian present unique challenges for standard statistical tools.
Key Insights and Ideas
- Prosodic Specificity: One of the most fascinating "aha moments" is the role of acoustics. Research shows that in spoken discourse, keywords are 40% more likely to occur at the end of a syntagma (a rhythmic-semantic unit) and are often preceded by specific pause patterns.
- The Domain Gap: As shown in Figure 2, KE research is heavily lopsided. The Biomedical and Clinical Sciences dominate the field, likely due to the high demand for indexing massive databases like PubMed.
- Terminology Overload: The paper clarifies the "terminological landscape" (see Figure 5). In Russian literature, researchers use nuanced terms like "semantic milestones" or "lexical dominants," whereas English literature tends to stick to "keywords" or "key items."
- From Extraction to Creation: The shift to LLMs represents a paradigm shift. We are moving from "extractive" systems (finding the best words in the text) to "abstractive" systems (understanding the concept and naming it).
Key Results
While this is a review paper rather than a primary study, it highlights several significant findings from the literature:
- Performance Gains: Modern ML models like Random Forest have been shown to improve F-measures from 22% (traditional KEA) to 30% on the same datasets.
- Human Perception: Experiments confirmed that providing listeners with a set of keywords before they hear a noisy audio clip significantly improves their ability to reconstruct and understand the speech.
- LLM Superiority: In tasks involving both present and absent keyphrases, generative models like CopyRNN consistently outperform traditional baseline methods like TextRank or TF-IDF.
Limitations and Future Work
The authors identify three primary "bottlenecks" for the next generation of KE:
- Hallucinations: Generative LLMs are prone to creating keyphrases that are topically relevant but factually incorrect or not grounded in the source text.
- Low-Resource Languages: There is a massive scarcity of annotated data for languages other than English, making it difficult to train robust models for global applications.
- Multimodal Fusion: The "holy grail" of the field is a truly integrated system that fuses textual features from ASR with raw acoustic-prosodic features (pitch, energy, duration) in a single end-to-end model.
Summary updated