Automatic Keyphrase Extraction: From Statistical Methods to LLMs and End-to-End Speech

The explosive growth of textual and audiovisual information has turned automatic keyphrase extraction (KE) from an auxiliary task into a core component of information retrieval and knowledge organization. A recent survey systematically reviews the theoretical methods and practical solutions for KE across both text and speech modalities, attempting to integrate long-fragmented research threads into a unified analytical framework. The value of this article lies not in proposing new algorithms, but in outlining a clear roadmap of the field's evolution from statistical methods toward large language models (LLMs) and end-to-end speech processing.

The Basic Pipeline of Keyphrase Extraction and Linguistic Differences

The survey first breaks down the typical stages of KE: candidate generation, feature construction, ranking or classification, and final keyphrase annotation. This pipeline is relatively mature in the text modality but faces structural challenges when transferred to speech. Written-language keyphrases tend to be predominantly noun phrases with clear boundaries; spoken-language keyphrases may be interrupted by filler words, repetitions, and repairs, and prosodic information (stress, pauses) carries boundary cues that written punctuation cannot replace. The authors particularly emphasize terminological nuances—keyphrase, keyword, and term are often used interchangeably in the literature, yet the granularity and linguistic properties they refer to differ, and this ambiguity directly affects the comparability of evaluations.

The Technical Spectrum of Four Method Categories

The survey classifies automatic KE methods into four categories: statistical, hybrid, machine learning, and structured. Statistical methods rely on signals such as TF-IDF and co-occurrence frequency, require no training data, and remain strong baselines in low-resource settings. Machine learning methods introduce classification or sequence labeling models trained on annotated corpora, improving domain adaptability. Hybrid methods attempt to combine statistical candidate generation with learned ranking. Structured methods exploit structural information within documents, such as sections, citation networks, or syntactic trees. The value of this classification lies in revealing a trend: the boundaries between methods are blurring, and modern systems often invoke multiple types of signals simultaneously.

TF-IDF (term frequency–inverse document frequency) is the most fundamental weighting scheme among statistical methods: term frequency measures how many times a word appears in a single document, while inverse document frequency penalizes high-frequency words that appear commonly across the entire corpus; multiplying the two yields a representativeness score for that word with respect to the document. Its core assumption is that "words that recur in a small number of documents better represent the topic of that document." Co-occurrence frequency methods are based on another idea: words that frequently co-occur within the same sliding window or the same sentence are often semantically related, so centrality metrics on a co-occurrence graph (such as degree centrality or PageRank) can be used to filter candidate keyphrases. The advantages of such methods are that they are fully unsupervised, computationally cheap, and require no retraining for cross-domain transfer, making them hard-to-avoid strong baselines in domains and languages lacking annotated data; their limitations are that they cannot capture word order or deep semantics, and they handle synonymous expressions and multi-word phrases rather crudely.

The Generative Turn Brought by LLMs

The survey devotes considerable space to LLM-based keyphrase generation. Unlike traditional extractive methods, the generative paradigm does not require keyphrases to appear verbatim in the source text; instead, the model produces semantically equivalent but more standardized phrasings. This alleviates extractive methods' over-reliance on source wording but introduces new problems: generated results may deviate from the source semantics, and evaluation difficulty rises significantly—traditional matching-based metrics struggle to measure the semantic correctness of generated keyphrases. The authors note that evaluating generative KE requires more careful human judgment or semantic similarity measures, which constitutes an open gap in current methodology.

The fundamental difference between extractive and generative approaches lies in the output space: extractive methods frame keyphrase selection as a labeling or ranking problem over the source word sequence, so the output is necessarily a substring of the source, naturally guaranteeing that keyphrases "appear in the text" and offering strong interpretability; generative methods, by contrast, formulate the task as conditional text generation, with the model autoregressively producing keyphrases word by word, unconstrained by source vocabulary. The cost of this shift is the loss of faithfulness guarantees to the source—the model may generate semantically related phrases not expressed in the source, or even introduce factual drift. At the evaluation level, extractive methods can be compared directly using exact match, recall, and F1; generative methods typically rely on soft-matching metrics such as BERTScore and semantic embedding cosine similarity, or introduce LLMs as judges, but the consistency of these metrics with human judgment remains an active research question.

The Bottleneck of Corpus Annotation

Automatic annotation is another focus of the survey. The performance of KE systems depends heavily on annotation quality, but keyphrase annotation is inherently subjective: different annotators may make noticeably different keyphrase choices for the same document. The survey reviews methods and challenges of automatic corpus annotation, including weak supervision using existing metadata (such as author keywords and titles) and cross-lingual transfer annotation. Data scarcity is particularly acute in low-resource languages, which is also one of the future directions the authors repeatedly emphasize in their conclusion.

Two Routes for Speech Keyphrase Extraction

For spoken scenarios, the survey compares two routes: transcription-based text and end-to-end acoustic. The transcription route reuses text KE methods but is constrained by the propagation of speech recognition errors—misrecognition of proper nouns and domain terms directly harms keyphrase quality. The end-to-end acoustic route learns keyphrase representations directly from audio signals, theoretically able to exploit prosodic and pronunciation information, but is currently constrained by the scarcity of speech-keyphrase alignment data. The authors argue that the two are not substitutes; multimodal fusion is the pragmatic direction, and effective multimodal fusion is precisely the core unsolved challenge.

End-to-end speech processing refers to bypassing the cascaded "recognize first, process later" structure, with a single neural network mapping directly from raw audio waveforms or acoustic features to the target task output. In keyphrase extraction, this means the model learns directly from audio which speech segments correspond to keyphrases, without intermediate transcribed text. Its theoretical appeal is that in cascaded systems, speech recognition errors propagate irreversibly downstream, whereas end-to-end training allows acoustic representations to be optimized for the keyphrase task and preserves prosodic cues such as stress, intonation, and pauses. However, end-to-end approaches impose stricter requirements on annotated data—they need time-aligned annotations of audio and keyphrases, and such data is far scarcer than "audio + transcribed text," which is also the main bottleneck preventing this route from being deployed at scale.

Toward Integration and Context Awareness

The survey's conclusion points to an integrated, context-aware paradigm: KE for text and speech should not operate in isolation, and generative and extractive methods need to be compared within a unified evaluation framework. Future progress depends on solving three key challenges—data scarcity in low-resource languages, effective fusion of multimodal information, and careful evaluation of generative KE systems. For practitioners, this means that when selecting approaches, one should not view any single method category in isolation, but rather focus on whether the system can flexibly combine different signals across candidate generation, ranking, and evaluation.