Abstractive Generation of Absent Keyphrases

49 concepts35 papers1 proposal

When we index a document, our natural instinct is to look for the most important words already written on the page. Traditional keyphrase extraction algorithms do exactly this: they score and select spans of text directly from the source document. However, this approach hits a hard theoretical limit. As established during the foundational SemEval-2010 shared task by Su Nam Kim et al. in SemEval-2010 Task 5 : Automatic Keyphrase Extraction from Scientific Articles, a significant portion of human-assigned keyphrases do not appear verbatim anywhere in the source text.

These are absent keyphrases—high-level conceptual labels, synonyms, or broader categories that summarize a document's essence. For example, a paper discussing "convolutional networks and backpropagation" might be naturally tagged with the absent keyphrase "Deep Learning". Because purely extractive models are structurally incapable of producing words outside the source text, they face an unyielding recall ceiling. To bridge this gap, researchers have shifted toward abstractive keyphrase generation.

The Shift to Sequence-to-Sequence Generation

How do you design a model that can "hallucinate" relevant terms not found in the text? The paradigm shift occurred when Rui Meng et al. reframed keyphrase identification as a sequence-to-sequence generation task in Deep Keyphrase Generation. Instead of treating the document as a sequence to be labeled, they deployed an encoder-decoder architecture equipped with a copying mechanism.

This dual-mode architecture provides an elegant solution: the model can copy explicit phrases directly from the text (present keyphrases) while retaining the ability to generate entirely new tokens from its vocabulary based on the latent semantic representation of the document (absent keyphrases). As noted by Binbin Xie et al. in From statistical methods to deep learning, automatic keyphrase prediction: A survey, this transition allowed models to move beyond statistical pattern matching and begin mapping documents to abstract semantic concepts.

The Challenge of Topical Drift

While sequence-to-sequence models can generate absent keyphrases, doing so reliably is incredibly difficult. Without the guardrails of the source text, unconstrained generation often suffers from what Anna V. Glazkova et al. describe in Key Algorithms for Keyphrase Generation: Instruction-Based LLMs for Russian Scientific Keyphrases as semantic insufficiency—producing phrases that are grammatically sound but fail to capture the core contribution of the text. Furthermore, Jing Zhao et al. point out in Incorporating Linguistic Constraints into Keyphrase Generation that models lack a structural template for valid phrases, leading to topically adjacent but ultimately irrelevant outputs.

To ground the generation process, researchers introduce structural anchors. For instance, Wang Chen et al. developed TG-Net in Title-Guided Encoding for Keyphrase Generation, which treats a document's title as a high-level query that guides the encoder to focus on the most salient themes in the body text. Similarly, Huanqin Wu et al. utilized overlapping keywords as anchors within a prompt-based learning framework in Fast and Constrained Absent Keyphrase Generation by Prompt-Based Learning, using a non-autoregressive "mask-predict" decoder to ensure the generated absent phrases remain tightly bound to the document's actual context.

Forcing Abstraction via Data Augmentation

A major reason models struggle to generate absent keyphrases is that standard training datasets are overwhelmingly dominated by present keyphrases. The model naturally takes the path of least resistance: it relies heavily on its copying mechanism, failing to develop deep semantic abstraction capabilities.

To force models to make the creative leap, Jishnu Ray Chowdhury et al. introduced KPDROP in KPDROP: Improving Absent Keyphrase Generation. This data augmentation technique randomly masks present keyphrases from the source text during training, transforming them into "artificial" absent keyphrases. Because the model can no longer copy these terms, it is forced to rely on the surrounding context to infer them, significantly boosting its ability to synthesize genuine absent keyphrases at test time.

In a similar vein of expanding context, Krishna Garg et al. demonstrated in Keyphrase Generation Beyond the Boundaries of Title and Abstract that incorporating extractive summaries of the full text—rather than just relying on the title and abstract—provides the necessary high-level thematic overview required to infer abstractive terms.

Borrowing Vocabulary from External Knowledge

If an absent keyphrase does not exist in the source text, where does the model find the vocabulary to describe it? While internal model weights store some semantic associations, low-resource or highly specialized domains often lack sufficient training data to map these connections, as observed by Di Wu et al. in Representation Learning for Resource-Constrained Keyphrase Generation.

One powerful solution is to look outside the document boundaries. In An Integrated Approach for Keyphrase Generation via Exploring the Power of Retrieval and Extraction, Hou Pong Chan designed a retrieval mechanism that pulls topically similar documents from a corpus, allowing the generator to use these external references as a semantic guide. Taking this further, Jiacheng Ye et al. constructed heterogeneous graphs connecting the source document to retrieved reference texts in Heterogeneous Graph Neural Networks for Keyphrase Generation, enabling the model to explicitly "borrow" relevant terms from related works to predict absent labels. For completely unsupervised settings, Xianjie Shen et al. leveraged a corpus-wide phrase bank combined with a dual-metric ranking system in Unsupervised Deep Keyphrase Generation to discover absent descriptors without requiring human-annotated targets.

Broader Applications and Modern Frontiers

The ability to generate absent keyphrases has profound implications for downstream tasks. In text summarization, as demonstrated by Fajri Koto in LipKey: A Large-Scale News Dataset for Absent Keyphrases Generation and Abstractive Summarization, absent keyphrases serve as critical semantic anchors, acting as a conceptual bridge that helps abstractive summarizers maintain thematic consistency.

As the field expands into multilingual and specialized domains—such as the legal corpus introduced by Olivier Salaün et al. in EUROPA: A Legal Multilingual Keyphrase Generation Dataset—the reliance on massive labeled datasets remains a bottleneck. To address this, recent work has turned to Large Language Models (LLMs). By investigating zero-shot capabilities, Jishnu Ray Chowdhury et al. found in Zero-Shot Keyphrase Generation: Investigating Specialized Instructions and Multi-sample Aggregation on Large Language Models that while specialized instructions yield marginal gains, aggregating multiple LLM outputs via a frequency-order strategy allows zero-shot models to effectively synthesize abstractive descriptors, opening new pathways for unsupervised conceptual indexing.

Key Foundational Work

SemEval-2010 Task 5 : Automatic Keyphrase Extraction from Scientific Articles (2010) — contributed 4 concept elements to this direction.
Automatic keyword extraction for learning object repositories (2008) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_absent_keyphrase_generation_22237800,
  title = {absent keyphrase generation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/absent-keyphrase-generation/rdr_22237800268444288b2a59ce06eb2c1c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_absent_keyphrase_generation_22237800_v1,
  title = {absent keyphrase generation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/absent-keyphrase-generation/rdr_22237800268444288b2a59ce06eb2c1c/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake