Topic-Keyword Semantic Guidance (TKSG)

Updated 4 July 2026

TKSG is a design principle that decomposes semantic control into a global layer (topic vectors) providing persistent context and a local layer (keyword cues) refining fine-grained decisions.
It integrates topic-level signals to bias overall decoding while using keyword-based attention to enhance local evidence in applications like radiology report generation and multimodal retrieval.
Empirical studies demonstrate that combining global and local semantic guidance improves performance in topic modeling, dialogue systems, and image-text retrieval tasks.

Topic-Keyword Semantic Guidance (TKSG) denotes a family of methods in which topic-level signals provide global semantic control and keyword-level signals provide finer local control over representation learning, retrieval, generation, or fusion. In the most explicit formulation, topic words are treated as disease classifications and keywords as common symptom or finding concepts, so that a topic vector guides the entire decoding process while keyword embeddings refine local decoding decisions (Xiao et al., 13 Sep 2025). Closely related work extends the same design principle to seed-guided topic discovery, semantic graph-guided topic models, remote-sensing retrieval, target-guided dialogue, and multimodal fusion, even when the papers use adjacent rather than identical terminology (Harandizadeh et al., 2021, Zhang et al., 2022, Li et al., 26 Jan 2026, Li et al., 12 May 2026).

1. Conceptual structure

A recurring TKSG pattern is a two-level semantic decomposition. The global layer supplies persistent context, typically as a topic distribution, topic vector, scenario bias, or sentence-level semantic embedding. The local layer supplies more selective cues, typically as keywords, key object words, masks, region proposals, or token-level semantic units. In automated radiology report generation, the global topic probabilities $P_T$ are mapped into a dense topic vector $l$ , which is added to every decoder input embedding, while the top predicted keywords are embedded and injected through semantic-guided attention over $[\mathbf{X};\mathbf{E}]$ (Xiao et al., 13 Sep 2025).

An adjacent but structurally similar decomposition appears in proactive dialogue. There, conversational scenario modeling uses user profile and domain knowledge to produce a scenario bias over the output vocabulary, while intent-keyword bridging predicts future-turn keyword-type and keyword-topic pairs and injects their pooled representation into decoder cross-attention (Li et al., 12 May 2026). The distinction is important: topic-like control governs discourse-level direction, while keyword-like control governs the next semantic waypoint.

A related two-level design also appears in semantic-guided image fusion. TeSG derives mask semantics and text semantics from automatically generated textual descriptions, then uses the mask-guided cross-attention module for region-precise fusion and the text-driven attentional fusion module for later global semantic conditioning (Zhu et al., 20 Jun 2025). This suggests that TKSG is not restricted to lexical supervision in text-only models; it more generally describes a separation between broad semantic intent and fine-grained semantic evidence.

2. Topic modeling and seed-guided discovery

The topic-modeling lineage of TKSG ranges from automatic keyword-centered topic assignment to explicit seed-guided semantic priors. Early work on "Topic Modeling based on Keywords and Context" treats certain words as characteristic keywords of a topic and lets nearby words inherit topic assignments from those keywords within a local window, thereby reducing unnatural topic switching within documents (Schneider, 2017). "Keyword-based Topic Modeling and Keyword Selection" instead models keyword subsets as upstream control variables: keywords are selected first, then document-topic mixtures and topic-word matrices are conditioned on the chosen subset, making keyword choice a mechanism for forecasting future latent topics (Wang et al., 2020).

"Keyword Assisted Embedded Topic Model" provides one of the clearest direct realizations of TKSG inside a neural topic model. It defines user seed sets $\omega_k$ , forms a semantic centroid $\omega_{ks}$ from fixed word embeddings, builds a binary prior matrix $\gamma^{prior}$ by exact seed membership or cosine-similarity thresholding, and regularizes both the topic-word pathway and the document inference pathway through $L_\alpha$ and $L_\mu$ (Harandizadeh et al., 2021). "Effective Seed-Guided Topic Discovery by Integrating Multiple Types of Contexts" extends the idea by combining three context sources—local-context word embeddings, pre-trained LLM representations, and topic-indicative sentences retrieved from current seed expansions—inside an iterative ensemble ranking framework (Zhang et al., 2022).

A stricter boundary case is "TopicNet: Semantic Graph-Guided Topic Discovery" (Duan et al., 2021). TopicNet injects a TopicTree derived from WordNet or user customization and constrains Gaussian topic embeddings through asymmetric thresholded KL divergence. This is semantically guided topic discovery, but the guidance is primarily topic-topic and ontology-structured rather than topic-keyword in the seed-word sense.

Paper	Guidance unit	Distinctive mechanism
"Topic Modeling based on Keywords and Context" (Schneider, 2017)	Automatically learned keywords	Nearby words inherit topic scores from characteristic keywords
"Keyword-based Topic Modeling and Keyword Selection" (Wang et al., 2020)	Keyword subsets	Keywords condition future topic mixtures and topic-word matrices
"Keyword Assisted Embedded Topic Model" (Harandizadeh et al., 2021)	User seed words	Embedding-expanded prior regularizes topic-word and inference pathways
"Effective Seed-Guided Topic Discovery by Integrating Multiple Types of Contexts" (Zhang et al., 2022)	Seeds plus retrieved topic sentences	Iterative ensemble of embedding, PLM, and sentence-level evidence
"TopicNet: Semantic Graph-Guided Topic Discovery" (Duan et al., 2021)	TopicTree concepts	Graph-structured topic-topic semantic guidance via Gaussian embeddings

The main conceptual distinction is therefore between methods that directly ground topics in keywords and methods that constrain topics through semantic graphs or latent geometry. Both are semantically guided, but only the former are strict TKSG formulations.

3. Generation and dialogue systems

Topic guidance entered neural generation early in video captioning. "Generating Video Descriptions with Topic Guidance" mines soft topic distributions with LDA, predicts those topics from multimodal video features, and conditions caption generation either implicitly through input concatenation or explicitly through topic-modulated decoder weights in the Topic-Guided Model, where the effective decoder parameters are a soft mixture over topic-specific components (Chen et al., 2017). This is topic-semantic guidance rather than keyword guidance, but it established a persistent design principle: semantic control is stronger when it modulates the decoder rather than merely decorating the input.

A more explicit TKSG instantiation appears in automated radiology report generation. The TKSG framework retrieves top $N_R=30$ similar historical reports using BiomedCLIP, predicts 14 disease topics from pooled image features, predicts keyword probabilities from pooled image and retrieved-report features, maps the topic probabilities into a dense topic vector $l$ , and then injects $l$ 0 at every decoding step while using keyword embeddings inside semantic-guided attention (Xiao et al., 13 Sep 2025). The model is trained jointly with report-generation, topic-detection, and keyword-detection losses, $l$ 1, making global and local semantic guidance part of the same decoding pipeline.

Dialogue systems show a parallel evolution. "Dynamic Knowledge Routing Network for Target-Guided Open-Domain Conversation" uses coarse-grained keywords as intermediate semantic waypoints, predicts the next keyword from dialogue context, filters candidates through a directed keyword relation graph, and then performs keyword-augmented response retrieval with an additional discourse-level target-guided strategy (Qin et al., 2020). "Enhancing Target-Guided Proactive Dialogue Systems via Conversational Scenario Modeling and Intent-Keyword Bridging" adds a richer decomposition: a scenario bias derived from user profile and domain knowledge for discourse-level control, plus multi-turn intent-keyword prediction that separates keyword-type from keyword-topic and injects the resulting bridge representation into generation (Li et al., 12 May 2026).

A latent, non-keyword variant appears in "Unsupervised Abstractive Opinion Summarization by Generating Sentences with Tree-Structured Topic Guidance" (Isonuma et al., 2021). RecurSum replaces a unimodal Gaussian prior with a recursive Gaussian mixture whose components correspond to nodes in a topic tree, then decodes topic means into root, internal, and leaf topic sentences. This is topic-semantic guidance rather than keyword guidance, but it demonstrates that hierarchical semantic control can organize generation from generic to specific content.

4. Retrieval, fusion, and semantic expansion beyond text generation

TKSG-style mechanisms also appear in retrieval and multimodal fusion, where guidance shapes what evidence is even considered before final matching. "Multi-Perspective Subimage CLIP with Keyword Guidance for Remote Sensing Image-Text Retrieval" shifts remote-sensing retrieval from global CLIP matching to keyword-guided fine-grained alignment: DeepSeek V3.2 extracts core semantic keywords, SamGeo uses them to generate semantically relevant sub-perspectives, a Gated Global Attention adapter adapts the frozen CLIP backbone, and a Multi-Perspective Representation module turns local cues into multiple embeddings whose maximum-response similarity with the text drives training (Li et al., 26 Jan 2026). The critical matching operation is $l$ 2, which suppresses noisy perspectives and preserves the strongest local semantic correspondence.

In software topic recommendation, semantic guidance is supplied by a curated topic graph rather than by a decoder. "Semantically-enhanced Topic Recommendation System for Software Projects" builds SED-KGraph with 863 topics, 2,234 verified relationships, and 13 relation types, then uses it in two modes: KGRec for missing-topic augmentation from an existing topic set, and KGRec+ for topic recommendation from scratch by combining TF-IDF text classification with graph-based semantic expansion (Izadi et al., 2022). Here the graph acts as a label-space prior: topic inference is not flat multi-label classification but graph-constrained semantic completion.

A related pattern appears in semantic summarization and matching. "Semantic Similarity Measure of Natural Language Text through Machine Learning and a Keyword-Aware Cross-Encoder-Ranking Summarizer" uses human-authored keywords to rank sentences from long GIS&T topic descriptions and build a semantic summary that improves downstream topic-linking performance (Tian et al., 2023). "Multi-view Semantic Matching of Question retrieval using Fine-grained Semantic Representations" learns keyword importance from question-pair supervision, constructs multiple keyword subsets by iteratively removing the least important term, and matches questions through multiple salience-guided views (Chong et al., 2022). Both systems are keyword-centric without explicit topic variables, but both embody the TKSG intuition that semantic guidance should restructure representation and evidence selection rather than appear only at the final scoring layer.

TeSG occupies an intermediate position. It is semantic-guided and text-guided rather than formally topic-keyword guided: a BLIP-generated caption is modified by removing a key object word $l$ 3, diffusion-noise differences are normalized and binarized into mask semantics, and the resulting masks guide cross-modal fusion before a sentence-level text embedding refines the fused representation (Zhu et al., 20 Jun 2025). The paper itself notes that this is not a full topic-keyword representation, but it is directly relevant because it operationalizes the passage from one key word to localized spatial priors.

5. Empirical patterns

The empirical literature reports improvements when semantic guidance simultaneously constrains global semantics and local discrimination. In topic modeling, KeyETM improves topic quality and human topic intrusion performance by augmenting user seeds with embedding-based semantic neighborhoods and by regularizing both topic-word structure and document inference (Harandizadeh et al., 2021). In topic recommendation, adding SED-KGraph to text classification yields substantial gains over non-semantic baselines (Izadi et al., 2022). In multimodal retrieval and dialogue, maximum-response local matching or keyword-transition planning improves both semantic precision and end-task success (Li et al., 26 Jan 2026, Qin et al., 2020).

Paper	Setting	Reported result
"Keyword Assisted Embedded Topic Model" (Harandizadeh et al., 2021)	AYLIEN	Topic quality $l$ 4, intrusion $l$ 5
"Automated Radiology Report Generation Based on Topic-Keyword Semantic Guidance" (Xiao et al., 13 Sep 2025)	IU X-Ray / MIMIC-CXR	BLEU-4 $l$ 6; F1 $l$ 7
"Dynamic Knowledge Routing Network For Target-Guided Open-Domain Conversation" (Qin et al., 2020)	TGPC / CWC self-play	Success $l$ 8 / $l$ 9
"Semantically-enhanced Topic Recommendation System for Software Projects" (Izadi et al., 2022)	KGRec+	ASR@5 $[\mathbf{X};\mathbf{E}]$ 0, MAP@5 $[\mathbf{X};\mathbf{E}]$ 1
"Multi-Perspective Subimage CLIP with Keyword Guidance for Remote Sensing Image-Text Retrieval" (Li et al., 26 Jan 2026)	RSICD / RSITMD	mR $[\mathbf{X};\mathbf{E}]$ 2 / $[\mathbf{X};\mathbf{E}]$ 3
"TeSG: Textual Semantic Guidance for Infrared and Visible Image Fusion" (Zhu et al., 20 Jun 2025)	Detection / segmentation on fused MSRS	[email protected] $[\mathbf{X};\mathbf{E}]$ 4, mIoU $[\mathbf{X};\mathbf{E}]$ 5

Across these results, semantically guided methods are strongest in settings where global representations are underdetermined by the task. Underrepresented topics in imbalanced corpora, hard negatives in retrieval, dense overhead imagery, multi-turn target steering, and multimodal fusion of foreground versus background cues all benefit when semantics are decomposed into more structured control variables. This suggests that TKSG is particularly useful when the target semantics are sparse, asymmetric, or easily diluted by dominant background patterns.

6. Boundaries, limitations, and misconceptions

A common misconception is that any semantic-guided method is automatically a TKSG method. The literature draws finer distinctions. TopicNet is guided by a semantic ontology, but its main regularization is over topic-topic entailment across layers rather than direct keyword-topic anchoring (Duan et al., 2021). RecurSum is topic-guided through recursive Gaussian components, but it lacks explicit keyword units (Isonuma et al., 2021). The question-retrieval model based on keyword importance is strongly keyword-centric, yet its topic structure remains implicit and salience-based rather than explicitly modeled (Chong et al., 2022). TeSG uses one key object word and a sentence embedding, but it does not build a full topic-keyword hierarchy (Zhu et al., 20 Jun 2025).

A second limitation concerns dependence on external semantic sources. KeyETM depends on seed quality, embedding quality, and the balance between $[\mathbf{X};\mathbf{E}]$ 6 and $[\mathbf{X};\mathbf{E}]$ 7 (Harandizadeh et al., 2021). TopicNet depends on the quality of the constructed TopicTree and can suffer when the semantic prior does not align with downstream tasks (Duan et al., 2021). KGRec depends on the correctness and coverage of SED-KGraph, while relation types are curated but not yet exploited during propagation (Izadi et al., 2022). Radiology TKSG depends on CheXbert pseudo-labels, retrieved historical reports, and a fixed keyword vocabulary of frequent content words (Xiao et al., 13 Sep 2025). MPS-CLIP depends on LLM keyword extraction and a SamGeo interface that is conceptually clear but operationally under-specified in the reported method (Li et al., 26 Jan 2026).

A third limitation is reproducibility. Several papers omit exact low-level interfaces between semantic units and downstream modules: TeSG does not fully specify how infrared masks are computed from text-derived differences (Zhu et al., 20 Jun 2025); MPS-CLIP does not specify the exact LLM prompt, keyword count, or SamGeo prompting scheme (Li et al., 26 Jan 2026); DKRN omits exact loss formulas despite clearly describing its keyword-relation graph and discourse-level strategy (Qin et al., 2020). There is also a nomenclature problem: “semantic guidance” is broader than TKSG, and at least one nominally related paper, "Semantic Guidance Tuning for Text-To-Image Diffusion Models" (Kang et al., 2023), cannot be assessed as TKSG-related from the accessible content because the paper text is unavailable.

7. Synthesis and research directions

A plausible synthesis is that TKSG is converging toward a multi-level control stack. At the top level, systems use topic distributions, scenario bias, graph priors, or historical-case retrieval to define broad semantic direction. At the middle level, they introduce structured semantic carriers—seed words, intent keywords, key object words, graph neighbors, topic-indicative sentences, or sub-perspectives—that bridge abstract intent and local evidence. At the bottom level, they alter a concrete computation: topic-word probabilities, decoder embeddings, cross-attention keys and values, region selection, response retrieval, or ranking scores (Xiao et al., 13 Sep 2025, Li et al., 12 May 2026, Li et al., 26 Jan 2026, Izadi et al., 2022).

This suggests that future TKSG systems are likely to combine elements that have so far been developed separately. A plausible direction is to integrate explicit seed-word priors from KeyETM, ontology-level or graph-structured constraints from TopicNet and SED-KGraph, local grounding mechanisms from TeSG and MPS-CLIP, and discourse planning from DKRN and intent-keyword bridging (Harandizadeh et al., 2021, Duan et al., 2021, Zhu et al., 20 Jun 2025, Qin et al., 2020). Another plausible direction is to move from single-keyword or single-caption guidance to structured multi-keyword, token-level, and relation-aware control, since several adjacent methods already identify the limitations of one-word ablation, untyped graph propagation, or purely global sentence embeddings (Zhu et al., 20 Jun 2025, Izadi et al., 2022).

In that broader sense, TKSG is best understood not as one fixed architecture but as a design principle: semantic control is most effective when topics provide persistent global bias, keywords provide discriminative local evidence, and the model contains an explicit mechanism that links the two.