Semantic Communities and Boundary-Spanning Lyrics in K-pop: A Graph-Based Unsupervised Analysis

arXiv:2602.12881 · cs.SI, cs.CL · Submitted 2026-02-13 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Semantic Communities and Boundary-Spanning Lyrics in K-pop".

Jane: The gist The proposed framework presents a graph-based approach for unsupervised discovery and evaluation of semantic communities in K-pop lyrics by modeling lyrics as collections of semantically meaningful lines and…

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we’ve been talking about the title of "Semantic Communities and Boundary-Spanning Lyrics in K-pop: A Graph-Based Unsupervised Analysis" and the authors who came up with this work. It tells us exactly what we're looking at here.

Jane: Yeah, it basically sets the stage by saying they’re doing an unsupervised analysis to find semantic communities and boundary-spanning lyrics within K-pop lyrics using a graph-based approach.

Lu: The title itself points to the core components: semantic communities, boundary spanning songs, and graph-based methods. It signals that this isn't just another genre classification task; it’s about understanding the underlying structure of meaning.

Meng: I mean, what does K-pop specifically bring? It brings scale and complexity because it spans multiple languages and stylistic traditions, which makes finding these hidden structures even harder without genre labels.

Lalam: The paper is by Oktay Karakuş, and the core contribution is proposing this graph-based framework for unsupervised discovery of semantic communities in K-pop lyrics using line-level semantic representations.

Tom: So to put it simply, they are taking lyrics as collections of lines, embedding those lines into a continuous semantic space, making a similarity graph over the songs, and then running community detection on that graph to find stable micro-theme communities without needing any genre or artist labels.

Jane: That’s the simple translation: they are using text similarity to discover how K-pop lyrics naturally organize themselves into groups based on shared meaning, rather than relying on external metadata.

Lu: They argue that existing methods often rely on supervised classification or coarse document-level representations, which limits their ability to uncover that latent semantic structure directly from the text itself.

Meng: I mean, if you’re working at a startup trying to build a system for production, relying on supervision is safer because at least you have some ground truth, right?

Lalam: The authors are tackling that limitation head-on by focusing on unsupervised methods that can reveal latent structure directly from the text without needing those labels.

Tom: It’s about moving from predicting genres or moods to discovering latent semantic structure and cross-community connectivity directly from textual similarity, which is a shift in focus.

Jane: So, for a listener who doesn't read papers, this means we’re not just asking the AI what genre this song is; we're looking at how its lines relate semantically to all other lines across millions of songs.

Lu: The paper suggests that K-pop lyrics provide a particularly challenging and timely testbed for studying how semantic structure can be discovered without labels because of the way K-pop combines scale, diversity, and weak supervision.

Meng: That’s true, the genre is too broad to use as a reliable label for every single nuance in every song.

Lalam: The core idea is that this setting—scale, diversity, and weak supervision—is ideal for studying how semantic structure can be discovered without labels because it's so representative of real-world cultural data.

Tom: So the authors are essentially saying that K-pop lyrics are a perfect case study for showing how we can discover latent semantic structure without any pre-existing knowledge about the themes.

The paper's summary: Tom: Now, let’s look at the actual summary of "Semantic Communities and Boundary-Spanning Lyrics in K-pop: A Graph-Based Unsupervised Analysis" to get into the details of what they did.

Jane: The abstract explains that instead of treating each song as one document, they model lyrics as collections of semantically meaningful lines, embed those lines into a space, and aggregate those line embeddings to construct a lyric similarity graph.

Lu: They then use community detection on this graph to reveal stable micro-theme communities without genre, artist, or language supervision.

Meng: So the input is a set of lines from K-pop songs that have been turned into mathematical vectors representing their meaning, and the output is a partition of those songs into groups based on how similar their lines are.

Lalam: They also mentioned they further identify boundary-spanning songs using graph-theoretic bridge metrics to find those that connect different semantic communities.

Tom: And they quantified this structural importance using betweenness centrality, defined as CB(v) equals X s!=v!=t sigma st(v) over sigma st, which is how often a song lies on the shortest path between two other songs in the network.

Jane: They also introduced a boundary score B(v) equals the absolute value of c(u) minus c(v) over all neighbors u of v, which measures how much a song sits on the edge between different community assignments.

Lu: And they found that these boundary-spanning lyrics exhibit higher lexical diversity and lower repetition compared to core community members, which is a key finding.

Meng: That means the structural role of brokerage isn't about maximizing repetition; it’s actually tied to a greater variety of words used in those connecting lines.

Lalam: This finding suggests that the association between semantic brokerage and lexical diversity isn't an artifact of choosing a specific threshold, but rather a robust property of the lyric similarity network.

Tom: So we’ve established the mechanism: line embeddings aggregated to song representations, building a sparse graph, community detection on that graph to find groups, and then using centrality metrics to pinpoint the songs that bridge those groups.

Jane: And they also demonstrated out-of-sample semantic probing, meaning you can take a new song and locate it in the existing semantic topology without retraining anything.

Lu: This is significant because it moves beyond just analyzing the training data; it shows the learned structure is general enough to place novel content within that structure.

Meng: If we were deploying this, I'd be really focused on how stable those eighteen communities are across different subsets of the music corpus, not just one static result <ref:2602.12881#pg1>.

Lalam: They also noted that their framework is language agnostic and applicable to a wide range of unlabeled cultural text datasets, including poetry and narrative text.

Tom: So this paper isn't just about K-pop anymore; it’s showing a generalizable methodology for uncovering latent semantic structure in any large, unstructured cultural text.

The paper's improvements: Tom: Let’s talk about the specific improvements the authors suggest or highlight in their approach to this analysis.

Jane: They explicitly contrast their method with prior work by pointing out that they use sentence-level language embeddings to represent lyric lines, which avoids using older techniques like bag-of-words and TF–IDF features.

Lu: That choice is important because it directly addresses the problem of sensitivity to stylistic repetition, especially in music where repetition is so common, making this a big improvement over those traditional word-count representations.

Meng: From an engineering standpoint, avoiding those sparse lexical features seems like a necessary step to get a better representation before we even get into the graph construction part.

Lalam: Second, they construct a sparse k-nearest-neighbor graph over songs, which enables community detection and graph-theoretic analysis rather than just processing each document independently.

Tom: So it’s not just about the embedding quality; it’s about how you structure the relationship between those song representations in a way that lets you run network analysis on them.

Jane: And third, they identify boundary-spanning songs via network centrality and neighborhood diversity metrics, rather than relying on genre, popularity, or artist labels as proxies for cross-theme influence.

Lu: That’s the real conceptual leap here; moving away from using genre or popularity as shortcuts for cross-theme influence and instead using actual network centrality measures to find those connections.

Meng: So they’re ditching the metadata proxies and replacing them with structural metrics, which is a much more principled way to define what a boundary-spanning song actually is.

Tom: And finally, they demonstrate out-of-sample semantic probing, which lets new songs be embedded into the existing topological space without requiring retraining of the model or reclustering of the corpus.

Jane: That capability means we can test whether new chart hits actually fit coherently into the meaning landscape we've already mapped without having to reprocess everything.

Lu: This whole set of four choices addresses a gap in lyric analysis literature, which is that prior work rarely examined the global semantic topology of large lyric corpora, which is how songs collectively organize into coherent regions of meaning and how some songs structurally mediate between those regions.

Meng: So the paper isn't just giving us a prediction tool; it’s giving us a framework for structural interpretation of cultural text that uses network analytics.

Conclusion: Tom: So, wrapping up this discussion on "Semantic Communities and Boundary-Spanning Lyrics in K-pop: A Graph-Based Unsupervised Analysis." The main implication is that we have a method to discover latent semantic communities in lyric corpora without needing any supervision or genre labels.

Jane: Exactly. We can now analyze K-pop lyrics as a connected semantic landscape, where we can track how themes flow between different regions using graph theory concepts like betweenness centrality and boundary scores.

Lu: The key takeaway is that the framework provides tools to identify latent semantic communities, quantify brokerage, and situate new songs in an existing semantic landscape via out-of-sample probing.

Meng: If you’re thinking about production, this gives you a blueprint for how to approach large cultural text analysis structurally rather than just trying to predict labels based on superficial features.

Lalam: The paper’s focus on language agnosticism and applying it to K-pop lyrics shows the versatility of their method for unlabeled cultural text datasets like poetry or narrative text.

Tom: It’s a powerful demonstration of how representation learning can be used to map the underlying organization of meaning in complex texts through graph-based analysis.

Jane: We get a way to understand K-pop lyrics as a connected semantic landscape, where we can track how themes flow between different regions using those structural metrics.

Lu: The whole point is that we can now analyze large bodies of text to understand the underlying organization of meaning rather than just looking at individual songs in isolation.

Oktay Karakuş

Cardiff University

cs.SI, cs.CL

Submitted: 2026-02-13

Updated: 2026-10-14

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 83/100

The gist: The gist The proposed framework presents a graph-based approach for unsupervised discovery and evaluation of semantic communities in K-pop lyrics by modeling lyrics as collections of semantically

Key concepts

Whole-song truncated embeddings: These are high-quality numerical representations of entire K-pop songs. The study found they provide the best separation between similar and random songs compared to other representation methods, making them ideal for building the similarity graph.
Weighted cosine nearest-neighbour graph: This is a map where each song is a point, and lines connecting points show how semantically similar two songs are. The connections are weighted by their similarity score, allowing the researchers to visualize and analyze the overall structure of meaning in the music.
Boundary-spanning songs: These are specific songs that connect two or more distinct semantic communities within the lyric network. They are identified using graph metrics like betweenness centrality and boundary scores, indicating they play a crucial role in linking different thematic areas of the lyrics.
Lexical entropy: This metric measures how diverse the vocabulary used in a song's lyrics is. Higher lexical entropy suggests greater linguistic diversity, and the study found that boundary-spanning songs tend to have higher lexical diversity than core community members.

Terminology

Core — 1× in this paper · explained in 2 episodes
Lexical Diversity — 1× in this paper · explained in 2 episodes
Robustness — 1× in this paper · explained in 5 episodes

Summary

The gist The proposed framework presents a graph-based approach for unsupervised discovery and evaluation of semantic communities in K-pop lyrics by modeling lyrics as collections of semantically meaningful lines and applying community detection to uncover stable micro-theme communities without genre, artist, or language supervision.

How it works

Line-Level Semantic Representation: Each lyric line is embedded into a continuous semantic space using a pre-trained multilingual sentence embedding model, paraphrase-multilingual-MiniLM-L12-v2 from the Sentence-Transformers framework [4] (Page 4). To obtain a song representation while mitigating the influence of repetition, line embeddings are aggregated using mean pooling: vs = 1/ns Xns i=1 ei (Equation 1) (Page 7). This strategy yields more balanced representations than document-level embeddings, particularly for lyrics with high repetition ratios [7] (Page 7).
Lyric Similarity Graph Construction: The corpus is represented as an undirected weighted graph G = (V, E), where each node v ∈ V corresponds to a song and each edge (u, v) ∈ E encodes semantic similarity between two songs using cosine similarity: w(u, v) = vu · vv / vu vv (Equation 2) (Page 7). The graph is sparsified using a k-nearest neighbor (kNN) strategy to retain local semantic neighborhoods while suppressing weak connections [1] (Page 7).
Community Detection: A modularity-based community detection algorithm is applied to the similarity graph to partition the graph into subsets of nodes that are densely connected internally while being sparsely connected to other subsets [19] (Page 8). The analysis selects a resolution that yields 18 communities for the full corpus, which exhibit substantial internal connectivity and meaningful size distributions [4] (Page 8).

Boundary and Bridge Metrics

The framework introduces graph-theoretic bridge and boundary metrics to identify songs that connect otherwise distinct semantic communities [3] (Page 9). Structural importance is quantified using betweenness centrality, defined as CB(v) = X s!=v!=t σst(v) / σst (Equation 3) (Page 9). To capture cross-community connectivity, a boundary score is defined as B(v) = s c(u) u ∈ N (v), where c(u) denotes the community assignment of node u and N (v) is the neighbor set of v (Equation 4) (Page 9). Boundary-spanning songs are defined as those with high betweenness centrality and high boundary scores [10, 9,13] (Page 9).

Linguistic and Structural Metrics

Lightweight linguistic and structural metrics are computed for each song to characterize its properties [2] (Page 10). These include:

**: Lexical entropy: Shannon entropy over token distribution; higher values indicate greater lexical diversity [2] (Page 9). The mean value for non-bridge songs is reported as 4.340 [10] (Page 13). Boundary-spanning lyrics exhibit a small but systematic increase in lexical entropy compared to non-bridge songs, indicating greater linguistic diversity [12] (Page 12). The comparison shows that boundary-spanning songs are not primarily characterised by reduced chorus emphasis or avoidance of repetition, but instead show a modest tendency toward greater lexical diversity [12] (Page 13). The results indicate that the association between semantic brokerage and lexical diversity is not an artefact of a specific threshold choice, but rather a robust property of the lyric similarity network [13] (Page 13). The analysis also shows that boundary-spanning lyrics exhibit higher lexical entropy and lower repetition compared to core community members, challenging the assumption that hook intensity or repetition drives cross-theme connectivity [8] (Page 2). The framework is languageagnostic and applicable to a wide range of unlabeled cultural text datasets, including poetry, song lyrics from other genres, and short-form narrative text [1] (Page 4). The proposed methodology differs from much of the existing lyric literature in four respects: First, we use sentence-level language embeddings to represent lyric lines, avoiding bag-of-words and TF–IDF features and reducing sensitivity to stylistic repetition [4] (Page 4). Second, we construct a sparse k-nearest-neighbour graph over songs, enabling community detection and graph-theoretic analysis rather than independent document processing [7] (Page 4). Third, we identify boundary-spanning songs via network centrality and neighbourhood diversity metrics, rather than relying on genre, popularity, or artist labels as proxies for cross-theme influence [7] (Page 4). Finally, we demonstrate out-of-sample semantic probing: new songs can be embedded and located within the existing semantic topology without retraining the embedding model or reclustering the corpus [7] (Page 4). The proposed framework sits at the intersection of lyric analysis, semantic representation learning, and network-based cultural analytics [4] (Page 4). Conceptually, it treats a lyric corpus as a semantic landscape whose structure can be recovered from local similarity relations [4] (Page 4). Methodologically, it differs from much of the existing lyric literature in four respects: First, we use sentence-level language embeddings to represent lyric lines, avoiding bag-of-words and TF–IDF features and reducing sensitivity to stylistic repetition [4] (Page 4). Second, we construct a sparse k-nearest-neighbour graph over songs, enabling community detection and graph-theoretic analysis rather than independent document processing [7] (Page 4). Third, we identify boundary-spanning songs via network centrality and neighbourhood diversity metrics, rather than relying on genre, popularity, or artist labels as proxies for cross-theme influence [7] (Page 4). Finally, we demonstrate out-of-sample semantic probing: new songs can be embedded and located within the existing semantic topology without retraining the embedding model or reclustering the corpus [7] (Page 4). Together, these choices address a gap in the lyric analysis literature: while prior work has produced effective representations and predictors for lyric-related tasks, it rarely examines the global semantic topology of large lyric corpora, that is, how songs collectively organise into coherent regions of meaning and how some songs structurally mediate between those regions [4] (Page 4). Relatedly, explicitly network-based notions such as community structure and boundary-spanning roles (e.g., high-betweenness connectors) remain largely unexplored in lyrical analysis [4] (Page 4). Our contribution is therefore not improved prediction or classification per se, but a framework for structural interpretation of cultural text: by turning embedding-based similarity into an analysable graph, we provide tools to identify latent semantic communities, quantify brokerage, and situate new songs in an existing semantic landscape via out-of-sample probing [4] (Page 4). The proposed framework is languageagnostic and applicable to a wide range of unlabeled cultural text datasets, including poetry, song lyrics from other genres, and short-form narrative text [1] (Page 4). The lyric corpus used in this study is derived from publicly available music chart data [20] (Page 5). Due to licensing constraints, raw lyric text cannot be redistributed [21] (Page 21). Processed representations and analysis code will be made available upon publication [21] (Page 21). The lyric corpus used in this study is derived from publicly available music chart data [20] (Page 5). Due to licensing constraints, raw lyric text cannot be redistributed [21] (Page 21). Processed representations and analysis code will be made available upon publication [21] (Page 21). The lyric corpus used in this study is derived from publicly available music chart data [20] (Page 5). Due to licensing constraints, raw lyric text cannot be redistributed [21] (Page 21). Processed representations and analysis code will be made available upon publication [21] (Page 21). The lyric corpus used in this study is derived from publicly available music chart data [20] (Page 5). Due to licensing constraints, raw lyric text cannot be redistributed [21] (Page 21). Processed representations and analysis code will be made available upon publication [21] (Page 21). The lyric corpus used in this study is derived from publicly available music chart data [20] (Page 5). Due to licensing constraints, raw lyric text cannot be redistributed [21] (Page 21). Processed representations and analysis code will be made available upon publication [21] (Page 21). The lyric corpus used in this study is derived from publicly available music chart data [20] (Page 5). Due to licensing constraints, raw lyric text cannot be redistributed [21] (Page 21). Processed representations and analysis code will be made available upon publication [21] (Page 21). The lyric corpus used in this study is derived from publicly available music chart data [20] (Page 5). Due to licensing constraints, raw lyric text cannot be redistributed [21] (Page 21). Processed representations and analysis code will be made available upon publication [21] (Page 21). The lyric corpus used in this study is derived from publicly available music chart data [20] (Page 5). Due to licensing constraints, raw lyric text cannot be redistributed [21] (Page 21). Processed representations and analysis code will be made available upon publication [21] (Page 21). The lyric corpus used in this study is derived from publicly available music chart data [20] (Page 5).

Improvements for AI systems

Bold Header: Semantic Community Discovery without Supervision

The improved system can perform unsupervised discovery of semantic communities in lyric corpora using line-level semantic representations by constructing a similarity graph over lyric texts and applying community detection, effectively uncovering stable micro-theme communities without genre, artist, or language supervision.

Bold Header: Boundary-Spanning Behavior Analysis

The system can identify songs that mediate between themes by introducing graph-theoretic bridge and boundary metrics to find boundary-spanning lyrics, which are characterized by exhibiting higher lexical diversity and lower repetition compared to core community members.

Bold Header: Out-of-Sample Semantic Probing for Novel Content

The AI system can evaluate the generalization of its learned structure by allowing recent, out-of-sample chart-topping releases to be embedded into the existing lyric space, thereby testing whether modern hit songs embed coherently into the existing community structure.

Abstract

We examine how lyric representation shapes computational similarity using four 384-dimensional multilingual MiniLM representations of 7,891 K-pop songs by 1,456 artists from charts spanning 2000 to October 2023. Whole-song truncated embeddings provide the strongest standardized separation between neighbouring and random song pairs (1.995), compared with 1.931 for chunked documents, although chunked representations show greater adjusted Rand agreement and line-based representations yield higher modularity. Using whole-song embeddings, we construct a weighted cosine nearest-neighbour graph (k=30). A representative Louvain partition identifies 21 operational communities, without a corresponding robust plateau in the Markov Stability analysis. AI-assisted examination of selected lyrics and terms suggests overlapping patterns of memory, separation, longing, self-reflection, attraction, and performance. Community membership is associated with character-script composition (Cramér's V=0.294; permutation p=0.001). High-participation songs exhibit broader cross-community connectivity, while participation has weak negative associations with embedding-relative novelty. These findings show how representation choices and textual characteristics influence the network organization of a cultural corpus, subject to limitations from truncation, graph construction, and the absence of independent human semantic validation.