Associative Syntax and Maximal Repetitions reveal context-dependent complexity in fruit bat communication

AI summary

Overview

Research area: Computational bioacoustics and machine learning for animal communication, specifically unsupervised inference of vocal repertoires and syntax in a graded (non-discrete) vocal system, using fruit bats as a case study.

Technical level: Intermediate. The paper combines standard unsupervised pipelines (mel-spectrograms, UMAP, HDBSCAN, agglomerative clustering, Random Forest classification) with sequence-analysis tools (Maximal Repeats, transition networks). The concepts are explained accessibly, but familiarity with clustering metrics (ARI, NMI, Silhouette) and information-theoretic terms helps.

Scope in one sentence: The paper refines an existing unsupervised pipeline for labeling fruit bat syllables, then uses those labels to test whether bat syntax is associative or combinatorial and to measure communicative complexity across behavioral contexts via Maximal Repetitions and syllabic transition networks.

What This Paper Is About

Quantifying communication complexity in species with graded vocalizations is difficult because vocal units blend into one another, so existing manifold-learning pipelines (which assume clean unit boundaries) degrade on species like fruit bats. This paper refines that pipeline by systematically testing how dimensionality reduction of mel-spectrograms affects unsupervised labeling, then uses the resulting syllable labels to ask whether syllable order carries meaning and whether complexity varies by behavioral context. The authors benchmark against the fruit bat dataset of Prat et al. (2017), for which expert reference on repertoire size and syntax type exists.

Key Contributions

1. A refined unsupervised pipeline for repertoire quantification in graded vocal systems, improving on Sainburg et al. (2020) and producing results consistent with expert knowledge (Amit and Yovel, 2023).
2. An analysis of context-dependent syntax that adapts the method of Zhang et al. (2019) to use automatically labeled syllables across multiple context-dependent repertoires, rather than expert ground-truth labels.
3. The first application (to the author's knowledge) of Maximal Repeats (MRs) to animal communication, providing evidence for heavy-tailed MR distributions and proposing MR length as a metric of combinatorial complexity.
4. Findings suggesting higher communicative complexity (longer MRs, denser transition networks) in conflict behaviors than in cooperative ones.

Main Findings

• Improved clustering: Coarse-graining the temporal dimension of spectrograms from dynamically segmented vocal units yielded the best results (Silhouette > 0.5, 95% assignment accuracy), identifying seven types of vocal units and improving on the previous baseline, which discriminated only two (utterances of mothers and pups in Isolation versus utterances between adults in all other contexts).
• Repertoire estimates: An acoustic-similarity proxy (agglomerative clustering on Dynamic Time Warping distance over MFCCs, with quantile distance threshold q = 0.05) yielded 27 ± 2 syllable types per emitter, consistent with known fruit bat repertoire sizes. Agreement between this proxy and the best HDBSCAN clustering (mel-spectrograms retaining higher dimensionality, as in the original experiment) was moderate: mean ARI = 0.12 ± 0.01 and mean NMI = 0.30 ± 0.01, suggesting a repertoire of 14 syllables.
• Associative, not combinatorial, syntax: Syllable order did not affect behavioral context classification (F1-score > 0.9 for both original and permuted sequences). Failing to reject the null hypothesis HP1₀ supports associative rather than combinatorial syntax, consistent with Amit and Yovel (2023).
• Context-dependent syllable use: Syllable distribution was significantly different between Isolation and other contexts (p < 0.05, Wilcoxon rank-sum test). No significant evidence was found to reject HP2₀ for the cooperative contexts of Feeding, Grooming and Kissing in the majority of pairwise comparisons, suggesting more uniform syllable usage across those behaviors. Heatmaps also suggested that emitters raised in the same colony may not differ in syllable use.
• Heavy-tailed Maximal Repetitions: The likelihood ratio test rejected HP3₀ (p < 0.05). MR length distribution was best described by a truncated power-law (α = 1.79, reported in the abstract as α < 2), indicating long-range temporal structure rather than memory-less exponential decay.
• Higher complexity in conflict contexts: Average MR length was greater in conflict-related contexts (Mating Protest, Fighting, Threat-like) than in cooperative ones. Conflict-related networks showed small-world architecture (ω ≈ 0, average clustering coefficient Avg C > 0.4) with efficient global connectivity, while cooperative contexts showed metrics suggesting more random, less structured networks (ω > 0.5).
• Concrete network example (Bat #215): Mating Protest had support 629, σ = 1.00, ω = 0.00, 17 big cliques, 25 total cliques, density 0.81, Avg C 0.62; Kissing had support 48, σ = 0.86, ω = 0.63, 4 big cliques, 39 total cliques, density 0.13, Avg C 0.12. Isolation showed a simple repetition pattern (support 78, density 0.10, Avg C 0.00).
• Classifier feature importance: Features representing richness of contextual syntax, unpredictability of sequences, commonness of patterns and strength of short transitions (features e, f, g and p) accounted for about 50% of total Random Forest feature importance.

Methodology in Plain English

Dataset. The authors used the annotated fruit bat vocalization dataset from Prat et al. (2017), containing 41 specimens with emitter, addressee and behavioral context labels; vocal units were automatically segmented by the original dataset authors. Contexts analyzed were Mating Protests, Fighting, Threat-like, Biting, Feeding, Grooming, Kissing and Isolation (mother-pup interactions). Contexts labeled Generic, Sleeping or Unknown were excluded as ambiguous.

Experiment 1 (repertoire size and diversity). Following Sainburg et al. (2020), spectrograms (or autoencoder latent representations of them) are projected into a low-dimensional space with UMAP and clustered with HDBSCAN. The authors systematically varied the input: spectrogram time-frequency settings, dimensionality reduction (PCA on autoencoder latents, testing different autoencoder architectures), and segmentation method (the original fixed-noise-floor amplitude thresholding versus Dynamic Threshold Segmentation, which estimates the noise floor dynamically and isolates shorter sub-units). This analysis ran on the top-5 emitters, and the best configuration was scaled to the full dataset. Figure 2 diagnostic used 152,578 data points from all bats (41 individuals).

Evaluation without ground truth. Because there are no expert syllable labels, the authors used two strategies: (1) internal validation via Silhouette Score for HDBSCAN cluster consistency, and (2) agreement with an acoustic-similarity proxy built from agglomerative clustering on DTW distances between MFCCs, measured with ARI and NMI.

Experiment 2 (syntax and temporal structure). Vocalizations were encoded as syllabic sequences using the unsupervised labels. A Random Forest classifier adapted from Zhang et al. (2019) predicted behavioral context from 18 sequence-derived features (syllable richness, sequence length, transition count, entropy-based measures, Markov predictability, perplexity and others listed in Table 1). To test whether order matters, classification performance on original sequences was compared with permuted sequences (HP1₀). Syllable frequency distributions were compared between pairs of behaviors using Wilcoxon rank-sum tests (HP2₀). Maximal Repeats, the longest repeating subsequences, were extracted with a prefix-suffix tree algorithm, and a likelihood ratio test compared exponential versus power-law fits (HP3₀). Finally, mean MR length was compared across behaviors and syllabic transition networks were inspected both quantitatively (small-world coefficients, clique counts, density, clustering) and qualitatively.

Why This Matters

The paper addresses a known weakness in unsupervised animal-communication pipelines, namely their poor performance on species whose vocal units blend together, and it introduces a metric (Maximal Repeats) borrowed from computational linguistics that captures long-range combinatorial structure which Shannon entropy alone does not. It also offers an explicit alternative to the social complexity hypothesis for communicative complexity, which the authors describe as risking circularity between sociality and communication variables.

Potential real-world applications:

• Automated monitoring of animal vocal behavior from long passive-acoustic recordings, where human labeling does not scale.
• Welfare and management assessments in captive or farmed settings, where detecting conflict-related vocal signatures could serve as an early indicator of stress or aggression.
• Conservation bioacoustics, where repertoire and complexity metrics could complement population or biodiversity monitoring.
• Cross-species comparative research and bio-inspired sequence modeling, since the same MR and network toolkit could be applied to other graded vocal systems.

Industry relevance: The pipeline components (autoencoder compression of spectrograms, UMAP plus HDBSCAN clustering, sequence feature engineering, transition-network analysis) are directly transferable to commercial bioacoustics platforms, acoustic event detection, and any domain that needs unsupervised discovery of recurring structure in continuous time-series signals with heavy-tailed repetition patterns.

Future Directions

• Test Maximal Repeats as a proxy for combinatorial capacity in other species and across graded versus discrete vocal systems, as the authors explicitly propose.
• Resolve the discrepancy between the two repertoire estimates reported here (14 syllables from the best HDBSCAN clustering versus 27 ± 2 from the acoustic-similarity proxy), since the authors note that specific outcomes depend on the clustering method defining the repertoire.
• Extend the behavioral classifier and syntax analysis beyond the multiple behavioral classes tested here, and validate the interpretation of contexts as "conflictual" or "cooperative" against independent behavioral data, since the authors acknowledge these terms are their own interpretation of the original dataset's annotations.
• Refine and refactor the released codebase (the repository currently contains the unrefactored Master's thesis implementation, with a cleaner version under development).

Target Audience

This paper is most useful to computational bioacousticians and machine learning researchers working on animal vocal communication, particularly those dealing with graded vocal systems where standard clustering pipelines underperform. It also suits comparative psychologists and behavioral ecologists interested in quantitative measures of communicative complexity, computational linguists interested in Maximal Repeats and information-decay comparisons across species, and graduate students looking for a worked example of an unsupervised pipeline that is validated against both internal metrics and expert reference knowledge.

Authors’ abstract

This study presents an unsupervised method to infer discreteness, syntax and temporal structures of fruit-bats vocalizations, as a case study of graded vocal systems, and evaluates the complexity of communication patterns in relation with behavioral context. The method improved the baseline for unsupervised labeling of vocal units (i.e. syllables) through manifold learning, by investigating how dimensionality reduction on mel-spectrograms affects labeling, and comparing it with unsupervised labels based on acoustic similarity. We then encoded vocalizations as syllabic sequences to analyze the type of syntax, and extracted the Maximal Repetitions (MRs) to evaluate syntactical structures. We found evidence for: i) associative syntax, rather than combinatorial (context classification is unaffected by permutation of sequences, F 1 &gt; 0.9); ii) context-dependent use of syllables (Wilcoxon rank-sum tests, p-value &lt; 0.05); iii) heavy-tail distribution of MRs (truncated power-law, exponent α &lt; 2), indicative of mechanism encoding combinatorial complexity. Analysis of MRs and syllabic transition networks revealed that mother-pupil interactions were characterized by repetitions, while communication in conflict-contexts exhibited higher complexity (longer MRs and more interconnected vocal sequences) than non-agonistic contexts. We propose that communicative complexity is higher in scenarios of disagreement, reflecting lower compressibility of information.