Comparing context-dependent call sequences employing machine learning methods: an indication of syntactic structure of greater horseshoe bats

Understanding how human language evolved from earlier forms of animal communication is the key to fully appreciating its unique capacity in human communication (Scarantino and Clay, 2015). One of the most powerful features of human language is grammatical rules such as word ordering under various backgrounds (Fitch, 2010). Multisyllabic vocalizations with syntax (the set of rules for combining words into phrases as in human sentences) have been found in several species of birds and mammals (Collier et al., 2014; Schlenker et al., 2016). Furthermore, social context has been shown to be essential for song structure and note use in songbirds (Catchpole and Slater, 2003; Byers and Kroodsma, 2009). In mammals, although the ability to construct calls/songs from vocal units has been reported in various species (Clarke et al., 2006; Bohn et al., 2009; Fedurek and Slocombe, 2011; Green et al., 2011; Filatova et al., 2012; Kershenbaum et al., 2012), evidence of altering vocal composition and structure owing to social context has been reported in only a few species (Bohn et al., 2008; Candiotti et al., 2012; Cäsar et al., 2013; Chabout et al., 2015). Exploring the ways in which vocal elements are ordered and combined under various social cues or behavioural contexts in mammalian species is critical to advancing our understanding of the function and evolutionary origin of syntax.

Previous research described and compared animal syntax based on various sound features, including the spectral and temporal parameters of syllables, numbers and types of syllables, probability of different syllable types occurring, temporal emission patterns and probability of syllable transitions (Briefer et al., 2013; Deslandes et al., 2014; Chabout et al., 2015; Lin et al., 2016a). These features are based on different subunits such as syllables, phrases and bouts of animal vocalizations, and they can reveal the call structure and the rules of syllable ordering in animal vocalizations. However, most studies have used components of these features in one research paradigm (Gadziola et al., 2012; Bohn et al., 2013; Suzuki et al., 2018). For example, Chabout et al. (2015) mainly used the probability of occurrence of a transition type between syllables to compare male mice song syntax depending on social contexts, while Cäsar et al. (2013) compared the types of call sequences between study groups to state that Titi monkey calls varied with predator location and type. Moreover, analyses that are conducted at the level of sequences, such as Markov chains and Zipf's law, have usually only adopted call sequences of sufficient length (Berwick et al., 2011; Deslandes et al., 2014), which might cause information to be neglected owing to the different lengths of sequences. In addition, limited work has been conducted for exploring the contribution of different features for syntactic classification. Therefore, before further analysis of the context-dependent syntax in animal vocalization can proceed, it would be efficient if a method could be constructed to classify the call sequences in different contexts by integrating the features of different subunit levels and evaluating the contribution of each feature to the classification from a large quantity of sound recording data.

Machine learning is a subset of artificial intelligence in the field of computer science, often using statistical techniques to give computers the ability to ‘learn’ from data (Michalski et al., 2013). In recent years, machine learning has been shown to be a powerful tool and has been widely applied to make predictions and discover hidden structure within large datasets, as well as to deal with classification problems in many circumstances (Skowronski and Harris, 2006; Acevedo et al., 2009; Stathopoulos et al., 2018). In bioacoustics, numerous studies have applied various machine learning tools for species recognition in environmental monitoring (Huang et al., 2009; Walters et al., 2012; Shamir et al., 2014; Aodha et al., 2018) and automatic classification of animal vocalizations (Ranjard and Ross, 2008; Armitage and Ober, 2010; Pozzi et al., 2010; Turesson et al., 2016). The successful application of machine learning tools indicates their importance for constructing a transparent, fast and accurate algorithm for animal acoustic signals analysis.

Bats have a suite of features that indicate a neural substrate supporting vocal plasticity and complexity, such as neural adaptations to support laryngeal echolocation (Siemers et al., 2011; Fenton et al., 2012), vocal learning (Knörnschild, 2014) and geographical divergence of vocalizations (Lin et al., 2015; Prat et al., 2017). Furthermore, some species have been documented to possess the capacity to vary social calls in response to social cues and behavioural context. For example, Brazilian free-tailed bats, Tadarida brasiliensis, quickly varied song composition to meet the specific demands of different social functions (Bohn et al., 2013). Big brown bats, Eptersicus fuscus, emitted bouts of vocalizations that could be assigned to specific aggressive behaviours (Gadziola et al., 2012). Mexican free-tailed bats (also known as Brazilian free-tailed bats) produced fixed vocal compositions, including irritation calls, protest calls and warning calls during agonistic interactions (Bohn et al., 2008). The complexity and plasticity of acoustic communication systems observed in bats make them an important template for studies of acoustic communication. However, to date, syntactic structures of only songs and aggressive calls have been reported in bats. It remains unclear whether syntax exists in other social contexts of other bat species and whether the syntax is the same or differs between different contexts.

Aggressive encounters and distress calls occur in a wide range of animal groups. Aggressive encounters occur when the individuals compete for limited resources such as mates, food, shelter or territories (Bradbury and Vehrencamp, 2011). Distress calls as a categorical alarm signal are usually produced by vertebrates when cornered, attacked or captured by a predator (Magrath et al., 2015). In echolocating bats that have weak vision and that normally live in dark environments, acoustic signals play a primary role in information exchange in the two behavioural contexts (Gillam and Fenton, 2016). As mentioned above, syntax of aggressive calls in Mexican free-tailed bats and big brown bats has been reported (Bohn et al., 2008; Gadziola et al., 2012), but it is still unclear in other species. For distress calls, previous studies have revealed both interspecific and conspecific acoustic similarity of the call structures, and the bats could recognize and respond to the acoustic similarity (Russ et al., 2004; Eckenweber and Knörnschild, 2016; Huang et al., 2018). These findings suggested the potential of existing rules in which the vocal units were ordered and combined, that is, existing syntax in a distress context.

Greater horseshoe bats (Rhinolophus ferrumequinum) have a large vocal repertoire (Ma et al., 2006; Jiang et al., 2017). The bats roost in the tens to hundreds and have all-female or mixed-sex colonies that fluctuate in size across seasons. Previous studies revealed a broad diversity of the vocalizations in this species that may reflect the existing rules for syllable ordering and combining responses to different social and behavioural contexts (Jones and Siemers, 2011; Luo et al., 2013; Lin et al., 2016b). Therefore, we recorded the social calls of greater horseshoe bats in aggressive and distress contexts as study templates. Our aim was to employ machine learning methods to classify the call sequences in different behavioural contexts by integrating the sound features of all subunit levels (syllables, transitions and sequences) and extracting the features that play large roles in the classification. The analysis may provide a fast and efficient path to mining useful information from large amounts of vocalizational data, so that the experimenter can select the important features for further analysis.

In May 2016, we captured eight adult Rhinolophus ferrumequinum (Schreber 1774) (4 males, 4 females) with mist nets from Dalazi Cave in Zhi'an Village, Jilin, Peoples Republic of China. Bats were housed in a laboratory with regulated temperature (20−25°C), humidity (50%−70%) and light:dark cycles (natural photoperiod in Changchun). Experimental bats had free access to sufficient mealworms and fresh water in dishes every day. All experimental procedures complied with the ABS/ASAB guidelines for the Use of Animals in Research and were approved by the Committee on the Use and Care of Animals at the Northeast Normal University (approval number: NENU-W-2010–101). All bats were released into their roosts after the experiment.

Bats often emit echolocation calls and distress calls when they are captured by predators or experimenters. Distress calls are usually recorded from handheld bats, as predation events of bats are very rare (Russ et al., 2004; Luo et al., 2013; Lin et al., 2015; Huang et al., 2018). Therefore, we recorded distress calls from the handheld bats. Distress calls were recorded using an UltrasoundGate 116 (Avisoft Bioacoustics, Berlin, Germany) connected to a laptop computer (at a sampling rate of 250 kHz at 16 bits per sample). The condenser microphone, with a flat frequency response between 10 Hz and 200 kHz (±3 dB), was set on a small tripod 1 m from the hand-held bat. During recording, each bat was held gently and its lower back was gently massaged by the researcher. In this case, a 4 min sound file of distress calls was produced for every individual.

After distress call recording, each bat was marked with 4.2-mm numbered aluminium alloy band (Porzana Ltd, East Sussex, UK). Recent studies in our laboratory (Jiang et al., 2017; Sun et al., 2019) have confirmed that the bands do not change the normal behaviour of the bats. For each recording trial, four individuals (one male and three females) were randomly selected and housed in a small cage (33×33×25 cm). To obtain natural calls, we adopted the natural paradigm (Gadziola et al., 2012) in which bats were undisturbed and recorded for several hours during their active period. Each recording trial lasted from 17:00 to 06:00 h on the next day, and data were recorded via the UltrasoundGate system at a sampling rate of 250 kHz at 16 bits per sample. Synchronized videos were filmed via an infrared digital video camera (Sony HDR PJ760E). Bats typically produce aggressive calls when one individual (intruder) disturbs another (resident) as it might jostle for a roost position within the group (Zhao et al., 2018, 2019; Sun et al., 2019), but the aggressive call structure varies with the degree of agonistic encounter (Gadziola et al., 2012). To obtain the aggressive calls under relatively the same aggression degree, we only used the call sequences produced during physical conflict for analysis. During the physical conflict, only the resident that was disturbed vocalized, while the intruder or other undisturbed residents did not produce any sounds. Thus, vocalizations recorded during aggressive contexts rarely contained overlapping signals from multiple animals. In addition, the caller could be visually identified in the video, because it would open its mouth when producing aggressive calls. The recording trials were repeated until no new syllable types were found in the call sequences.

We analysed the recorded call sequences using Avisoft SASLab Pro (Avisoft Bioacoustics, Berlin, Germany). The social vocalizations were described and classified following the nomenclature given by Kanwal et al. (1994) and followed by others, e.g. Ma et al. (2006) and Gadziola et al. (2012). Simple syllable types were named according to call structure, with a prefix denoting secondary spectral features [e.g. SFM, sinusoidal FM; BNB, broadband noise burst (NB); DFM, downward FM; UFM, upward FM] and a suffix denoting secondary temporal features (e.g. BNBl, long BNB; BNBs, short BNB). The composite syllable types were named according to the empirically established combination of simple syllables and abbreviated accordingly (e.g. NB-DFM, noise burst-DFM). The composited syllables with more than two components were named according to the combination of the first letter of names of each syllable (e.g. NSND for NB-SFM-NB-DFM).

Because the purpose of the present study was to compare the sequence structures, we used each call sequence as an analysis unit. During distress contexts, the held bats emitted distress call sequences with multiple syllables separated by inter-syllable intervals. A distress call sequence was determined if it was separated from other calls by intervals exceeding four times the average inter-syllable intervals within this sequence (Jiang et al., 2017).

During the physical contact in aggressive contexts, the disturbed residents usually fought back with wing flapping or boxing moves. We determined a separate agonistic behaviour when it started from the first wing flap or boxing move given by any disturbed individual and ended when every individual had calmed down. A sequence of multiple syllables emitted by the disturbed resident during each separate agonistic behaviour was defined as an aggressive call sequence for the following analyses.

All sequences recorded in aggressive and distress contexts were used for machine learning classification models. We extracted 12 features used frequently in bat acoustic analysis from each sequence using a self-written Python program. These features were as follows: (a) total number of syllable types occurring in a sequence; (b) total number of syllables of all types in a sequence; (c) total number of transitions (transitions between two adjacent syllables) types in a sequence; (d) a/c, addressing the linearity of the way syllables are ordered in a sequence (Scharff and Nottebohm, 1991); (e) c/total number of transition types under the behaviour context, expressing the consistency of the occurrence frequency of one transition type (Scharff and Nottebohm, 1991); (f) entropy, calculated with ⁠, where P_i is the probability of occurrence of the ith syllable type, and n is the number of syllable types; (g) product of the probabilities of each syllable occurring in a certain position of the sequence; (h) the product of probabilities of each transition occurring in the current context; (i) a/b, representing the versatility of a sequence; (j) uncertainty of transitions, calculated as ⁠, where H (x) measures a given syllable x to the rest of the n syllables that follow, and P_i represents the probability of the transition from x to i (Hailman et al., 1985); (k) gender of bats that emitted the sequence; and (l) marker label of the bats, representing the bat individual. Features a and b were related to syllables; features c, e, h and j describe transitions in one sequence; features d, f and i describe characteristics of sequence structure; feature g describes positions of syllables occurring in a sequence; and features k and l concern individuals' information. The class names representing aggressive and distress contexts were labelled with integers (0, 1) as training targets.

Nominal features such as gender and marker label were converted to new dummy variables via the one-hot encoding technique to avoid technical glitches. Because many linear models, such as the logistic regression and SVM, initialize the weights to 0 or small random values close to 0, standardization was used to center the feature columns at a mean of 0 with a standard deviation of 1 so that the feature columns take the form of a normal distribution, which makes it easier to learn the weights.

All the classification algorithms were implemented with Scikit-learn (version 0.19.1), a Python module integrating classical machine learning algorithms (Pedregosa et al., 2011; Raschka and Mirjalili, 2017).

The logistic regression classifier is a powerful and widely used algorithm for linear and binary classification problems. As the name suggests, in the logistic regression model, the weighted input features are fed to the logistic function. Then, the probability of each sequence belonging to a given context was calculated with the odds ratio. Finally, the predicted probability can then be converted into a binary outcome. The logistic regression model has been used for research on habitat selection of animal populations (Prugh et al., 2008; Duchesne et al., 2010) and other scenarios of data mining. Here, we trained the classifier with the regularization parameter C=1000.0 and regularized using the L2 norm of the classifier weights. We used the implementation in sklearn.linear_model. LogisticRegression.

The SVM is a sophisticated kernel-based machine learning classifier and has attracted much attention as a new classification technique with good generalization ability (Cristianini and Shawe-Taylor, 2000). SVMs have been widely applied to species recognition and acoustical classification in animals (Fagerlund, 2007; Chen et al., 2012). In SVMs, the optimization objective is to maximize the margin, the distance between the separating hyperplane (each hyperplane representing the call sequences of each context) and the training samples that are closest to this hyperplane, the so-called support vectors. The SVM induction process aims to establish an optimal discriminative function between two classes of call sequences in two contexts while accomplishing the trade-off between generalization and overfitting. Here, we used the sklearn.svm.SVC method with kernel=ʻlinear’ and hyperparameter C=1.0.

Decision trees are a non-parametric supervised learning method used for classification and regression. The goal is to create a model that predicts the value of a target variable by learning simple decision rules inferred from the data features. Using the decision algorithm, we started at the tree root and split the data on the feature that results in the largest information gain. Feature scaling (standardization) is not a requirement for decision tree algorithms. Decision trees are readily interpretable, directly used to generate rules, and computationally inexpensive to train, evaluate and store (Valletta et al., 2017). Here, we used DecisionTreeClassifier in sklearn.tree to train a decision tree with criterion=ʻentropy’ and max_depth=5.

After the model construction and feature selection, all 12 features of each call sequence in two contexts were standardized by the StandardScaler function in the scikit-learn package and collected as a dataset. Then the dataset was randomly partitioned into a separate test dataset (804 samples) and a training dataset (1875 samples) with the ratio of 3:7. The three models for sequence classification were trained with the training set and tested with the test set. Hyperparameters specific to each machine learning algorithm were tuned by cross-validation to strike a balance between underfitting and overfitting.

The performance of the machine learning models was characterized with area under the curve in receiver operator characteristic graphs, a useful tool for estimating model performance with respect to the false positive and true positive rates (Raschka and Mirjalili, 2017).

The contribution of each feature to the classification of call sequences between contexts was measured as the average impurity decrease in the random forest model. It was computed from all decision trees in the forest without making any assumptions, e.g. whether our data were linearly separable. Random forests can handle thousands of mixed categorical and continuous predictors and are robust to outliers, and therefore they are often used to compute an estimate of the importance of every predictor (Valletta et al., 2017). We assessed the importance of the features via the feature_importance attribute after fitting a random forest classifier with 10,000 trees.

Statistical analyses were conducted to test the difference of sequence structures between aggressive and distress calls. Pearson's chi-square statistic (chisquare and chi2_consistency) were used to test the occurrence frequencies of different syllable types, the probabilities of occurrences of transition types and the probabilities of syllable types occurring in each position of a sequence. Features related to sequence characteristics were tested with Wilcoxon rank-sum statistics (ranksums in scipy). The statistical analyses were performed using the scipy.stats package in IPython for Windows version 5.1.0. The significance level of all tests was set at 0.05.

A total of 2014 aggressive call sequences were collected from 12,337-min recordings, consisting of 4692 syllables belonging to 30 different types. In the distress context, 665 call sequences were collected containing 2552 syllables belonging to 19 types (Table 1, Fig. 1). First, we employed machine learning models to classify the syntactic structures in different behavioural contexts and extracted the most important features for the classification. Then, statistical analysis was employed to test the results of the machine learning models.

The 12 sound features extracted from various levels of call subunits including syllables, phrases and sequences were used to train the logistic regression, SVM and decision trees algorithms and perform the classification. The classification performance of the three algorithms showed impressively high accuracies on categorizing all sequences to two contexts: logistic regression 98%, SVM 97% and decision trees 96% (Fig. 2).

The contribution of each feature to the discrepancy was measured by the random forest model (Fig. 3). The most important features were related to syllable transitions (features e, h, j and c, 35.7%). The second most important features concerned the characteristics of call sequences (features d, f and i, 21.4%), and the third most important were about the positions where each syllable occurred in a sequence (feature g, 16.1%). The features concerning the sequences structure were more important than features about number or types of syllables (features a and b, 4.7%), bat individuals (feature l, 14.3%) and gender (feature k, 7.7%).

Because syllables were fundamental units of call sequences, numbers and types of syllables in two contexts were compared first, although features related to syllables contributed the least percentages. The percentages of shared syllable types in the two contexts were both greater than 90% (aggressive: 95.0%, distress: 90.5%; each syllable type with a frequency of occurrence greater than two was counted), which showed high similarity. However, in both aggressive and distress contexts, greater horseshoe bats had intra-context selective preferences in using different syllable types (chi-square test: aggressive: χ²=26,275, d.f.=29, P<0.001, distress: χ²=8524.1, d.f.=18, P<0.001). NB-SFM, NB-DFM and SFM were used the most by bats (aggressive: 77.8%, distress: 69.9%). Some syllable types such as SNU and SNSNS were only found several times. The syllable types occurring most frequently between two contexts were different. Bats tended to use NB-SFM in aggressive contexts but tended to use NB-DFM in distress contexts (chi-square test: χ²=632.8, d.f.=29, P<0.001; Fig. 4).

Transition types of two adjacent syllables in each context were also selectively used by greater horseshoe bats (Fig. 5). In aggressive calls, the transitions used frequently were self-transitions of NB-DFM (n=402), NB-SFM (n=368) and SFM (n=220) (chi-square test: χ²=24,858.4, d.f.=155, P<0.001; Fig. 5A). In distress calls, the transitions used most frequently were self-transitions of NB-DFM (n=485), NB-SFM (n=126) and BNBl (n=138) (chi-square test: χ²=19,485, d.f.=121, P<0.001; Fig. 5B). Although self-transition of NB-DFM was the most frequently used transition type in both contexts, it occupied a higher percentage in distress calls (25.7%) than in aggressive calls (15%) (chi-square test: χ²=856, d.f.=184, P<0.001).

Features related to characteristics of call sequences were sequence entropy (d), linearity (f) and versatility (i) (Fig. 6). Difference in entropy indicated more variable syllables occurring in distress contexts than in aggressive contexts (F=467,414, P<0.001). Sequences in aggressive context had low linearity (F=506,663, P<0.001), meaning that the ways in which syllables were ordered were more variable (or had more transition types), but high versatility (F=876,051, P<0.001), meaning there were more syllable types or shorter sequence lengths in call sequences of the aggressive context than of the distress context. These results were consistent with the overview of sequences under the two contexts presented in Table 1.

To obtain the occurrence frequency of syllable types in different positions of a sequence, we counted the numbers of the first three positions of call sequences in the aggressive context and the first four positions of call sequences in the distress context based on the average length of all sequences (Fig. 7). In aggressive calls, NB-SFM had the highest occurrence frequency in the first three positions (Fig. 7). In distress calls, NB-DFM had the highest occurrence frequency in the first four positions (Fig. 7). The probability of occurrence of syllable types in each position within a context were significantly different (aggressive context: chi-square test: position 1: χ²=7695.4, d.f.=21, P<0.001, position 2: χ²=3735.2, d.f.=18, P<0.001, position 3: χ²=1745.8, d.f.=16, P<0.001; distress context: position 1: χ²=2496.5, d.f.=18, P<0.001, position 2: χ²=1358.2, d.f.=14, P<0.001, position 3: χ²=674.75, d.f.=13, P<0.001, position 4: χ²=603.4, d.f.=13, P<0.001). Significant differences in the occurrence probability of syllable types in the first three positions also existed between the two contexts (chi-square test: position 1: χ²=228.7, d.f.=22, P<0.001, position 2: χ²=144.8, d.f.=19, P<0.001, position 3: χ²=93.8, d.f.=18, P<0.001).

This study employed machine learning methods to classify the call sequences between aggressive and distress contexts in greater horseshoe bats and to extract the features that play important roles in the classification. Logistic regression, SVM and decision trees were trained using 12 features, and each method obtained accurate classification rates greater than 95% (logistic regression 98%, SVM 97% and decision trees 96%; Fig. 2). The top three most important features for classification were all related to the structures of call sequences, including syllable transitions, positions and characteristics of call sequences (versatility, consistency and entropy). Moreover, the statistical comparison of selective preferences of syllable types, transitions, syllables occurring in different positions of a sequence and the differences in the characteristic of sequences highlighted the point that discrepancy in sequence structure existed between the two behavioural contexts. The good performance of the algorithms and the extracted important features indicated that machine learning algorithms could be a powerful tool for classifying call sequences of social vocalizations between different contexts. The method presented should enable data mining from large sound datasets in the initial step of studies on the syntax of social vocalizations in bats.

Machine learning methods could be an ideal choice for acoustic research owing to their good generalization to numerous studies. The methods have been used to solve complex problems that were previously intractable, such as dealing with large datasets and acoustic recognition in multi-species and complicated environments (Walters et al., 2012; Shamir et al., 2014; Priyadarshani et al., 2018). Recently, a few studies have applied machine learning tools to acoustic detection and species classification in bats by analysing their echolocation calls (Skowronski and Harris, 2006; Armitage and Ober, 2010; Aodha et al., 2018). Our results indicated that machine learning could also be used to classify call sequences of social vocalizations in different behavioural contexts by integrating sound features of vocal subunits. Although discriminant analysis has also been frequently used for classification (Lachenbruch and Goldstein, 1979), it is more suitable for displaying the functions of fewer than three features (Mika et al., 1999; Huberty and Olejnik, 2005). Compared with discriminant analysis, the machine learning algorithms employed in the present study could clearly rank the importance of 12 features simultaneously. In addition, the logistic regression model has advantages over discriminant analysis and Hotelling's T² test in not needing normally distributed variables (Hoffman, 2019).

Because the aim of the present study was to compare context-dependent call sequences employing machine learning methods, our results showed successful classification of call sequences of two distinct kinds of behaviour, distress and aggression. Actually, machine learning methods also could deal with tasks of varying complexity. For example, four supervised machine learning methods were conducted on barks of domestic dogs, which obtained high percentages of correct classifications on sex (85.13%), age (80.25%), individual (67.63%) and context (55.50%) (Larrañaga et al., 2015). Prat et al. (2016) employed the Gaussian mixture model-universal background model algorithm for vocalization classifications of different aggressive contexts and different emitters in Egyptian fruit bats and obtained high balanced accuracy (different aggressive contexts for each emitter: 75%, emitters: 71%). Although the above studies focused on the call classification but not extraction of syntactic structures, they suggested the potential ability of machine learning methods for call classification of similar contexts. But further study is still invited to test the appropriate algorithms and their performances in comparing the structures of call sequences when the context is less distinct.

An important factor to consider is the ease of implementing a given method. All the algorithms we adopted could be utilized easily and widely (Kotsiantis et al., 2007; Armitage and Ober, 2010). None need computing resources beyond an ordinary personal computer. Moreover, the features in our study were not species-specific and could be appropriate to other animal categories (Priyadarshani et al., 2018). The sound features were extracted based on syllables, which could be obtained from vocalizations of most animal species. In addition, the types of features integrated and estimated by the machine learning algorithms were variable, including not only features of sequence structure but also features of sound emitters such as individual label and gender. This demonstrates another advantage, that is, that machine learning can handle combinations of different parameters regardless of their units (Alice and Amanda, 2018). The proposed methods in our study were mainly suited for comparison and classification of call sequences under different conditions. Robust extraction and description of the structures of call sequences would require much improved processing.

The selective preference in the sound features concerning sequence structures suggested that greater horseshoe bats might order and arrange syllables according to certain rules, that is, syntactic structure in variable behavioural contexts. Using syntax composed of different structures of calls under specific situations may be a common phenomenon in many animal species. For example, syntax depending on social context was found in the ‘chick-a-dee’ calls of chickadees and in the songs of male mice (Clucas et al., 2004; Chabout et al., 2015). When confronted with conspecifics or predators, it was reasonable that greater horseshoe bats tended to have different reactions. Although a number of syllable types were found in all recording files, the sum of percentages for the five most frequently used types was greater than 80% in both the aggressive and distress contexts. This demonstrated that greater horseshoe bats in non-breeding periods could make good use of the order and arrangement of limited types instead of emitting complex composite syllables such as songs for mating (Davidson and Wilkinson, 2004; Bohn et al., 2009). Pioneering research in non-human primates and birds has indicated that animal signals can be functionally referential (Townsend et al., 2013; Scarantino and Clay, 2015). The sequence compositions of Titi monkeys (Callicebus nigrifrons) and the specificity note combinations of discrete alarm calls of Japanese great tits (Parus major minor) can both be used to communicate predator type (Cäsar et al., 2013; Suzuki, 2014).

In summary, our research reveals a tangible instance for employing machine learning methods to explore vocalization data. The results provided three useful and efficient models for analysing syntactic variation in bioacoustics. Using the power of machine learning, researchers can extract useful information from many vocalizations before they design behavioural experiments for further analysis. This study also demonstrated the presence of complex vocalization and potential syntactic structures of call sequences in non-breeding contexts of bats. Further experimentation, such as using playback of calls, is necessary to investigate the information encoded in different syntactic patterns.

We thank LetPub (www.letpub.com) for its linguistic assistance during the preparation of the manuscript.