Study On Structure And Function Of Syntax In Social Calls Of Greater Horseshoe Bats (Rhinolophus Ferrumequinum) Employing Machine Learning Methods

A profitable approach to gain insights into the origin and evolution of language is to find and use appropriate analogies of these capabilities in other animals for comparative studies.Syntax is the set of rules for combining words into phrases,providing the basis for the generative power of linguistic expressions.A key step in understanding the evolution of human language involves unravelling the origins of language's syntactic structure.In so many kinds of contexts which were very vertical for animals' life-story,aggressive and distress contexts play a major role in resource competition,creation of dominance hierarchy and survival.However,it is largely unclear that what the structure and function of social call syntax in these contexts,what's more,the proper methods to explore big data of animal social call is urgently needed nowadays.Therefore,the first aim of the present study is to investigate the application of multiple machine learning methods in analysing the syntax in social calls of greater horseshoe bats(Rhinolophus ferrumequinum).And then with the help of machine learning methods and playback experiments,the next purpose is to study the structure and function of social call syntax under aggressive and distress contexts.Firstly,to deal with the problem of overlapping calls in animal acoustic research,this study constructed bi-direction long-short memory network(BLSTMs)based on the technology of speech separation in deep neural networks.The BLSTMs could be used to separate the overlapping echolocation-communication calls of bats.The echolocation pulsed and communication calls of three constant frequency(CF)bats,including greater horseshoe bats,great Himalayan leaf-nosed bats(Hipposideros armiger),least horseshoe bats(Rhinolophus pusillus)and three frequency modulation(FM)bats,including Asian particoloured bats(Vespertilio sinensis),great evening bats(Ia io)and big?footed myotis(Myotis macrodactylus)were used to test the generalization of the BLSTMs.The acoustic parameters of original recorded and separated pulses/syllables were extracted and compared to test the model performance.The results showed that the model could separate the echolocation pulses and communication calls without changing their qualities.What's more,clustering analysis was conducted using the separated echolocation pulses of six bats species outputted the high corrected rand index(82.79%),which indicated that the separated calls could be used in the acoustic analysis.Secondly,an automatic call sequence structure analysis system was developed to analyse the aggressive calls of greater horseshoe bats effectively.The system was built with PHP,Java Script,HTML,CSS,and My SQL database.The results showed that there were selective preferences in syllable types,transitions between two adjacent syllables,syllable types that occurred in each position of a sequence,and sequence types in aggressive call sequences.The most probable syllable type was NB-SFM,the most probable transition was NB-DFM/NB-DFM and the most probable syllable types that occurred in the first three positions of a sequence were NB-SFM.Besides,the most probable sequence type was NB-SFM/NB-SFM/NB-SFM.Thirdly,the distress call of greater horseshoe bats was recorded and compared with aggressive calls applying three machine learning models: logistic regression,support vector machine and decision trees.Then the contributions of the features used for comparison were analysed using a random forest model and important features were investigated to compare the differences in syntax structure of aggressive and distress contexts.The results showed that all three machine learning models produced high accuracy(more than 95%)when classifying the call sequences of two different contexts.Moreover,the important features were relative to sequence structure instead of emitters or gender,which indicated a high discrepancy in syntax between aggressive and distress contexts.Furthermore,the results of playback experiments suggested that the behaviour response of greater horseshoe bats to the most probable sequence types of two contexts were significantly different.Fourthly,based on the previous study of this paper,different types of call sequences were played back to greater horseshoe bats and then the response behaviours of bats were analysed by Deep Lab Cut system to investigate the function of syntax structures.The results revealed that the moving speed and total distance of the right forearm of greater horseshoe bats responding to sequences containing different syllable types were significantly different.When sequences composed of syllables in opposite order were played back to bats,the moving speed and total distance of bats' left ear were different.In the last playback experiments,the moving speed and total distance of bats' forearms,ears,mouth,head,and numbers of echolocation pulsed responding to call sequences having different lengths(or repetitions)were significantly different.Finally,the present study suggested that(1)machine learning methods could be effectively used in animal acoustic researches,such as overlapping call separation,call sequence classification and quantifying behaviour;(2)the syntax existed in the call sequences in the aggressive context of greater horseshoe bats and the structures of syntax were different between aggressive context and distress context;(3)the syllables in aggressive calls of greater horseshoe bats have difference meanings and the sequences of difference structure also have discrepancy meanings,which indicate that the syntax in social calls of greater horseshoe bats was consisted with compositional syntax.Here,multiple machine learning techniques and playback experiments were adopted to investigate the structure and function of syntax in social calls of greater horseshoe bats,this study provided effective analysing tools for big scaled sequence data of animal acoustics and given insights for applying machine learning methods in animal acoustic research.Besides,the results offered important fundamental data for further studies on understanding the mechanisms of animal communication and the evolution of communication signals,which had contributions to reveal the origin and evolution of human language.