Towards a multimodal knowledge base for Indian art music: A case study with melodic intonation

This thesis is a result of our research efforts in building a multi-modal knowledge-base for the specific case of Carnatic music. Besides making use of metadata and symbolic notations, we process natural language text and audio data to extract culturally relevant and musically meaningful information and structuring it with formal knowledge representations. This process broadly consists of two parts. In the first part, we analyze the audio recordings for intonation description of pitches used in the performances. We conduct a thorough survey and evaluation of the previously proposed pitch distribution based approaches on a common dataset, outlining their merits and limitations. We propose a new data model to describe pitches to overcome the shortcomings identified. This expands the perspective of the note model in-vogue to cater to the conceptualization of melodic space in Carnatic music. We put forward three different approaches to retrieve compact description of pitches used in a given recording employing our data model. We qualitatively evaluate our approaches comparing the representations of pitched obtained from our approach with those from a manually labeled dataset, showing that our data model and approaches have resulted in representations that are very similar to the latter. Further, in a raaga classification task on the largest Carnatic music dataset so far, two of our approaches are shown to outperform the state-of-the-art by a statistically significant margin.

In the second part, we develop knowledge representations for various concepts in Carnatic music, with a particular emphasis on the melodic framework. We discuss the limitations of the current semantic web technologies in expressing the order in sequential data that curtails the application of logical inference. We present our use of rule languages to overcome this limitation to a certain extent. We then use open information extraction systems to retrieve concepts, entities and their relationships from natural language text concerning Carnatic music. We evaluate these systems using the concepts and relations from knowledge representations we have developed, and groundtruth curated using Wikipedia data. Thematic domains like Carnatic music have limited volume of data available online. Considering that these systems are built for web-scale data where repetitions are taken advantage of, we compare their performances qualitatively and quantitatively, emphasizing characteristics desired for cases such as this. The retrieved concepts and entities are mapped to those in the metadata. In the final step, using the knowledge representations developed, we publish and integrate the information obtained from different modalities to a knowledge-base. On this resource, we demonstrate how linking information from different modalities allows us to deduce conclusions which otherwise would not have been possible.