IMPROVISING HIP HOP LYRICS VIA TRANSDUCTION GRAMMAR INDUCTION

PhD Thesis Proposal Defence


Title: "IMPROVISING HIP HOP LYRICS VIA TRANSDUCTION GRAMMAR INDUCTION"

by

Mr. Karteek ADDANKI


Abstract:

Among the many genres of language that have been studied in computational 
linguistics and spoken language processing, there has been a dearth of work on 
lyrics in music, despite the major impact that this form of language has across 
almost all human cultures. In this proposal, we propose theoretically motivated 
symbolic and deep learning models for improvising lyrics in music and we choose 
the genre of hip hop lyrics as our domain. Unlike most other approaches, all 
our models are completely unsupervised and do not make use of any linguistic or 
phonetic information. Through our work, we model the issues in song lyric 
improvisation using modern statistical language technologies and attempt to 
bridge the gap between language and music from a computational perspective.

We improvise hip hop lyrics by generating responses to challenges similar to a 
freestyle rap battle, modeling it as a ''transduction'' problem where the 
challenges need to be translated into a response. We propose a novel hidden 
Markov Model (HMM) based rhyme scheme detection module, which identifies the 
rhyme scheme within a stanza in a completely unsupervised fashion and use it to 
select training data for our models so as to generate fluent and rhyming 
responses. We choose the framework of transduction grammars, in particular 
inversion transduction grammars (ITGs), as our transduction model given their 
representational capacity and empirical performance across a spectrum of NLP 
tasks. We propose two symbolic models for improvisation based on 1) a token 
based bracketing inversion transduction grammar, and 2) interpolated grammar 
induced using bottom-up token based rule induction and top-down rule 
segmentation strategies. We demonstrate that the interpolated grammar generates 
responses that are more fluent and rhyme better with the challenges according 
to human evaluators. We also compare the performance of our models against the 
widely used off-the-shelf phrase base SMT (PBSMT) model upon the same task and 
show that both our models outperform the PBSMT baseline. We also present 
similar results on Maghrebi French hip hop lyrics demonstrating the language 
independence of our models.

We also present a novel deep learning improvisation model based on a fully 
bilingual generalization of the monolingual Recursive Auto-Associative Memory 
(RAAM) known as Transduction Recursive Auto-Associative Memory (TRAAM). TRAAM 
models learn soft, context-sensitive generalizations over the structural 
relationships between associated parts of challenge and response raps, while 
avoiding the exponential complexity costs that symbolic models would require. 
In TRAAM, feature vectors are learned simultaneously using context from both 
the challenge and the response, such that challenge-response association 
patterns with similar structure tend to have similar vectors. We argue that the 
TRAAM models capture the context-sensitive generalizations that symbolic models 
fail to capture and will therefore produce better quality responses.

Lastly, we discuss the challenges in evaluating the performance on the 
improvisation task of evaluating hip hop lyrics as a first step toward 
designing robust evaluation strategies for improvisation tasks, a relatively 
neglected area to date. We discuss our observations regarding inter-evaluator 
agreement on judging improvisation quality as a means to better understand the 
high degree of subjectivity at play in improvisation tasks, thereby enabling 
the design of more discriminative evaluation strategies to drive future model 
development.


Date:			Tuesday, 16 June 2015

Time:                   2:00pm - 4:00pm

Venue:                  Room 3584
                         lifts 27/28

Committee Members:	Prof. Dekai Wu (Supervisor)
  			Prof. Qiang Yang (Chairperson)
 			Prof. Dit-Yan Yeung
  			Prof. Pascale Fung (ECE)


**** ALL are Welcome ****