LaBB-CAT: an Annotation Store

LaBB-CAT: an Annotation Store Robert Fromont Jennifer Hay NZILBB, University of Canterbury NZILBB, University of Canterbury Private Bag 4800 Private Bag 4800 Christchurch, New Zealand Christchurch, New Zealand robert.fromont@canterbury.ac.nz jen.hay@canterbury.ac.nz Abstract “ONZE Miner”, an open-source tool for storing and automatically annotating Transcriber transcripts, has been redeveloped to use “annotation graphs” as its data model. The annotation graph framework provides the new software, “LaBB-CAT”, greater flexibility for automatic and manual annotation of corpus data at various independent levels of granularity, and allows more sophisticated annotation structures, opening up new possibilities for corpus mining and conversion between tool formats. 1 Introduction “ONZE Miner” (Fromont & Hay 2008) was a browser-based, searchable database tool for time- aligned transcripts of speech produced using Transcriber, a transcription and annotation tool developed by Barras et al. (2000). It has been used for a variety of research projects in various universities and labs, primarily for sociophonetic research. ONZE Miner's original data structure was designed to closely mirror that of Transcriber, so transcripts are divided into topic-tagged sections, which contain speaker turns, divided up into utterance lines containing text and other 'event' annotations such as noises, comments, etc. In order to allow automatic annotation of lexical data from CELEX (Baayen et al., 1995), and to facilitate storage for forced-alignments produced by the Hidden Markov Model Toolkit, HTK (Young et al. 2006), lines were tokenized into words that were stored as separately annotatable units, which could be further divided into segments for storage of phones produced by HTK. For researchers with large collections of recordings and Transcriber transcripts, ONZE Miner was very useful for mining corpus data, but it had certain limitations related to its data structures, which are explained below. 1.1 Other Formats Many corpora exist, or are being produced, using tools other than Transcriber. For example the Buckeye corpus (Pitt et al. 2007) includes aligned transcription files in the Xwaves (Hawkins 2008) format, and transcriptions for various other corpora are available as Praat (Boersma & Weenink 2005) textgrids. For ONZE Miner, support for these was only available via conversion from these formats to Transcriber files before importing the data. The problem was that, in many cases, the data was not structured in a way that was compatible with the Transcriber model. For example, some formats include much finer-grained synchronisation than is typically available with Transcriber. Simultaneous speech also presented problems for data conversion. In Transcriber, overlapping speech is modelled using a 'simultaneous speech' turn – i.e. a single turn that has multiple speakers attached to it, and multiple corresponding transcriptions. For example, if a second speaker started their turn before a first speaker finished theirs, this would be modelled as three turns: 1. a turn containing words spoken while only the first speaker is speaking, 2. a 'simultaneous speech' turn containing words spoken by both speakers, during the time that they are both speaking, and 3. a turn containing words spoken by the only second speaker, once the first speaker is no longer talking. Robert Fromont and …

Content truncated to 500 words