Due to its many interrelated components, the TDT3 Text Corpus requires additional quality checking to verify that the dependencies between the components hold. There are two types of TDT3 sources, broadcast and newswire, in other words, audio and text. The audio files created from the broadcast sources make up the TDT3 Audio Corpus, but text information is also derived from this audio data. Therefore the TDT3 Text Corpus is derived from both broadcast and newswire. Newswire data, along with transcripts of the audio data, are tagged with SGML. Each of these SGML files is used to create a token file, which, although still tagged with SGML, is basically one word per line and has no sentence or paragraph structure. Similar text token files are created from the audio files using speech recognition. TDT3 contains both English and Mandarin, and Mandarin token files, whether created from the transcripts or the audio, are translated using machine translation software. And finally, because each source file may contain several stories, all the different token files have corresponding boundary table files, which indicate which tokens make up each story.