Self-validated Story Segmentation of Chinese Broadcast News.

Automatic story segmentation is an important prerequisite for semantic-level applications. The normalized cuts (NCuts) method has recently shown great promise for segmenting English spoken lectures. However, the availability assumption of the exact story number per file significantly limits its capability to handle a large number of transcripts. Besides, how to apply such method to Chinese language in the presence of speech recognition errors is unclear yet. Addressesing these two problems, we propose a self-validated NCuts (SNCuts) algorithm for segmenting Chinese broadcast news via inaccurate lexical cues, generated by the Chinese large vocabulary continuous speech recognizer (LVCSR). Due to the specialty of Chinese language, we present a subword-level graph embedding for the erroneous LVCSR transcripts. We regularize the NCuts criterion by a general exponential prior of story numbers, respecting the principle of Occam’s razor. Given the maximum story number as a general parameter, we can automatically obtain reasonable segmentations for a large number of news transcripts, with the story numbers automatically determined for each file, and with comparable complexity to alternative non-self-validated methods. Extensive experiments on benchmark corpus show that: (i) the proposed SNCuts algorithm can efficiently produce comparable or even better segmentation quality, as compared to other state-of-the-art methods with true story number as an input parameter; and (ii) the subword-level embedding always helps to recovering lexical cohesion in Chinese erroneous transcripts, thus improving both segmentation accuracy and robustness to LVCSR errors.