Research on Knowledge Elements in Exponential Language Model

This paper presents an exponential language model (ELM) for modeling and managing knowledge elements. The model has been developed based on Minimum Sample Risk (MSR) algorithm, which is a discriminative training method. ELM uses features to capture global, domain, or sentential language phenomena that is composed of name entities, part of speech strings, personal usage words, positions of words, sentence mood, sentence tense etc. We study kinds of knowledge elements' performances on the task of Chinese Pinyin to Chinese character (PTC) conversion in Internet language (Chinese mobile short messages and Chinese QQ' chat records). By combining different kind of knowledge elements to ELM, the model performs different, but all ELMs with more knowledge elements outperform the ELM only using probability knowledge calculated by baseline n-gram models which use Kneser-Ney smoothing technology.