Search

In this paper, we study language models based on recurrent neural networks on three databases in two languages. We implement basic recurrent neural networks (RNN) and refined RNNs with long short-term memory (LSTM) cells. We use the corpora of Penn Tree Bank (PTB) and AMI in English, and the Academia Sinica Balanced Corpus (ASBC) in Chinese. On ASBC, we investigate word-based and character-based language models. For character-based language models, we look into the cases where the inter-word space is treated or not treated as a token. In summary, we report and comment on the performance of RNN language models with different databases, network topology, language, and granularity.

It has been argued that recurrent neural network language models are better in capturing long-range dependency than n-gram language models. In this paper, we attempt to verify this claim by investigating the prediction accuracy and the perplexity of these language models as a function of word position, i.e., the position of a word in a sentence. It is expected that as word position increases, the advantage of using recurrent neural network language models over n-gram language models will become more and more evident. On the text corpus of Penn Tree Bank (PTB), a recurrent neural network language model outperforms a trigram language model in both perplexity and word prediction. However, on the AMI meeting corpus, a trigram outperforms a recurrent neural network language model.