The effect of word similarity on N-gram language models in Northern and Southern Dutch

In this paper we examine several combinations of classical N-gram language models with more advanced and well known techniques based on word similarity such as cache models and Latent Semantic Analysis. We compare the efficiency of these combined models to a model that combines N-grams with the recently proposed, state-of-the-art neural network-based continuous skip-gram. We discuss the strengths and weaknesses of each of these models, based on their predictive power of the Dutch language and find that a linear interpolation of a 3-gram, a cache model and a continuous skip-gram is capable of reducing perplexity by up to 18.63%, compared to a 3-gram baseline. This is three times the reduction achieved with a 5-gram.In addition, we investigate whether and in what way the effect of Southern Dutch training material on these combined models differs when evaluated on Northern and Southern Dutch material. Experiments on Dutch newspaper and magazine material suggest that N-grams are mostly influenced by the register and not so much by the language (variety) of the training material. Word similarity models on the other hand seem to perform best when they are trained on material in the same language (variety).In this paper we examine several combinations of classical N-gram language models with more advanced and well known techniques based on word similarity such as cache models and Latent Semantic Analysis. We compare the efficiency of these combined models to a model that combines N-grams with the recently proposed, state-of-the-art neural network-based continuous skip-gram. We discuss the strengths and weaknesses of each of these models, based on their predictive power of the Dutch language and find that a linear interpolation of a 3-gram, a cache model and a continuous skip-gram is capable of reducing perplexity by up to 18.63%, compared to a 3-gram baseline. This is three times the reduction achieved with a 5-gram.In addition, we investigate whether and in what way the effect of Southern Dutch training material on these combined models differs when evaluated on Northern and Southern Dutch material. Experiments on Dutch newspaper and magazine material suggest that N-grams are mostly influenced by the register and not so much by the language (variety) of the training material. Word similarity models on the other hand seem to perform best when they are trained on material in the same language (variety).A