No abstract

The performance of deep learning in natural language processing has been spectacular, but the reasons for this success remain unclear because of the inherent complexity of deep learning. This paper provides empirical evidence of its effectiveness and of a limitation of neural networks for language engineering. Precisely, we demonstrate that a neural language model based on long short-term memory (LSTM) effectively reproduces Zipf’s law and Heaps’ law, two representative statistical properties underlying natural language. We discuss the quality of reproducibility and the emergence of Zipf’s law and Heaps’ law as training progresses. We also point out that the neural language model has a limitation in reproducing long-range correlation, another statistical property of natural language. This understanding could provide a direction for improving the architectures of neural networks.

In this article, we evaluate computational models of natural language with respect to the universal statistical behaviors of natural language. Statistical mechanical analyses have revealed that natural language text is characterized by scaling properties, which quantify the global structure in the vocabulary population and the long memory of a text. We study whether five scaling properties (given by Zipf's law, Heaps' law, Ebeling's method, Taylor's law, and long-range correlation analysis) can serve for evaluation of computational models. Specifically, we test ngram language models, a probabilistic context-free grammar (PCFG), language models based on Simon/Pitman-Yor processes, neural language models, and generative adversarial networks (GANs) for text generation. Our analysis reveals that language models based on recurrent neural networks (RNNs) with a gating mechanism (i.e., long short-term memory, LSTM; a gated recurrent unit, GRU; and quasi-recurrent neural networks, QRNNs) are the only computational models that can reproduce the long memory behavior of natural language. Furthermore, through comparison with recently proposed model-based evaluation methods, we find that the exponent of Taylor's law is a good indicator of model quality. *

The entropy rate h of a natural language quantifies the complexity underlying the language. While recent studies have used computational approaches to estimate this rate, their results rely fundamentally on the performance of the language model used for prediction. On the other hand, in 1951, Shannon conducted a cognitive experiment to estimate the rate without the use of any such artifact. Shannon’s experiment, however, used only one subject, bringing into question the statistical validity of his value of h = 1.3 bits per character for the English language entropy rate. In this study, we conducted Shannon’s experiment on a much larger scale to reevaluate the entropy rate h via Amazon’s Mechanical Turk, a crowd-sourcing service. The online subjects recruited through Mechanical Turk were each asked to guess the succeeding character after being given the preceding characters until obtaining the correct answer. We collected 172,954 character predictions and analyzed these predictions with a bootstrap technique. The analysis suggests that a large number of character predictions per context length, perhaps as many as 10 3 , would be necessary to obtain a convergent estimate of the entropy rate, and if fewer predictions are used, the resulting h value may be underestimated. Our final entropy estimate was h ≈ 1.22 bits per character.

Neural language models have drawn a lot of attention for their strong ability to predict natural language text. In this paper, we estimate the entropy rate of natural language with state-of-the-art neural language models. To obtain the estimate, we consider the cross entropy, a measure of the prediction accuracy of neural language models, under the theoretically ideal conditions that they are trained with an infinitely large dataset and receive an infinitely long context for prediction. We empirically verify that the effects of the two parameters, the training data size and context length, on the cross entropy consistently obey a power-law decay with a positive constant for two different state-of-the-art neural language models with different language datasets. Based on the verification, we obtained 1.12 bits per character for English by extrapolating the two parameters to infinity. This result suggests that the upper bound of the entropy rate of natural language is potentially smaller than the previously reported values.