Structural bias in inducing representations for probabilistic natural language parsing

We present a neural network based natural language parser. Training the neural network induces hidden representations of unbounded partial parse histories, which are used to estimate probabilities for parser decisions. This induction process is given domain-specific biases by matching the flow of information in the network to structural locality in the parse tree, without imposing any independence assumptions. The parser achieves performance on the benchmark datasets which is roughly equivalent to the best current parsers.