We take an end-to-end, sequence-to-sequence learning approach to the task of following natural language route instructions, i.e., mapping natural language instructions to action sequences. Our model is an alignment-based long short-term memory recurrent neural network (attention-based LSTM-RNN) that encodes the free-form navigational instruction sentence and the corresponding representation of the environment state. We integrate alignment as part of our network, thereby empowering the model to focus on important sentence "regions." The alignment-based LSTM then decodes the learned representation to obtain the inferred action sequence. Adding bidirectionality to the network helps further. In contrast with existing methods, our model uses no additional information or resources about the task or language at all (e.g., parsers or seed lexicons) and still achieves state-of-the-art on a single-sentence benchmark dataset and strong results in the limited-training multi-sentence setting. Moreover, our model is more stable across runs than previous work. We evaluate our model through a series of ablation studies that elucidate the contributions of the primary components of our model.
from cs.AI updates on arXiv.org http://ift.tt/1JSbxbL
via IFTTT
No comments:
Post a Comment