Building neural networks to query a knowledge base (a table) with natural language is an emerging research topic in NLP. The neural enquirer typically necessitates multiple steps of execution because of the compositionality of queries. In previous studies, researchers have developed either distributed enquirers or symbolic ones for table querying. The distributed enquirer is end-to-end learnable, but is weak in terms of execution efficiency and explicit interpretability. The symbolic enqurier, on the contrary, is efficient during execution; but it is very difficult to train especially at initial stages. In this paper, we propose to couple distributed and symbolic execution for natural language queries. The observation is that a fully distributed executor also exhibits meaningful, albeit imperfect, interpretation. We can thus pretrain the symbolic executor with the distributed one's intermediate execution results in a step-by-step fashion. Experiments show that our approach significantly outperforms either the distributed or symbolic executor; moreover, we have recovered more than 80% execution sequences with only groundtruth denotations during training. In summary, the coupled neural enquirer takes advantages of both distributed and symbolic executors, and has high performance, high learning efficiency, high execution efficiency, and high interpretability.
from cs.AI updates on arXiv.org http://ift.tt/2ggxMx4
via IFTTT
No comments:
Post a Comment