@article{mifflin:semismooth,
  author={Yoav Artzi and  Luke Zettlemoyer},
  title={Learning to Recover Meaning from Unannotated Conversational Interactions},
  year={2011},
  annote = {

There has been significant work on supervised learning for inducing semantic parsers. There are approaches for learning semantic parsers from questions paired with database answers and taking advantage of natural language sentences paired with demonstrations of actions. In this paper, the authors describe a novel method to induce semantic parsers that are self-improving dialog systems that learn from their mistakes, with little or no human intervention. Every user’s response to a question is mapped to the meaning representation (MR). When a system fails to understand the meaning of the user’s utterance it requests clarification. The ability to automatically learn from this indirect feedback will enable the rapid development of simple self-improving dialog systems that learn through interaction and continually acquire better understanding of language.
One unique challenge for building these representations is that utterance meanings are often not independent from the rest of the interaction. In the paper the authors focused on learning to analyze the subset of context-independent sentences.A loss function is used that measures how a given MR (a) aligns with system understanding (b) fits the domain of the dialog. The learning had two major methods, lexical induction and parameter updates based on least loss. This approach significantly outperforms a baseline lacking the conversational and has results close to that from a supervised model. In future work, the authors aim to build a semantic parser that incorporates contextual information into the parsing process.
Written by: Rohun Tripati
}
}