@InProceedings{kim-mooney-13_discriminative-reranking-for-grounded-language-learning,
  author    = {Kim, Joohyun  and  Mooney, Raymond},
  title     = {Adapting Discriminative Reranking to Grounded Language Learning},
  booktitle = {Proc. 51st ACL (Long papers)},
  month     = {August},
  year      = {2013},
  pages     = {218--227},
  url       = {http://www.aclweb.org/anthology/P13-1022}
  annote    = {
  Authors have applied the technique of adaptive discriminative reranking for the task of learning language using the context in which its used.Particularly the technique is used for  interpreting navigational instructions in NL.
They used the fact that the task of grounded language learning is equivalent to inducing PCFG  from corpora.
  The task of navigation is first described formaly as given,   ( ei, ai, wi) from the training instances  ,they wanted to learn to generate  aj  for  test instances ( ej, wj)    
where e is Sentence in NL ,a is action to be taken from  NL and w is the current world state.The grammer generated for navigational task is very large so they use previously learned semantic lexicon for reducing the size of grammer.
Since the sequence of actions ai provides a very weak supervision for learning language because same task can be done in numerous ways so they use response based feedback for learning.example:
whether or not task completed successfully.Also this approach was adopted since they dont have the "gold-standard" reference parses for this navigational task for training .The reranking techniques applied else where uses the parse trees.
Methodology:
1.Trained the parser using generative model for generating candidate parse trees.
2.Now the discriminative model is trained for rescoring the candidate parses using global features.
3.For reranking they have used the averaged perceptron method.Some novel ideas used:
Since they dont have the gspt so they use psuedo gspt whose derived meaningful representation  guides you to the goal with highest success. For reranking , they  have modified the average perceptron algorithm
instead of using single parse , they use multiple parses for updating the  W as other parses with execution rate greater than the  current preferred parse as others also contain useful information
For reranking the parses generated from baseline generative model certain features are used by the discriminative model like:
1.Base features
2.Predicate-only Features
3.Descend Action features.
Results:
1.Paragraph level accuracies is low compared to sentence level .
2.The execution accuracy of response based feedback is greater than gold-standard based ,for parsing accuracy situation is reversed.But the response based feedback outperforms it on final task. 
3.Using multple reference parse trees gives best results in all task.( like execution )
4.Also they showed that descend action features are most effective for weight updation  but taking combination of all features gives best results.
Note: The navigational instructions corpus is taken from MacMohan et al(2006).
}}