Information Storage and Retrieval

Is that possible we can train a model on relevant feedback to know whether the users give the real feedback? I mean can we use this method to detect whether users give us a real feedback? I think giving feedback would distinguish on different culture people. For example, maybe U.S people would like to give feedback or not. I'm just wondering it is maybe a criteria to measure.

In the smoothing method, can you just skip the stop words? Why is the language model is hard to deal with relevance feedback? 

1. If we use the query likelihood model, how do we deal with the zero probability problem? 2. If we use the document likelihood model, Is not the same document length as the different contents and how do we deal with this problem?