"S1. I lived in Paris for many years. I, therefore, know many places in Paris"; "S2. Having lived in Paris for long, I know several places there". Can NLP techniques infer that S1 and S2 are similar semantically? If yes, which one (references)?
2 Answers
This is a problem that is very similar to the "Interpretable Semantic Textual Similarity" task of SemEval 2015 and 2016.
The task is to align two sentences and provide information which phrases of both sentences are related in which way with respect to semantic similarity.
For details see http://alt.qcri.org/semeval2016/task2/
-
Thanks. So from what you mentioned, we need to do the "chunking" and "alignment" manually? This is cumbersome task!!– SanjayOct 11, 2017 at 15:02
-
The competition featured two tracks: manual and automatic chunking. There are various chunkers available, the most famous is probably the one of the IXA pipeline. For you it might be best to train your own chunker with a standard CRF model on your own domain data. Alignment is done automatically and this was the objective of the competition. So depending on which track you are looking at, either nothing was done manually or only chunking.– peschüOct 12, 2017 at 16:48
You may look at deep learning approaches in general. This would be similar to what "reading comprehension tasks" are trying to achieve. Search for some papers and you shall find related ones, for example the paper "A Thorough Examination of the CNN/Daily Mail Reading Comprehension Task" at https://arxiv.org/abs/1606.02858.
There is also a paper list from a course I just had the last semester: http://www.sfs.uni-tuebingen.de/~ddekok/dl4nlp/ which you might find helpful.