The main tasks in many applications can be formalized as matching between heterogeneous objects, including search, recommendation, question answering, paraphrasing, and image retrieval. For example, search can be viewed as a problem of matching between a query and a document, and image retrieval can be viewed as a problem of matching between a text query and an image.
A variety of machine learning techniques have been developed for various matching tasks. We refer to them as `learning to match’. This article gives the definition, the importance and open problems of learning to match. It also introduces our work on learning to match conducted at Noah's Ark Lab.
This is a brief introduction to string re-writing kernel (SRK) proposed by Bu, Li, & Zhu in 2012. SRK measures the similarity between two re-writings of strings. SRK can capture the lexical and structural similarity between two pairs of sentences for paraphrasing, question answering, and short text conversation.
It might be intractable to compute a generic SRK; Bu et al. further propose a sub-class of SRK, called kb-SRK, which can be computed efficiently. Experimental results show that the use of kb-SRK can achieve results comparable with the state-of-the-art methods on paraphrase identification and recognizing textual entailment.
Their paper received the best student paper award at ACL 2012.
The article can be found here PDF.