Reading List - Question Classifcation

Question classification problem is briefly explained mainly based on [1]. Then analogy to CoreOnto is provided.

References:
[1] Concept-based Question Classification in Community Question Answering, http://files.me.com/kschoi/3r2c51
[2] S. Bergsma and Q. Qang, 2007. Learning noun phrase query segementation. In Proc. of EMNLP-CoNLL
[3] D. Blei, A. NG, and M. Jordan, 2003. Latent dirichlet allocation. In Journal of Machine Learning
[4] X. Phan, L. Ngyuyen, and S. Horiguchi. 2008. Learning to classify short and sparse text and web with hidden topics from large-scale data collections. In Prof. of WWW
[5] E.M. Voorhees. 2003. Overview of the TREC 2003 question answering track. In TExt REtrieval Conference.

For a given question, how to classify the question? One solution is described by exploiting the question structures beyond simple bag-of-words. Assume a question composed of a single concept summarizing the semantics of the question together with a bag of contextual words further specifying it. We develop a probabilistic method that achieves accurate question classficiation by identifying the conept in the question. Experiments using questions in Yahoo! Answers demonstrate that such concept-based approach outperforms competitors by10-30%.

mDLA [2] contributes to question classification when assuming each predefined class as a hidden topic, each concept as a virtual document, and each context (i.e., the remainder texts of each question containing the concept (=key phrase in question) as a bag-of-word paragraph. But modeling each concept as a mixture over classes, and each class as a mixture over words, the topic model can effectively capture these strucgtures in the data set.

Comments on [1]:

1. "class" and "category" should be the same term. Unify them into one.

2. "concept" in question is rather misleading. Consider "we assume a question is composed of a single concept summarizing the semantics of the question together with a bag of contextual words further specifying it." But in Table 1, "new york times" in "what was the first font used by the new york times" is hard to say that the term is to summarize the question. It is the object about what the question is. "New york times" is the object of the question, and the "first font" is the main summarized question about "New york times".

Analogy to CoreOnto: 무엇을 어노테이션 할 것인가의 문제, 그리고 왜 하느냐의 문제, 무엇에 기여할 것이가에 대한 문제, Wikipedia는 왜 중요한 것인가? DBPedia는 왜 중요한 것인가? Wikipedia에서 비슷한 항목간 비교가 쉽게 가능한 것인가? DBPedia로 정리가 되면 그 비슷한 항목이 비교가 되고 묶일 것인가? 주어진 문장의 중심 대상이 무엇이며, 그 중심 대상에 대하여 말하고자 하는 중심 질문은 무엇인가를 알 수 있도록 한다. 거꾸로 이야기하면, 이미 중심대상이 정하여져 있다면, 그 중심대상을 말하는 문장을 고를 수 있다는 것이다. CoreOnto의 목적이 어떤 중심대상에 대한 것만을 모아서 어노테이션을 하고 온톨로지에 매핑을 하거나 만드는 것이라면, 선택적 어노테이션에 대한 기준이 될 수가 있다. 또 다른 측면을 본다면, 각각의 문장에 대하여 중심 대상이 선정이 된다면, 관련 온톨로지나 factoid를 업데이트하는데에도 쓰인다.