3).WITH A RELATIVE HIGH TRANSLATION PROBABILITY.IN THIS PAPER, WE AR...

3.3).

with a relative high translation probability.

In this paper, we argue that it is beneficial to cap-

• finally, we conduct the experiments on

ture contextual information for question retrieval.

community-based Q&A data for question re-

To this end, inspired by the phrase-based statistical

trieval. The results show that our proposed ap-

machine translation (SMT) systems (Koehn et al.,

proach significantly outperforms the baseline

2003; Och and Ney, 2004), we propose a phrase-

methods (in Section 4).

based translation model (P-Trans) for question re-

trieval, and we assume that question retrieval should

The remainder of this paper is organized as fol-

be performed at the phrase level. This model learns

lows. Section 2 introduces the existing state-of-the-

the probability of translating one sequence of words

art methods. Section 3 describes our phrase-based

(e.g., phrase) into another sequence of words, e.g.,

translation model for question retrieval. Section 4

translating a phrase in a historical question into an-

presents the experimental results. In Section 5, we

other phrase in a queried question. Compared to the

conclude with ideas for future research.

traditional word-based translation models that ac-

count for translating single words in isolation, the

2 Preliminaries

phrase-based translation model is potentially more

effective because it captures some contextual infor-