使用類別資訊產生作為查詢詞概觀的網站片段敘述

Abstract

在這篇論文中, 我們探討一個如何改善搜尋結果頁面 所有結果底下的片段資訊以作為搜尋詞的整體概觀的問 題, 使使用者在點擊搜尋結果前可已對這個搜尋詞得到一 個大概的了解. 對於一個已知其類別的搜尋詞, 我們使用 了其類別所涵蓋的語意以及那些跟這類別息息相關的屬 性來組織這樣的片段資訊我們對類別從社群問答網站中 的問題抽取了那些跟類別息息相關的屬性, 並從那些問題 的答案中萃取了每個屬性的context information. 而我們 產生這些片段資訊主要依賴三個因素, 涵蓋搜尋詞的資訊 量, 涵蓋類別語意的程度, 以及涵蓋類別屬性的程度, 為了 能同時最佳化這三個因素, 我們採用了整數線性規劃來 模組化我們的問題, 實驗結果顯示我們產生出的片段資 訊, 在與傳統搜尋頁面以及一些基本的summarization 演 算法, 在表達搜尋詞概觀的程度上, 有不少的進步.

Previous work on snippet generation focuses mainly on how to produce one snippet for an individual search result. This paper aims to generate a comprehensive overview for an entity query in the search-result page. We assume each entity has its own category, whose attributes are regarded as the unique characteristics that the users might be interested in when searching for the entity. Given an entity as query (e.g., enterogastritis) and its category (e.g., disease), we want to organize the snippets that contain its attributes (e.g., symptoms and diagnoses) so that users can learn about the useful information with respect to the given query directly from the generated snippets without downloading documents. First, we extract the attributes of a category from a community-based question-answering (CQA) website. Next, the snippets are generated according to several factors, including how a sentence could be central to the meanings of the query, its category and corresponding attributes, and how well the snippets diversify the attributes. Finally, an Integer Linear Programming (ILP) is adopted to find an optimal sentence set as the snippet. The experiments are conducted on 100 common disease queries. Experimental results demonstrate the effectiveness and efficiency of the proposed approach, compared to an existing search engine and several summarization baselines.