基於半監督多任務學習與重排序的立場新聞檢索

Abstract

新聞立場檢索是從搜索中的到和搜索輸入具有相同立場的新聞，若能快速得到各種特定立場的新聞，有助於人們快速理解不同立場新聞下的價值觀，也有助於對社會輿論的長期趨勢進行分析，有相當的價值，而現有大多數搜索引擎的返回內容需要人工去分析判斷新聞的立場。對這類問題，本文提出了半監督多任務學習方法和重排序方法。半監督多任務學習方法在不引入標記資料的情況下，利用新聞結構信息與多任務學習提升了模型的性能。重排序挖掘排序對象之間的關係，不需要新的標記資料，不需對特定的任務的專家理解，得到更精準的排序結果，並可普適于其它排序任務。

Stance news retrieval aims to obtain news, which is related and having the same stance with query, from huge amounts of news. Retrieving news with specific stance can be beneficial, which helps to understand values from different stance, and also helps to analyze the long-term trend of public opinions. We introduce Semi-supervised multi-task learning for stance classification and an re-ranking method for news ranking. The semi-supervised multi-task Learning, a transfer learning method which leverages the structure information in news, significantly outperform the base model without new labeled data. The re-ranking method leverages the relationship between the ranking items, it does not require any human knowledge or any labeled data, improves the ranking performance and is applicable in other ranking task.