Learning on Structured Documents for Conditional Question Answering

Learning on Structured Documents for Conditional Question Answering Zihan Wang, Hongjin Qian, Zhicheng Dou∗ Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China Engineering Research Center of Next-Generation Intelligent Search and Recommendation, MOE {wangzihan0527, ian, dou}@ruc.edu.cn Abstract Conditional question answering (CQA) is an important task in natural language processing that involves answering questions that depend on specific conditions. CQA is crucial for domains that require the provision of personalized advice or making context-dependent analyses, such as legal consulting and medical diagnosis. However, existing CQA models struggle with generating multiple conditional answers due to two main challenges: (1) the lack of supervised training data with diverse conditions and corresponding answers, and (2) the difficulty to output in a complex format that involves multiple conditions and answers. To address the challenge of limited super- vision, we propose LSD (Learning on Structured Documents), a self-supervised learning method on structured documents for CQA. LSD involves a conditional problem generation method and a contrastive learning objective. The model is trained with LSD on massive unlabeled structured documents and is fine-tuned on labeled CQA dataset afterwards. To overcome the limitation of outputting answers with complex formats in CQA, we propose a pipeline that enables the gen- eration of multiple answers and conditions. Experimental results on the ConditionalQA dataset demonstrate that LSD outperforms previous CQA models in terms of accuracy both in providing answers and conditions. 1 Introduction Recently, question answering (QA) has gained increasing interest in the field of Natural Language Pro- cessing. Various types of question answering tasks, such as knowledge-based QA (Cui et al., 2017), open domain QA (Kwiatkowski et al., 2019), and multi-hop QA (Yang et al., 2018), have been exten- sively studied. Among them, conditional question answering (CQA) (Sun et al., 2022a) is becoming increasingly important in various contexts, such as medical diagnosis, legal consultation, financial analy- sis, and more. Unlike the traditional question answering problem that only accepts a question and returns an answer, CQA involves understanding a complex and lengthy document, finding all possible answers under different conditions, and determining under what condition the answer is applicable. Figure 1 shows an example for CQA, where the answer could be different when the questioner is under different conditions. The CQA task includes providing potential answers “yes” and “no” and their corresponding conditions based on the given question and scenario. Previous studies on CQA can be broadly categorized into two groups: extractive and generative meth- ods. Extractive methods (Ainslie et al., 2020) (Sun et al., 2021) extract the most relevant span from a document as answers and conditions. In contrast, generative methods (Izacard and Grave, 2021) (Sun et al., 2022b) use a generative model to generate answers along with their corresponding conditions di- rectly. However, current CQA models face two common challenges. Firstly, the supervised data for CQA is limited and expensive to obtain. Unlike traditional QA datasets, CQA requires specific annota- tions that include scenarios, answers, and conditions, making the data collection process more extensive and time-consuming. Secondly, current …

Content truncated to 500 words