Building robust and modular question answering systems
Over the past few years, significant progress has been made in QA systems due to the availability of annotated datasets on a large scale and the impressive advancements in large-scale pre-trained language models. Despite these successes, the black-box nature of end-to-end trained QA systems makes them hard to interpret and control. When these systems encounter inputs that deviate from their training data distribution or are subjected to adversarial perturbations, their performance tends to deteriorate by a large margin. Furthermore, they may occasionally produce unanticipated results, potentially leading to confusion among users. Additionally, this deficiency in robustness and interpretability poses challenges when deploying such models in real-world scenarios.
In this dissertation, we aim to build robust QA systems by explicitly decomposing various QA tasks into distinct sub-modules, each responsible for a particular aspect of the overall QA process. Through this decomposition, we seek to achieve improved performance in terms of both the system's ability to handle diverse and challenging inputs (robustness) and its capacity to provide transparent and explainable reasoning (interpretability).
To address the aforementioned limitations, in this dissertation, we aim to build robust QA models by explicitly decomposing different QA tasks into different sub-modules. We argue that utilizing these sub-modules can substantially improve the robustness and interpretability of different QA systems. In the first half of this dissertation, we introduce three sub-modules to mitigate the dataset artifacts that models learn from datasets. These sub-modules also enable us to examine and exert explicit control over the intermediate outputs. In the first work, to address question answering that requires multi-hop reasoning, we propose a chain extractor, which extracts the reasoning chains necessary for models to derive the final answer. The reasoning chains not only prevent the model from exploiting reasoning shortcuts but also provide an explanation of how the answer is derived. In the second work, we incorporate an alignment layer between the question and the context before generating the answer. This alignment layer can help us interpret the models' behavior and improve the robustness of adversarial settings. In the third work, we add an answer verifier after QA models generate the answer. This verifier can boost QA models' prediction confidence across several different domains and help us spot cases where QA models predict the right answer for the wrong reason by utilizing the external NLI datasets and models.
In the second half of this dissertation, we tackle the problem of complex fact-checking in the real world by treating it as a modularized QA task. We first decompose a complex claim into several yes-no subquestions whose answer directly contributes to the veracity of the claim. Then, each sub-question is fed into a commercial search engine to retrieve relevant documents. Additionally, we extract the relevant snippets in the retrieved documents and use a GPT3-based summarizer to generate the core evidence for checking the claim. We show that the decompositions can play an important role in both evidence retrieval and veracity composition of an explainable fact-checking system. Also, we show the GPT3-based evidence summarizer generates faithful summaries of documents most of the time indicating it can be used as an effective part of the pipeline. Moreover, we annotate a dataset -- ClaimDecomp, containing 1,200 complex claims and the decompositions. We believe that this dataset can further promote building explainable fact-checking systems and analyzing complex claims in the real world.