Reinforcement Learning Based Graph-to-Sequence Model for Natural Question Generation

Abstract

Natural question generation (QG) is a challenging yet rewarding task, that aims to generate questions given an input passage and a target answer. Previous works on QG, however, either (i) ignore the rich structure information hidden in the word sequence, (ii) fail to fully exploit the target answer, or (iii) solely rely on cross-entropy loss that leads to issues like exposure bias and evaluation discrepancy between training and testing. To address the above limitations, in this paper, we propose a reinforcement learning (RL) based graphto-sequence (Graph2Seq) architecture for the QG task. Our model consists of a Graph2Seq generator where a novel bidirectional graph neural network (GNN) based encoder is applied to embed the input passage incorporating the answer information via a simple yet effective Deep Alignment Network, and an evaluator where a mixed objective function combining both cross-entropy loss and RL loss is designed for ensuring the generation of semantically and syntactically valid text. The proposed model is end-to-end trainable, and achieves new state-of-the-art scores and outperforms all previous methods by a great margin on the SQuAD benchmark.

GraphSearchNet: Enhancing GNNs via Capturing Global Dependencies for Semantic Code Search

“…Recently, GNNs (Hamilton, Ying, and Leskovec 2017;Li et al 2015;Kipf and Welling 2016;Xu et al 2018) have become a hot research topic since their strengths in learning structure data. Various applications in different domains such as chemistry biology (Duvenaud et al 2015), computer vision (Norcliffe-Brown, Vafeias, andParisot 2018), natural language processing (Xu et al 2018;Chen, Wu, and Zaki 2019b) have demonstrated the effectiveness of GNNs. In program scenario, compared with the early works to represent programs with abstract syntax tree (Alon et al 2018(Alon et al , 2019Liu et al 2020b), more works have already attempted to use graphs (Allamanis, Brockschmidt, and Khademi 2017) to learn the semantics for various applications, e.g., source code summarization (Liu et al 2020a;Fernandes, Allamanis, and Brockschmidt 2018), vulnerability detection (Zhou et al 2019), type inference (Allamanis et al 2020).…”

“…In addition, many existing graph-based works (Zhou et al 2019;Liu et al 2020a;Fernandes, Allamanis, and Brockschmidt 2018) in program learning have proved the effectiveness of GNNs (Kipf and Welling 2016;Allamanis, Brockschmidt, and Khademi 2017) on capturing program semantics. Furthermore, due to the powerful relation learning capacity of GNNs, they have also been widely used in many NLP reasoning applications, e.g., natural question generation (QG) (Chen, Wu, and Zaki 2019b;Su et al 2020), conversational machine comprehension (MC) (Chen, Wu, and Zaki 2019a;Song et al 2018;De Cao, Aziz, and Titov 2018). Annervaz et al (Annervaz, Chowdhury, and Dukkipati 2018) further proved that augmenting graph information with LSTM can improve the performance of many NLP applications.…”

Iterative Deep Graph Learning for Graph Neural Networks: Better and Robust Node Embeddings

“…In the field of GNNs [27,15,17,33,58], there is a line of research on developing robust GNNs that are invulnerable to adversarial graphs by leveraging attention-based methods [5], Bayesian methods [13,59] and graph diffusion-based methods [29]. Recently, researchers have explored methods to automatically construct a graph of objects [42,8,32,14] or words [37,6,7] when applying GNNs to non-graph structured data. However, these methods merely optimize the graphs towards the downstream tasks without the explicit control of the quality of the learned graphs.…”

“…Another potential reason is that real-world graphs are often noisy or even incomplete due to the inevitably error-prone data measurement or collection. Furthermore, many applications such as those in natural language processing [7,53] may only have sequential data or even just the original feature matrix, requiring additional graph construction from the original data matrix.To address these limitations, we propose an end-to-end graph learning framework, namely Iterative Deep Graph Learning (IDGL), for jointly and iteratively learning the graph structure and the GNN parameters that are optimized towards the downstream prediction task. The key rationale of our ˚Corresponding author.Preprint.…”

“…Another potential reason is that real-world graphs are often noisy or even incomplete due to the inevitably error-prone data measurement or collection. Furthermore, many applications such as those in natural language processing [7,53] may only have sequential data or even just the original feature matrix, requiring additional graph construction from the original data matrix.…”