A Generative Model for End-to-End Argument Mining with Reconstructed Positional Encoding and Constrained Pointer Mechanism

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10437 - 10449 December 7-11, 2022 ©2022 Association for Computational Linguistics A Generative Model for End-to-End Argument Mining with Reconstructed Positional Encoding and Constrained Pointer Mechanism Abstract Argument mining (AM) is a challenging task as it requires recognizing the complex argumen- tation structures involving multiple subtasks. To handle all subtasks of AM in an end-to-end fashion, previous works generally transform AM into a dependency parsing task. However, such methods largely require complex pre- and post-processing to realize the task transforma- tion. In this paper, we investigate the end- to-end AM task from a novel perspective by proposing a generative framework, in which the expected outputs of AM are framed as a simple target sequence. Then, we employ a pre- trained sequence-to-sequence language model with a constrained pointer mechanism (CPM) to model the clues for all the subtasks of AM in the light of the target sequence. Furthermore, we devise a reconstructed positional encoding (RPE) to alleviate the order biases induced by the autoregressive generation paradigm. Exper- imental results show that our proposed frame- work achieves new state-of-the-art performance on two AM benchmarks.1 1 Introduction As a fundamental task of computational argumen- tation, argument mining (AM) has drawn much re- search attention recently (Schaefer and Stede, 2021; Vecchi et al., 2021; Lawrence and Reed, 2019). The ultimate goal of AM is to analyze and understand argumentative text, so as to obtain structured ar- gumentation knowledge that can support a diverse range of downstream tasks, such as argument per- suasiveness prediction (Li et al., 2020; Huang et al., 2021), automated essay scoring (Ghosh et al., 2016; Nguyen and Litman, 2018; Song et al., 2020), ar- gument generation (Hua et al., 2019; Slonim et al., ∗Equal Contribution †Corresponding Authors 1Code is available at https://github.com/HITSZ-HLT/ GMAM. I believe cloning is beneficial . For example , cloned organs can help patients . Claim Support Premise [ 𝟑, 𝟓, 𝑪𝒍𝒂𝒊𝒎, 𝟏𝟎, 𝟏𝟒, 𝑷𝒓𝒆𝒎𝒊𝒔𝒆, 𝑺𝒖𝒑𝒑𝒐𝒓𝒕 ] Target Sequence: Argumentative Text: Figure 1: A simplified example of AM task. Two argument components are marked in green and blue, respectively, where the former is a Claim and the latter is a Premise. In addition, there is a Support relation from the Premise to the Claim. Our proposed target sequence corresponding to this example is shown at the bottom. 2021; Khatib et al., 2021), text summarization (Fab- bri et al., 2021; Bar-Haim et al., 2020), etc. Given a piece of argumentative text as input, an end-to-end AM system needs to identify both the argument components (ACs) and the argumenta- tive relations (ARs) between them. An example is shown in Figure 1. Specifically, AM generally com- prises four fine-grained subtasks (Eger et al., 2017): 1) component segmentation detects the boundaries of fine-grained argumentative segments, which are known as ACs; 2) component classification classi- fies the ACs into the categories defined by argumen- tation schemes; 3) relation detection determines whether there is an AR between two ACs; 4) rela- …

Content truncated to 500 words