Aspect-Based Argument Mining (ABAM) is a critical
task in computational argumentation that aims extract the
argument units expressing support or opposing stances and the
aspect terms mentioned within these argument units, from a
given text discussing a controversial topic. However, existing approaches
have primarily treated ABAM as a nested named entity
recognition problem, overlooking the need for tailored strategies
to effectively address the specific challenges and requirements of
ABAM tasks. In this paper, we propose a layer-based Hierarchical
Reinforcement Framework (HRF) for ABAM, which consists
of four modules: basic module, argument unit enhancement
module, aspect term enhancement module and decision module.
In particular, within the basic module, we propose a Semantic and
Syntactic Fusion (SSF) component, which uses graph attention
network to compensate for the limitation of long short term
memory network in forgetting long distance information, thereby
constructing a robust underlying representation. Furthermore,
the argument unit enhancement module introduces a Batch-level
Heterogeneous Graph Attention Network (BHGAT) component,
which facilitates communication and learning of opinion expressions
among different argument units discussing the same topic.
This component enhances the model’s ability to capture nuanced
relationships and distinctions between argument units. To address
the specific challenges related to aspect term recognition, the
paper presents the Span Mask Interactive Attention (SMIA)
component. This component effectively constrains the recognition
range of aspect terms by leveraging span masks and interactive
attention mechanisms. By doing so, it enables more precise
identification and extraction of aspect terms within their specified
boundaries. Finally, the decision module consolidates the results
from the preceding modules to make the final decision regarding
the extracted argument units and aspect terms. The effectiveness
of the proposed framework and its individual components is
verified through extensive experiments conducted on multiple
datasets and various tasks, reaffirming their potential and value
in the field of ABAM.
License type:
Publisher Copyright
Funding Info:
This research is supported by core funding from: I2R CORE FUNDING
Grant Reference no. : Nil
The works described in this paper are supported by the National
Key Research and Development Program of China (2022QY0300-01), National Natural Science Foundation of China (62106130, 62076158, 62072294, 62272286), Natural Science Foundation of Shanxi
Province, China (20210302124084), Scientific and Technological Innovation Programs of Higher Education Institutions in Shanxi, China (2021L284), and CCF-Zhipu AI Large Model Foundation of China (CCF-Zhipu202310).