Argument Mining(AM) aims to uncover the argumentative structures within a text. Previous methods require several subtasks, such as span identification, component classification, and relation classification. Consequently, these methods need rule-based postprocessing to derive argumentative structures from the output of each subtask. This approach adds to the complexity of the model and expands the search space of the hyperparameters. To address this difficulty, we propose a simple yet strong method based on a text-to-text generation approach using a pretrained encoder-decoder language model. Our method simultaneously generates argumentatively annotated text for spans, components, and relations, eliminating the need for task-specific postprocessing and hyperparameter tuning. Furthermore, because it is a straightforward text-to-text generation method, we can easily adapt our approach to various types of argumentative structures. Experimental results demonstrate the effectiveness of our method, as it achieves state-of-the-art performance on three different types of benchmark datasets: the Argument-annotated Essays Corpus(AAEC), AbstRCT, and the Cornell eRulemaking Corpus(CDCP)
Figures & tables
Figure 1: Overview of our methods. For our methodology, we input text into a pretrained encoder-decoder, such as T5 and FLAN T5. This process generates an argumentatively annotated text with spans, components, and relations. We then postprocess the output text to extract the argumentative structure.
AAEC
AbstRCT
CDCP
# component
6,089
3,279
4,931
# relation
3,832
2,060
1,220
# components with multiple parents
0
31
160
% words in nonargumentative span 3 3 3 We used SpaCy 3.6.1 as a tokenizer for word counts.
28.09
49.30
0
Table 1: Statistics of AAEC, AbstRCT, CDCP.
Input Text
Advantages and disadvantages of the prevalent of English With the development of globalization , English became the dominated language in national trade , conference and many important events . This phenomenon has aroused a heated discussion in public . Some people claim that the prevalent of English brings a great number of benefits for people .
w/ nonargumentative span
Advantages and disadvantages of the prevalent of English With the development of globalization , English became the dominated language in national trade , conference and many important events . This phenomenon has aroused a heated discussion in public . Some people claim that [ the prevalent of English brings a great number of benefits for people | claim for ] .
w/o nonargumentative span
[ the prevalent of English brings a great number of benefits for people | claim for ]
Table 2: Example of input text and output in TANL and our format. The table shows that our output format reduces the number of tokens compared to the TANL format by removing tokens that do not contain any components or relations.
Table 3: Examples of repeated representation and serial representation in CDCP.
Essay
Paragraph
Params
Component
Relation
Component
Relation
ILP Stab and Gurevych (2017)
-
-
-
62.61
34.74
BLCC Eger et al. (2017)
-
63.23
34.82
66.69
39.83
LSTM-ER Eger et al. (2017)
-
66.21
29.56
70.83
45.52
BiPAM-syn Ye and Teufel (2021)
110M
-
-
73.5
46.4
BART-CPM Bao et al. (2022)
139M
-
-
75.94
50.08
Table 4: Evaluation results at both the essay and paragraph levels obtained from AAEC. “Params” indicates the model parameters of the pretrained language model used by each comparison model. Bold indicates the highest F1 score for each task.
C
R
ST Model Morio et al. (2022)
64.16
38.38
FLAN T5-Base
68.76
38.31
FLAN T5-Large
71.11
44.47
FLAN T5-XL
71.27
45.80
FLAN T5-XXL
72.86
47.66
Table 5: Evaluation results for Component-F1 (C) and Relation-F1 (R) in AbstRCT. Bold denotes the highest F1 score for each task.
C
R
BART-CPM Bao et al. (2022)
57.72
16.57
ST Model Morio et al. (2022)
68.90
31.94
FLAN T5-Base
66.80
23.19
FLAN T5-Large
68.94
28.42
FLAN T5-XL
72.12
31.01
FLAN T5-XXL
72.68
33.96
Table 6: Evaluation results for Component-F1 (C) and Relation-F1 (R) in CDCP. Bold denotes the highest F1 score for each task.
AAEC (Essay)
AAEC (Paragraph)
AbstRCT
Output Format
Model
Component
Relation
Component
Relation
Component
Relation
w/ nonargumentative span
FLAN-T5 Base
74.99
50.87
74.97
57.54
65.30
34.55
FLAN T5-Large
77.76
55.62
76.53
59.09
69.47
39.66
FLAN T5-XL
78.73
57.21
77.17
61.03
73.13
42.39
FLAN T5-XXL
80.59
60.37
79.06
62.38
72.78
47.11
w/o nonargumentative span
FLAN-T5 Base
75.17
51.99
75.55
58.51
68.76
38.31
Table 7: Comparison of Component-F1 and Relation-F1 scores with and without nonargumentative span output for essay-level and paragraph-level tasks in AAEC and AbstRCT.
Full Dataset
Multiple Parents
Output Format
Model
Component
Relation
Component
Relation
repeated representation
FLAN-T5 Base
66.94
22.40
60.12
22.64
FLAN T5-Large
66.80
23.19
64.15
27.43
FLAN T5-XL
72.12
31.01
68.67
32.81
FLAN T5-XXL
72.68
33.96
69.97
35.26
serial representation
FLAN-T5 Base
67.11
23.64
63.17
23.81
Table 8: Comparison of Component-F1 (C) and Relation-F1 (R) with different representations in CDCP. Full dataset shows results using all CDCP data, while Multiple Parent shows results using only data with multiple parents.
Figure 2: Comparison of inference time with and without nonargumentative spans in AbstRCT. We set the batch size to 2 for all models during inference and measured the time required to complete the process on the entire test dataset.
Figure 3: Comparison of F1 score output by label for the Component and Relation tasks.
Gold
[ the advertising expenses lead to a higher product price and some of them express fake information , creating information asymmetry between consumers and companies | claim against ] [ its merits still outweigh these downsides | premise | attack = the advertising expenses lead to a higher product price and some of them express fake information , creating information asymmetry between consumers and companies ]
FLAN T5-XL
[ the advertising expenses lead to a higher product price and some of them express fake information , creating information asymmetry between consumers and companies | premise | support = advertisements have no downsides ] [ its merits still outweigh these downsides | premise | attack = the advertising expenses lead to a higher product price and some of them express fake information , creating information asymmetry between consumers and companies ]
FLAN T5-XXL
[ the advertising expenses lead to a higher product price and some of them express fake information , creating information asymmetry between consumers and companies | claim against ] [ its merits still outweigh these downsides | premise | attack = the advertising expenses lead to a higher product price and some of them express fake information , creating information asymmetry between consumers and companies ]
Table 9: Example of output text from FLAN T5-XL and FLAN T5-XXL.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
AAEC (Paragraph)
AAEC (Essay)
Batch size
32
8
Max length
512
1024
Step
10,000
Dropout
0.1
Adam beta1
0.9
Adam beta2
0.998
Appendix
Table 10: Hyperparameters for AAEC.
AbstRCT
CDCP
Batch size
8
8
Max length
1024
1024
Step
10,000
Dropout
0.1
Adam beta1
0.9
Adam beta2
0.998
Appendix
Table 11: Hyperparameters for AbstRCT and CDCP.
r
16
lora alpha
32
lora dropout
0.05
bias
none
task type
SEQ_2_SEQ_LM
target modules
q, v, k, o,
wo, wi_0, wi_1
Appendix
Table 12: Hyperparameters for fine-tuning with QeLoRA.
Component
Rrelation
FLAN T5-Base
75.17
51.99
FLAN T5-Large
77.75
56.06
FLAN T5-XL
78.51
56.80
FLAN T5-XXL
80.15
61.19
GPT4-turbo 3-shot
48.99
24.78
GPT4-turbo 5-shot
51.37
26.43
Appendix
Table 13: Evaluation results of the few-shot learning using GPT-4-turbo. Bold denotes the highest F1 score for each task.
Argumentative component detection (ACD) is a core subtask of Argument(ation) Mining (AM) and one of its most challenging aspects, as it requires jointly delimiting argumentative spans and classifying them into components such as claims and premises. While research on this subtask remains relatively limited compared to other AM tasks, most existing approaches formulate it as a simplified sequence labeling problem, component classification, or a pipeline of component segmentation followed by classification. In this paper, we propose a novel approach based on instruction-tuned Large Language Models (LLMs) using compact instruction-based prompts, and reframe ACD as a language generation task, enabling arguments to be identified directly from plain text without relying on pre-segmented components. Experiments on standard benchmarks show that our approach achieves higher performance compared to state-of-the-art systems. To the best of our knowledge, this is one of the first attempts to fully model ACD as a generative task, highlighting the potential of instruction tuning for complex AM problems.
Sofiane Elguendouze, Erwan Hain, Elena Cabrio +1
Université Côte d’Azur, I3S, CNRS, Inria (Marianne), France
Argument Mining (AM) transforms natural language into its underlying argument structures. This transformation is typically realized through a sequence of AM tasks that form an end-to-end AM pipeline. However, AM approaches often differ in how they conceptualize these tasks, making direct comparisons between them difficult and opaque. This calls for a more nuanced, task-level analysis of AM approaches to enable clearer comparison and assessment. This work presents a preliminary meta-study that systematically reviews several state-of-the-art end-to-end AM works and analyzes their pipelines through a triple-perspective framework---a linguistic, computational and domain perspective---to understand how the pipelines model arguments as structures, computes them, and integrates domain knowledge. We further propose a general design to the linguistic and computational perspectives, illustrating how key AM tasks are designed for modeling and computation of argument structures. Our proposed framework lays the groundwork for methodology-centered descriptions across AM approaches, facilitating deeper understanding and more systematic comparisons in future research.
Siddharth Bhargava, Sara Tonelli, Patricia Martín-Rodilla
Fondazione Bruno Kessler, Trento, Italy · Universidade da Coruña, A Coruña, Spain · IEGPS-CSIC, Spanish National Research Council, Santiago de Compostela, Spain
Argumentative component detection (ACD) is a core subtask of Argument(ation) Mining (AM) and one of its most challenging aspects, as it requires jointly delimiting argumentative spans and classifying them into components such as claims and premises. While research on this subtask remains relatively limited compared to other AM tasks, most existing approaches formulate it as a simplified sequence labeling problem, component classification, or a pipeline of component segmentation followed by classification. In this paper, we propose ITFACD, a novel approach based on instruction-tuned Large Language Models (LLMs) using compact instruction-based prompts, and reframe ACD as a language generation task, enabling arguments to be identified directly from plain text without relying on pre-segmented components. Experiments on standard benchmarks show that our approach achieves higher performance compared to state-of-the-art systems. To the best of our knowledge, this is one of the first attempts to fully model ACD as a generative task, highlighting the potential of instruction tuning for complex AM problems. Our code and the datasets used are openly available in the following GitHub repository.