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.
Figures & tables
Dataset
Argument Components
Claim
Premise
USElecDeb60To16 Haddadan et al. (2019)
29k
26k
Persuasive Essays Stab and Gurevych (2017)
2257
3832
Web Discourse Habernal and Gurevych (2017)
195
538
Merge
31.5k
30.3k
Table 1: Datasets statistics
Model/Dataset
Macro F1
Acc
Human Upper Bound Stab and Gurevych (2017) - PE
0.8860
-
CRF Goudas et al. (2014)
0.4237
-
Heuristic Baseline Stab and Gurevych (2017) - PE
0.642
-
CRF with features Stab and Gurevych (2017) - PE
0.8670
-
MT-all Morio et al. (2022) - PE
0.7566
-
OPT-6.7B - PE
0.8518
0.8856
Table 2: ACD model assessment. PE denotes the Persuasive Essays dataset, Merge is the combination of all three datasets. Rows without references correspond to our proposed methods/models
Gold
And so <premise> if you believe the same thing </premise> , <claim> you just don’t want to raise taxes on people </claim> . And <claim> the reality is it’s not just wealthy people </claim>
OPT-6.7B ACD
And so <premise> if you believe the same thing </premise> , <claim> you just don’t want to raise </claim> taxes on people. And <premise> the reality is it’s not just wealthy people </premise>
Table 3: Example of ACD with argument type errors
Gold
<claim> We will do what we do best</claim>. <claim> It’s a strategy that we’ve been working on for a couple of years </claim> . <claim> It is going to take us to much better advantage in conventional forces </claim>
OPT-6.7B ACD
<claim> We will do what we did best </claim> . <premise> It’s a strategy that we’ve been working on for a few years </premise> . <claim> It is going to take us to much better advantage in conventional forces </claim>
Table 4: Example of ACD with argument type refinement and lexical adjustment
Gold
<claim> Maybe we need to do a better job in mental clinics to help them </claim> . Because <premise> there is a major problem there </premise>
OPT-6.7B ACD
<claim> Maybe we need to do better job in mental clinics </claim> <premise> to help them </premise> . Because <premise> there is a major problem there </premise> .
Table 5: Example of ACD with argument components discovery
Gold
<premise> She’s been doing this for 30 years </premise>
OPT-6.7B ACD
<premise> She’s been doing this job for 30 years </premise>
Table 6: Example of ACD with hallucination
Model
B-C
I-C
B-P
I-P
O
F1-Macro
OPT-6.7B (PE)
0.78
0.76
0.87
0.88
0.97
0.85
GPT-2-1.5B (PE)
0.78
0.76
0.87
0.89
0.96
0.85
Llama-3-8B (PE)
0.84
0.80
0.90
0.88
0.98
0.88
DeBERTa-v3 (PE)
0.6
0.74
0.58
0.79
0.82
0.71
RoBERTa (PE)
0.59
0.68
0.59
0.78
0.19
0.69
OPT-1.3B (Merge)
0.75
0.73
0.71
0.73
0.92
0.77
Table 7: Detailed f1-scores for ACD by experiment configuration
Dataset
O
B-Premise
I-Premise
B-Claim
I-Claim
USElecDeb60To16
566492
26055
350079
29624
338941
Persuasive Essays
35946
2257
29828
3832
59652
Web Discourse
61414
195
3491
538
20566
Table 8: Detailed statistics on the datasets used for fine-tuning and testing LLM models following the BIO-tagging scheme
Figure 1: Full prompt and generation example for ACD task in AM
Gold
As a former educator, I believe that every child deserves equal access to education in the least restrictive environment possible. As a parent, I support that concept as long as the inclusion doesn’t come at the expense of the other students in class. Inclusion can be successful rewarding for all involved but it also has drawbacks.
OPT-6.7B ACD
<claim> As a former educator, </claim> <claim> I believe that every child deserves equal access to education in the less restrictive environment possible. </claim><claim> As a parent, </claim> <premise> I support that concept as long as the inclusion doesn’t come at the expense of the other children in class. </premise> <claim> Inclusion can be successful rewarding for all involved but also has drawbacks. </claim>
Table 9: A second example of ACD with argument components discovery
Gold
<premise> I’m not a politician </premise> , but <claim> I think I could go to Washington in a week and get everybody holding hands and get this bill signed because I talk to the Democratic leaders and they want it </claim>
OPT-6.7B ACD
<premise> I’m not a politician </premise> , but <claim> I think I could go to Washington in a week and get Everybody holding hands and get this Bill signed </claim> <premise> because I talk to the Democratic Leaders and they want it </premise>
Table 10: A second example of ACD with argument type refinement
Faculty of Computer Science, Electronics, and Telecommunications AGH University of Krakow · Faculty of Humanities AGH University of Krakow · Institute of Mathematics University of Silesia