Embedding Models for Stance-Aware Argument Retrieval

Angelo Sparacino; Francesca Toni; Adam Dejl

cs.CLAug 28, 2026

Embedding Models for Stance-Aware Argument Retrieval

Authors: Angelo Sparacino, Francesca Toni, Adam Dejl

Organizations: Department of Computing, Imperial College London, UK

Abstract

In computational argumentation, obtaining arguments that explicitly support or attack given claims is a critical precursor to downstream reasoning tasks. When these supporting and attacking arguments are to be retrieved using semantic search methods, they need to be assessed for topic-relevance to the claims of interest as well as for correctness of their (positive or negative) stance towards the claims. In this paper we explore how dense embedding models (hereafter, models), powering modern retrieval pipelines, can serve as the basis of semantic search incorporating this dual assessment. We show experimentally that existing models struggle with asymmetric reasoning, exhibiting a strong bias toward topical overlap while ignoring instructional stance. We also show that correcting this bias via contrastive training triggers a new failure mode where models over-correct, over-fixating on polarity keywords (e.g., "supports" or "refutes") at the expense of the semantic topic. We thus introduce diagnostic word-ablation metrics to quantify this phenomenon and propose a data-centric solution. By implementing a balanced argument curriculum alongside LLM-augmented, stance-inverted arguments, we force the (embedding) models to learn deeper directional logic rather than exploiting superficial lexical shortcuts. Our evaluation demonstrates that, for sufficiently powerful models, this approach can alleviate the observed overcorrection, achieving further improvements in stance-aware argument retrieval.

Figures & tables

Figure 1: An illustration of the stance-aware argument retrieval problem addressed in this work. The Base Model (left) successfully matches on the semantic topic for all retrieved arguments but ignores the stance instruction (‘rejecting’), resulting in the retrieval of two ‘hard negatives’. Our Tuned Model (right) successfully overcomes this bias, retrieving topically relevant arguments with the correct opposing stance.

Figure 2: Average Precision@R for base and tuned models evaluated on unseen instructions. The base models exhibit severe instruction blindness, retrieving a high proportion of stance-violating hard negatives. While fine-tuning successfully reduces hard-negative retrievals, it also increases the retrieval of semi-hard negatives. This suggests that the used contrastive curriculum forces the models to overly focus on instructional stance at the expense of the claim topic, resulting in a phenomenon that we call topical collapse.

Figure 3: Retrieval composition across the ArgTumour, TFU Evaluation, and AVeriTeC datasets. The chart contrasts the homogeneous baseline against the data-centric interventions (Mixed and Mixed + Aug) across the three model architectures. The balanced curriculum (Mixed) mitigates topical collapse by uniformly reducing the proportion of semi-hard negatives (light grey segments) across all models. However, architectural responses diverge significantly: while Qwen3-Embedding-8B slightly expands its positive retrieval (green segments), both BGE-Large and Instructor-XL exhibit a severe stance regression (inflated red segments) when the hard-negative pressure is reduced. More detailed results can be found in Appendix D

Figure 4: Relative Claim Sensitivity (RCS) distributions for Qwen3-Embedding-8B on AVeriTeC data. Left : The Base model exhibits a severe noun bias, allocating the vast majority of its representational capacity to the claim (red distribution). Applying the homogeneous curriculum triggers topical collapse, shifting focus away from the claim and creating a notable secondary peak with low claim sensitivity (blue distribution). Right : The data-centric interventions regularise the model’s instruction focus. The Mixed (green) and Mixed + Augmentation (red) curricula progressively flatten the peak and pull semantic focus back toward the underlying topic.

Appendix figures & tables30 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Claims	Total Sup.	Avg Sup. (Min-Max)	Total Att.	Avg Att. (Min-Max)
TFU Training Arguments (80% for training)	283	1051	3.71 (0-7)	1146	4.05 (0-7)
TFU Training Arguments (20% held-out)	71	295	4.15 (1-6)	303	4.27 (2-6)
TFU Validation Arguments (GPT-5)	150	622	4.15 (0-12)	882	5.88 (0-13)
TFU Validation Arguments (Qwen3-8B)	150	374	2.49 (0-5)	485	3.23 (0-5)
ArgTumour Arguments	14	40	2.86 (0-16)	58	4.14 (2-10)
TFU Evaluation Arguments	750	1819	2.43 (0-5)	2321	3.09 (0-5)

Appendix

Table 1: Dataset statistics detailing the number of claims, total supporting/attacking arguments, and the average number of supporting/attacking arguments per claim (with minimum and maximum bounds).

Strategy	Stance	NDCG@10	Prec@R	Hard@R	Semi@R	Easy@R
GROUP 1: Dense Retrieval
Zero-Shot	Attack	$0.613\pm 0.28$	$0.438\pm 0.27$	$0.342\pm 0.26$	$0.181\pm 0.27$	$0.038\pm 0.12$
	Support	$0.605\pm 0.27$	$0.401\pm 0.27$	$0.420\pm 0.25$	$0.043\pm 0.12$	$0.136\pm 0.23$
	Overall	$0.609\pm 0.27$	$0.419\pm 0.27$	$0.381\pm 0.26$	$0.112\pm 0.22$	$0.087\pm 0.19$
Homogeneous	Attack	$0.883\pm 0.22$	$0.795\pm 0.27$	$\underline{\mathbf{0.087\pm 0.17}}$	$0.112\pm 0.19$	$0.006\pm 0.03$
	Support	$\underline{\mathbf{0.936\pm 0.12}}$	$\underline{\mathbf{0.850\pm 0.20}}$	$\underline{\mathbf{0.088\pm 0.16}}$	$0.053\pm 0.13$	$0.010\pm 0.05$

Appendix

Table 2: Retrieval performance on the TFU Training Arguments (20% held-out) dataset using BGE-Large .

Strategy	Stance	NDCG@10	Prec@R	Hard@R	Semi@R	Easy@R
GROUP 1: Dense Retrieval
Zero-Shot	Attack	$0.725\pm 0.14$	$0.482\pm 0.19$	$0.507\pm 0.20$	$\underline{\mathbf{0.008\pm 0.06}}$	$0.003\pm 0.02$
	Support	$0.804\pm 0.14$	$0.581\pm 0.19$	$0.410\pm 0.19$	$0.006\pm 0.04$	$\underline{0.003\pm 0.03}$
	Overall	$0.764\pm 0.15$	$0.531\pm 0.20$	$0.459\pm 0.20$	$\underline{\mathbf{0.007\pm 0.05}}$	$\underline{0.003\pm 0.02}$
Homogeneous	Attack	$\underline{\mathbf{0.872\pm 0.23}}$	$\underline{\mathbf{0.773\pm 0.29}}$	$\underline{\mathbf{0.139\pm 0.24}}$	$0.080\pm 0.17$	$0.009\pm 0.04$
	Support	$\underline{\mathbf{0.965\pm 0.09}}$	$\underline{\mathbf{0.905\pm 0.17}}$	$\underline{\mathbf{0.049\pm 0.12}}$	$0.030\pm 0.08$	$0.015\pm 0.07$

Appendix

Table 3: Retrieval performance on the TFU Training Arguments (20% held-out) dataset using Instructor-XL .

Strategy	Stance	NDCG@10	Prec@R	Hard@R	Semi@R	Easy@R
GROUP 1: Dense Retrieval
Zero-Shot	Attack	$0.792\pm 0.15$	$0.559\pm 0.22$	$0.430\pm 0.22$	$0.006\pm 0.03$	$0.004\pm 0.03$
	Support	$0.866\pm 0.13$	$0.634\pm 0.21$	$0.363\pm 0.21$	$0.003\pm 0.02$	$\underline{\mathbf{0.000\pm 0.00}}$
	Overall	$0.829\pm 0.14$	$0.597\pm 0.21$	$0.397\pm 0.22$	$0.005\pm 0.03$	$0.002\pm 0.02$
Zero-Shot + Reranker	Attack	$0.779\pm 0.16$	$0.535\pm 0.23$	$0.465\pm 0.24$	$\underline{\mathbf{0.000\pm 0.00}}$	$\underline{\mathbf{0.001\pm 0.01}}$
	Support	$0.874\pm 0.13$	$0.695\pm 0.22$	$0.305\pm 0.22$	$\underline{\mathbf{0.000\pm 0.00}}$	$\underline{\mathbf{0.000\pm 0.00}}$

Appendix

Table 4: Retrieval performance on the TFU Training Arguments (20% held-out) dataset using Qwen3-Embedding-8B .

Strategy	Stance	NDCG@10	Prec@R	Hard@R	Semi@R	Easy@R
GROUP 1: Dense Retrieval
Zero-Shot	Attack	$0.562\pm 0.31$	$0.448\pm 0.30$	$0.314\pm 0.28$	$\underline{\mathbf{0.131\pm 0.18}}$	$0.106\pm 0.15$
	Support	$0.678\pm 0.22$	$0.523\pm 0.21$	$0.261\pm 0.19$	$\underline{0.120\pm 0.18}$	$0.096\pm 0.15$
	Overall	$0.610\pm 0.28$	$0.480\pm 0.27$	$0.292\pm 0.25$	$\underline{0.126\pm 0.18}$	$0.102\pm 0.15$
Homogeneous	Attack	$0.761\pm 0.25$	$0.655\pm 0.27$	$\underline{\mathbf{0.022\pm 0.07}}$	$0.317\pm 0.26$	$\underline{\mathbf{0.006\pm 0.03}}$
	Support	$0.822\pm 0.19$	$0.730\pm 0.24$	$\underline{\mathbf{0.046\pm 0.11}}$	$0.207\pm 0.22$	$0.017\pm 0.07$

Appendix

Table 5: Retrieval performance on the TFU Validation Arguments generated by GPT-5 dataset using BGE-Large .

Strategy	Stance	NDCG@10	Prec@R	Hard@R	Semi@R	Easy@R
GROUP 1: Dense Retrieval
Zero-Shot	Attack	$0.662\pm 0.26$	$0.529\pm 0.32$	$0.336\pm 0.31$	$\underline{\mathbf{0.079\pm 0.14}}$	$0.056\pm 0.13$
	Support	$0.694\pm 0.24$	$0.558\pm 0.23$	$0.319\pm 0.20$	$\underline{\mathbf{0.070\pm 0.15}}$	$0.054\pm 0.11$
	Overall	$0.675\pm 0.25$	$0.541\pm 0.29$	$0.328\pm 0.27$	$\underline{\mathbf{0.075\pm 0.15}}$	$0.055\pm 0.12$
Homogeneous	Attack	$\underline{\mathbf{0.786\pm 0.24}}$	$\underline{\mathbf{0.687\pm 0.28}}$	$\underline{\mathbf{0.043\pm 0.13}}$	$0.263\pm 0.26$	$\underline{\mathbf{0.007\pm 0.05}}$
	Support	$\underline{\mathbf{0.841\pm 0.19}}$	$\underline{\mathbf{0.754\pm 0.23}}$	$\underline{\mathbf{0.048\pm 0.10}}$	$0.177\pm 0.21$	$0.021\pm 0.06$

Appendix

Table 6: Retrieval performance on the TFU Validation Arguments generated by GPT-5 dataset using Instructor-XL .

Strategy	Stance	NDCG@10	Prec@R	Hard@R	Semi@R	Easy@R
GROUP 1: Dense Retrieval
Zero-Shot	Attack	$0.730\pm 0.24$	$0.600\pm 0.31$	$0.314\pm 0.29$	$\underline{\mathbf{0.057\pm 0.13}}$	$0.030\pm 0.10$
	Support	$0.803\pm 0.18$	$0.667\pm 0.21$	$0.256\pm 0.19$	$\underline{\mathbf{0.055\pm 0.15}}$	$0.022\pm 0.06$
	Overall	$0.760\pm 0.22$	$0.628\pm 0.27$	$0.289\pm 0.26$	$\underline{\mathbf{0.056\pm 0.14}}$	$0.027\pm 0.09$
Homogeneous	Attack	$0.906\pm 0.14$	$0.828\pm 0.19$	$\underline{\mathbf{0.012\pm 0.06}}$	$0.154\pm 0.19$	$0.006\pm 0.04$
	Support	$0.926\pm 0.13$	$0.877\pm 0.19$	$\underline{\mathbf{0.020\pm 0.06}}$	$0.095\pm 0.18$	$0.008\pm 0.03$

Appendix

Table 7: Retrieval performance on the TFU Validation Arguments generated by GPT-5 dataset using Qwen3-Embedding-8B .

Strategy	Stance	NDCG@10	Prec@R	Hard@R	Semi@R	Easy@R
GROUP 1: Dense Retrieval
Zero-Shot	Attack	$0.619\pm 0.24$	$0.408\pm 0.31$	$0.481\pm 0.30$	$\underline{\mathbf{0.036\pm 0.11}}$	$0.075\pm 0.16$
	Support	$0.823\pm 0.22$	$0.611\pm 0.29$	$0.296\pm 0.29$	$\underline{0.067\pm 0.16}$	$0.026\pm 0.09$
	Overall	$0.715\pm 0.25$	$0.503\pm 0.32$	$0.394\pm 0.31$	$\underline{\mathbf{0.051\pm 0.14}}$	$0.052\pm 0.13$
Homogeneous	Attack	$0.807\pm 0.22$	$0.658\pm 0.26$	$\underline{\mathbf{0.064\pm 0.14}}$	$0.270\pm 0.27$	$\underline{\mathbf{0.008\pm 0.05}}$
	Support	$0.844\pm 0.24$	$0.734\pm 0.33$	$\underline{\mathbf{0.104\pm 0.23}}$	$0.134\pm 0.26$	$0.028\pm 0.12$

Appendix

Table 8: Retrieval performance on the TFU Validation Arguments generated by Qwen3-8B dataset using BGE-Large .

Strategy	Stance	NDCG@10	Prec@R	Hard@R	Semi@R	Easy@R
GROUP 1: Dense Retrieval
Zero-Shot	Attack	$0.710\pm 0.22$	$0.492\pm 0.33$	$0.445\pm 0.30$	$\underline{\mathbf{0.025\pm 0.09}}$	$0.038\pm 0.12$
	Support	$0.820\pm 0.21$	$0.595\pm 0.30$	$0.342\pm 0.31$	$\underline{0.041\pm 0.13}$	$0.022\pm 0.09$
	Overall	$0.762\pm 0.22$	$0.540\pm 0.32$	$0.397\pm 0.31$	$\underline{\mathbf{0.033\pm 0.11}}$	$0.030\pm 0.11$
Homogeneous	Attack	$\underline{\mathbf{0.844\pm 0.19}}$	$\underline{\mathbf{0.687\pm 0.26}}$	$\underline{\mathbf{0.100\pm 0.16}}$	$0.209\pm 0.25$	$\underline{\mathbf{0.004\pm 0.04}}$
	Support	$\underline{\mathbf{0.895\pm 0.21}}$	$\underline{\mathbf{0.783\pm 0.31}}$	$\underline{\mathbf{0.103\pm 0.23}}$	$0.106\pm 0.24$	$\underline{\mathbf{0.008\pm 0.06}}$

Appendix

Table 9: Retrieval performance on the TFU Validation Arguments generated by Qwen3-8B dataset using Instructor-XL .

Strategy	Stance	NDCG@10	Prec@R	Hard@R	Semi@R	Easy@R
GROUP 1: Dense Retrieval
Zero-Shot	Attack	$0.780\pm 0.20$	$0.566\pm 0.30$	$0.387\pm 0.28$	$\underline{\mathbf{0.027\pm 0.09}}$	$0.020\pm 0.09$
	Support	$0.881\pm 0.20$	$0.705\pm 0.31$	$0.242\pm 0.31$	$\underline{\mathbf{0.038\pm 0.11}}$	$0.015\pm 0.08$
	Overall	$0.828\pm 0.20$	$0.631\pm 0.31$	$0.319\pm 0.30$	$\underline{\mathbf{0.032\pm 0.10}}$	$0.018\pm 0.08$
Homogeneous	Attack	$0.932\pm 0.12$	$0.832\pm 0.21$	$0.025\pm 0.10$	$0.136\pm 0.20$	$0.007\pm 0.05$
	Support	$0.903\pm 0.25$	$0.836\pm 0.30$	$\underline{\mathbf{0.024\pm 0.12}}$	$0.125\pm 0.27$	$0.014\pm 0.09$

Appendix

Table 10: Retrieval performance on the TFU Validation Arguments generated by Qwen3-8B dataset using Qwen3-Embedding-8B .

Strategy	Stance	NDCG@10	Prec@R	Hard@R	Semi@R	Easy@R
GROUP 1: Dense Retrieval
Zero-Shot	Attack	$0.485\pm 0.40$	$0.317\pm 0.32$	$0.128\pm 0.25$	$0.513\pm 0.31$	$0.043\pm 0.09$
	Support	$0.338\pm 0.30$	$0.259\pm 0.29$	$0.089\pm 0.13$	$\underline{\mathbf{0.381\pm 0.12}}$	$0.271\pm 0.15$
	Overall	$0.441\pm 0.38$	$0.299\pm 0.31$	$0.116\pm 0.22$	$0.473\pm 0.27$	$0.111\pm 0.15$
Homogeneous	Attack	$0.567\pm 0.34$	$0.444\pm 0.31$	$\underline{\mathbf{0.021\pm 0.08}}$	$0.535\pm 0.30$	$\underline{\mathbf{0.000\pm 0.00}}$
	Support	$0.421\pm 0.32$	$\underline{0.321\pm 0.25}$	$0.094\pm 0.14$	$0.585\pm 0.23$	$\underline{\mathbf{0.000\pm 0.00}}$

Appendix

Table 11: Retrieval performance on the ArgTumour Arguments dataset using BGE-Large .

Strategy	Stance	NDCG@10	Prec@R	Hard@R	Semi@R	Easy@R
GROUP 1: Dense Retrieval
Zero-Shot	Attack	$0.579\pm 0.38$	$0.503\pm 0.39$	$0.157\pm 0.26$	$\underline{0.287\pm 0.30}$	$0.054\pm 0.10$
	Support	$0.473\pm 0.25$	$0.371\pm 0.27$	$0.121\pm 0.13$	$\underline{0.413\pm 0.22}$	$0.095\pm 0.18$
	Overall	$0.547\pm 0.35$	$0.463\pm 0.36$	$0.146\pm 0.23$	$\underline{0.324\pm 0.28}$	$0.066\pm 0.13$
Homogeneous	Attack	$0.567\pm 0.42$	$0.507\pm 0.41$	$\underline{\mathbf{0.053\pm 0.14}}$	$0.440\pm 0.37$	$\underline{\mathbf{0.000\pm 0.00}}$
	Support	$\underline{\mathbf{0.534\pm 0.30}}$	$\underline{\mathbf{0.419\pm 0.26}}$	$\underline{\mathbf{0.111\pm 0.18}}$	$0.470\pm 0.20$	$\underline{\mathbf{0.000\pm 0.00}}$

Appendix

Table 12: Retrieval performance on the ArgTumour Arguments dataset using Instructor-XL .

Strategy	Stance	NDCG@10	Prec@R	Hard@R	Semi@R	Easy@R
GROUP 1: Dense Retrieval
Zero-Shot	Attack	$0.705\pm 0.38$	$0.683\pm 0.39$	$0.183\pm 0.29$	$\underline{\mathbf{0.125\pm 0.23}}$	$0.009\pm 0.04$
	Support	$0.782\pm 0.19$	$0.625\pm 0.19$	$0.218\pm 0.19$	$\underline{\mathbf{0.118\pm 0.12}}$	$0.038\pm 0.13$
	Overall	$0.728\pm 0.33$	$0.666\pm 0.34$	$0.193\pm 0.26$	$\underline{\mathbf{0.123\pm 0.20}}$	$0.017\pm 0.08$
Homogeneous	Attack	$0.771\pm 0.23$	$0.626\pm 0.31$	$0.050\pm 0.18$	$0.324\pm 0.30$	$\underline{\mathbf{0.000\pm 0.00}}$
	Support	$0.783\pm 0.15$	$0.669\pm 0.17$	$0.006\pm 0.03$	$0.325\pm 0.17$	$\underline{\mathbf{0.000\pm 0.00}}$

Appendix

Table 13: Retrieval performance on the ArgTumour Arguments dataset using Qwen3-Embedding-8B .

Strategy	Stance	NDCG@10	Prec@R	Hard@R	Semi@R	Easy@R
GROUP 1: Dense Retrieval
Zero-Shot	Attack	$0.590\pm 0.27$	$0.411\pm 0.30$	$0.416\pm 0.30$	$\underline{0.107\pm 0.19}$	$0.066\pm 0.16$
	Support	$0.766\pm 0.26$	$0.591\pm 0.33$	$0.296\pm 0.31$	$0.071\pm 0.18$	$0.042\pm 0.13$
	Overall	$0.673\pm 0.28$	$0.497\pm 0.33$	$0.359\pm 0.31$	$0.090\pm 0.19$	$0.055\pm 0.14$
Homogeneous	Attack	$0.734\pm 0.29$	$0.599\pm 0.31$	$\underline{\mathbf{0.104\pm 0.20}}$	$0.261\pm 0.28$	$0.036\pm 0.12$
	Support	$0.775\pm 0.31$	$0.658\pm 0.36$	$\underline{\mathbf{0.147\pm 0.28}}$	$0.146\pm 0.27$	$0.049\pm 0.15$

Appendix

Table 14: Retrieval performance on the TFU Evaluation Arguments dataset using BGE-Large .

Strategy	Stance	NDCG@10	Prec@R	Hard@R	Semi@R	Easy@R
GROUP 1: Dense Retrieval
Zero-Shot	Attack	$0.740\pm 0.22$	$0.538\pm 0.30$	$0.413\pm 0.29$	$\underline{\mathbf{0.028\pm 0.10}}$	$0.021\pm 0.09$
	Support	$0.757\pm 0.27$	$0.548\pm 0.32$	$0.408\pm 0.32$	$\underline{\mathbf{0.025\pm 0.10}}$	$0.018\pm 0.08$
	Overall	$0.748\pm 0.24$	$0.543\pm 0.31$	$0.411\pm 0.31$	$\underline{\mathbf{0.027\pm 0.10}}$	$0.020\pm 0.09$
Homogeneous	Attack	$\underline{\mathbf{0.797\pm 0.25}}$	$\underline{\mathbf{0.649\pm 0.31}}$	$\underline{\mathbf{0.196\pm 0.27}}$	$0.137\pm 0.22$	$0.018\pm 0.08$
	Support	$\underline{\mathbf{0.861\pm 0.27}}$	$\underline{\mathbf{0.760\pm 0.32}}$	$\underline{\mathbf{0.152\pm 0.29}}$	$0.062\pm 0.17$	$0.026\pm 0.10$

Appendix

Table 15: Retrieval performance on the TFU Evaluation Arguments dataset using Instructor-XL .

Strategy	Stance	NDCG@10	Prec@R	Hard@R	Semi@R	Easy@R
GROUP 1: Dense Retrieval
Zero-Shot	Attack	$0.787\pm 0.20$	$0.581\pm 0.30$	$0.381\pm 0.30$	$\underline{\mathbf{0.025\pm 0.09}}$	$0.013\pm 0.07$
	Support	$0.843\pm 0.26$	$0.680\pm 0.33$	$0.285\pm 0.32$	$\underline{\mathbf{0.021\pm 0.10}}$	$\underline{\mathbf{0.014\pm 0.07}}$
	Overall	$0.814\pm 0.23$	$0.628\pm 0.32$	$0.335\pm 0.32$	$\underline{\mathbf{0.023\pm 0.10}}$	$\underline{\mathbf{0.014\pm 0.07}}$
Homogeneous	Attack	$0.918\pm 0.17$	$0.835\pm 0.24$	$\underline{\mathbf{0.038\pm 0.13}}$	$0.114\pm 0.21$	$0.014\pm 0.07$
	Support	$0.871\pm 0.29$	$0.813\pm 0.32$	$\underline{\mathbf{0.040\pm 0.17}}$	$0.121\pm 0.27$	$0.026\pm 0.12$

Appendix

Table 16: Retrieval performance on the TFU Evaluation Arguments dataset using Qwen3-Embedding-8B .

Strategy	Stance	NDCG@10	Prec@R	Hard@R	Semi@R	Easy@R
GROUP 1: Dense Retrieval
Zero-Shot	Attack	$0.457\pm 0.30$	$0.322\pm 0.28$	$0.315\pm 0.29$	$\underline{\mathbf{0.202\pm 0.26}}$	$0.161\pm 0.24$
	Support	$0.617\pm 0.33$	$0.453\pm 0.36$	$0.300\pm 0.36$	$\underline{0.139\pm 0.26}$	$0.108\pm 0.24$
	Overall	$0.530\pm 0.33$	$0.382\pm 0.33$	$0.308\pm 0.33$	$\underline{0.173\pm 0.26}$	$0.137\pm 0.24$
Homogeneous	Attack	$0.590\pm 0.31$	$0.464\pm 0.31$	$\underline{\mathbf{0.059\pm 0.15}}$	$0.452\pm 0.32$	$0.024\pm 0.09$
	Support	$0.636\pm 0.38$	$0.519\pm 0.40$	$\underline{\mathbf{0.065\pm 0.20}}$	$0.340\pm 0.39$	$0.075\pm 0.22$

Appendix

Table 17: Retrieval performance on the AVeriTeC Arguments dataset using BGE-Large .

Strategy	Stance	NDCG@10	Prec@R	Hard@R	Semi@R	Easy@R
GROUP 1: Dense Retrieval
Zero-Shot	Attack	$0.625\pm 0.27$	$0.452\pm 0.30$	$0.346\pm 0.30$	$\underline{\mathbf{0.120\pm 0.21}}$	$0.081\pm 0.18$
	Support	$0.689\pm 0.31$	$0.488\pm 0.36$	$0.383\pm 0.36$	$\underline{\mathbf{0.069\pm 0.19}}$	$0.060\pm 0.19$
	Overall	$0.655\pm 0.29$	$0.469\pm 0.33$	$0.363\pm 0.33$	$\underline{\mathbf{0.097\pm 0.20}}$	$0.072\pm 0.18$
Homogeneous	Attack	$0.637\pm 0.31$	$\underline{\mathbf{0.496\pm 0.31}}$	$\underline{\mathbf{0.122\pm 0.22}}$	$0.357\pm 0.32$	$\underline{\mathbf{0.025\pm 0.10}}$
	Support	$0.737\pm 0.35$	$\underline{\mathbf{0.623\pm 0.40}}$	$\underline{\mathbf{0.079\pm 0.23}}$	$0.246\pm 0.36$	$0.052\pm 0.18$

Appendix

Table 18: Retrieval performance on the AVeriTeC Arguments dataset using Instructor-XL .

Strategy	Stance	NDCG@10	Prec@R	Hard@R	Semi@R	Easy@R
GROUP 1: Dense Retrieval
Zero-Shot	Attack	$0.689\pm 0.25$	$0.497\pm 0.30$	$0.334\pm 0.30$	$\underline{\mathbf{0.116\pm 0.21}}$	$0.053\pm 0.14$
	Support	$0.756\pm 0.30$	$0.571\pm 0.37$	$0.324\pm 0.37$	$\underline{\mathbf{0.065\pm 0.19}}$	$0.040\pm 0.15$
	Overall	$0.720\pm 0.28$	$0.531\pm 0.33$	$0.330\pm 0.33$	$\underline{\mathbf{0.093\pm 0.20}}$	$0.047\pm 0.15$
Homogeneous	Attack	$0.778\pm 0.25$	$0.642\pm 0.30$	$\underline{\mathbf{0.042\pm 0.13}}$	$0.295\pm 0.29$	$0.020\pm 0.09$
	Support	$0.741\pm 0.37$	$0.658\pm 0.40$	$\underline{\mathbf{0.016\pm 0.11}}$	$0.289\pm 0.39$	$0.037\pm 0.16$

Appendix

Table 19: Retrieval performance on the AVeriTeC Arguments dataset using Qwen3-Embedding-8B .

Model	Strategy	RIS	RCS	DI (Pos)	DI (Hard Neg)
BGE-Large	Zero-Shot	$0.195\pm 0.10$	$0.805\pm 0.10$	$-0.010\pm 0.01$	$-0.002\pm 0.01$
	Homogeneous	$0.354\pm 0.25$	$0.646\pm 0.25$	$-0.123\pm 0.16$	$0.179\pm 0.17$
	Mixed	$0.331\pm 0.22$	$0.669\pm 0.22$	$-0.092\pm 0.13$	$0.157\pm 0.15$
	Mixed + Aug	$0.315\pm 0.21$	$0.685\pm 0.21$	$-0.077\pm 0.12$	$0.138\pm 0.13$
Instructor-XL	Zero-Shot	$0.139\pm 0.09$	$0.861\pm 0.09$	$-0.004\pm 0.01$	$0.001\pm 0.01$
	Homogeneous	$0.374\pm 0.22$	$0.626\pm 0.22$	$-0.085\pm 0.10$	$0.124\pm 0.12$

Appendix

Table 20: Ablation metrics on the TFU Training Arguments (20% held-out) dataset. Results present the mean (

\mu

) and standard deviation (

\sigma

). Metrics include Relative Instruction Sensitivity (RIS), Relative Claim Sensitivity (RCS), and Directional Impact (DI) calculated for positive and hard negative pairs.

Model	Strategy	RIS	RCS	DI (Pos)	DI (Hard Neg)
BGE-Large	Zero-Shot	$0.187\pm 0.13$	$0.813\pm 0.13$	$-0.009\pm 0.01$	$-0.003\pm 0.01$
	Homogeneous	$0.362\pm 0.25$	$0.638\pm 0.25$	$-0.135\pm 0.14$	$0.171\pm 0.16$
	Mixed	$0.318\pm 0.21$	$0.682\pm 0.21$	$-0.091\pm 0.10$	$0.137\pm 0.12$
	Mixed + Aug	$0.295\pm 0.20$	$0.705\pm 0.20$	$-0.075\pm 0.09$	$0.117\pm 0.11$
Instructor-XL	Zero-Shot	$0.134\pm 0.09$	$0.866\pm 0.09$	$-0.004\pm 0.01$	$0.001\pm 0.01$
	Homogeneous	$0.389\pm 0.21$	$0.611\pm 0.21$	$-0.086\pm 0.09$	$0.125\pm 0.10$

Appendix

Table 21: Ablation metrics on the TFU Validation Arguments generated by GPT-5 dataset. Results present the mean (

\mu

) and standard deviation (

\sigma

). Metrics include Relative Instruction Sensitivity (RIS), Relative Claim Sensitivity (RCS), and Directional Impact (DI) calculated for positive and hard negative pairs.

Model	Strategy	RIS	RCS	DI (Pos)	DI (Hard Neg)
BGE-Large	Zero-Shot	$0.134\pm 0.11$	$0.866\pm 0.11$	$-0.007\pm 0.01$	$-0.002\pm 0.01$
	Homogeneous	$0.323\pm 0.24$	$0.677\pm 0.24$	$-0.069\pm 0.12$	$0.157\pm 0.17$
	Mixed	$0.283\pm 0.21$	$0.717\pm 0.21$	$-0.037\pm 0.08$	$0.130\pm 0.14$
	Mixed + Aug	$0.251\pm 0.20$	$0.749\pm 0.20$	$-0.020\pm 0.07$	$0.106\pm 0.11$
Instructor-XL	Zero-Shot	$0.103\pm 0.07$	$0.897\pm 0.07$	$-0.002\pm 0.01$	$0.003\pm 0.01$
	Homogeneous	$0.341\pm 0.20$	$0.659\pm 0.20$	$-0.051\pm 0.08$	$0.115\pm 0.10$

Appendix

Table 22: Ablation metrics on the TFU Validation Arguments generated by Qwen3-8B dataset. Results present the mean (

\mu

) and standard deviation (

\sigma

). Metrics include Relative Instruction Sensitivity (RIS), Relative Claim Sensitivity (RCS), and Directional Impact (DI) calculated for positive and hard negative pairs.

Model	Strategy	RIS	RCS	DI (Pos)	DI (Hard Neg)
BGE-Large	Zero-Shot	$0.152\pm 0.08$	$0.848\pm 0.08$	$-0.004\pm 0.01$	$-0.001\pm 0.01$
	Homogeneous	$0.372\pm 0.24$	$0.628\pm 0.24$	$-0.090\pm 0.13$	$0.091\pm 0.10$
	Mixed	$0.330\pm 0.21$	$0.670\pm 0.21$	$-0.036\pm 0.10$	$0.092\pm 0.09$
	Mixed + Aug	$0.301\pm 0.19$	$0.699\pm 0.19$	$-0.030\pm 0.09$	$0.077\pm 0.07$
Instructor-XL	Zero-Shot	$0.141\pm 0.08$	$0.859\pm 0.08$	$-0.004\pm 0.01$	$0.004\pm 0.01$
	Homogeneous	$0.395\pm 0.21$	$0.605\pm 0.21$	$-0.076\pm 0.12$	$0.064\pm 0.08$

Appendix

Table 23: Ablation metrics on the ArgTumour Arguments dataset. Results present the mean (

\mu

) and standard deviation (

\sigma

). Metrics include Relative Instruction Sensitivity (RIS), Relative Claim Sensitivity (RCS), and Directional Impact (DI) calculated for positive and hard negative pairs.

Model	Strategy	RIS	RCS	DI (Pos)	DI (Hard Neg)
BGE-Large	Zero-Shot	$0.216\pm 0.11$	$0.784\pm 0.11$	$-0.004\pm 0.02$	$0.004\pm 0.02$
	Homogeneous	$0.313\pm 0.22$	$0.687\pm 0.22$	$-0.053\pm 0.14$	$0.131\pm 0.18$
	Mixed	$0.265\pm 0.20$	$0.735\pm 0.20$	$-0.022\pm 0.10$	$0.114\pm 0.14$
	Mixed + Aug	$0.238\pm 0.18$	$0.762\pm 0.18$	$0.000\pm 0.08$	$0.098\pm 0.12$
Instructor-XL	Zero-Shot	$0.133\pm 0.08$	$0.867\pm 0.08$	$0.000\pm 0.01$	$0.005\pm 0.01$
	Homogeneous	$0.315\pm 0.21$	$0.685\pm 0.21$	$-0.022\pm 0.08$	$0.103\pm 0.11$

Appendix

Table 24: Ablation metrics on the TFU Evaluation Arguments dataset. Results present the mean (

\mu

) and standard deviation (

\sigma

). Metrics include Relative Instruction Sensitivity (RIS), Relative Claim Sensitivity (RCS), and Directional Impact (DI) calculated for positive and hard negative pairs.

Model	Strategy	RIS	RCS	DI (Pos)	DI (Hard Neg)
BGE-Large	Zero-Shot	$0.127\pm 0.09$	$0.873\pm 0.09$	$-0.006\pm 0.01$	$-0.002\pm 0.01$
	Homogeneous	$0.311\pm 0.21$	$0.689\pm 0.21$	$-0.083\pm 0.13$	$0.103\pm 0.17$
	Mixed	$0.275\pm 0.19$	$0.725\pm 0.19$	$-0.047\pm 0.09$	$0.098\pm 0.14$
	Mixed + Aug	$0.228\pm 0.17$	$0.772\pm 0.17$	$-0.027\pm 0.07$	$0.074\pm 0.10$
Instructor-XL	Zero-Shot	$0.107\pm 0.07$	$0.893\pm 0.07$	$-0.002\pm 0.01$	$0.002\pm 0.01$
	Homogeneous	$0.312\pm 0.19$	$0.688\pm 0.19$	$-0.047\pm 0.08$	$0.093\pm 0.11$

Appendix

Table 25: Ablation metrics on the AVeriTeC Arguments dataset. Results present the mean (

\mu

) and standard deviation (

\sigma

). Metrics include Relative Instruction Sensitivity (RIS), Relative Claim Sensitivity (RCS), and Directional Impact (DI) calculated for positive and hard negative pairs.

Figure 5: Relative Claim Sensitivity (RCS) distributions for Qwen3-Embedding-8B on the 20% held-out TFU Training Arguments dataset. Left : The Base model exhibits a strong noun bias, focusing primarily on the claim (red distribution). The Homogeneous curriculum triggers topical collapse, shifting attention away from the claim to hyper-fixate on the instruction (blue distribution). Right : The data-centric interventions (Mixed and Mixed + Augmentation) successfully mitigate this collapse, flattening the secondary peak and organically restoring attention to the underlying topic.

Figure 6: Directional Impact (DI) distributions for Qwen3-Embedding-8B evaluated on the 20% held-out TFU Training Arguments dataset. The Base model exhibits near-zero similarity shifts when stance keywords are ablated, confirming its blindness to relational logic. Following fine-tuning, all training paradigms (Homogeneous, Mixed, and Mixed + Augmentation) demonstrate strong asymmetric stance boundaries, successfully pulling positive documents closer (green) while repelling stance-violating hard negatives (red).

Figure 7: Directional Impact (DI) distributions for Qwen3-Embedding-8B on the out-of-domain AVeriTeC Arguments dataset. Mirroring the in-domain results, the Base model fails to process directional instructions. Crucially, the Mixed and Mixed + Augmentation models sustain their wide distributional splits between positive and hard-negative documents, showing that the data-centric interventions maintain strict, generalisable stance boundaries even when recovering from topical collapse.

Figure 8: Word-ablation grids analysing the impact of removing individual words on document retrieval scores for base and tuned (homogeneous) Qwen3-Embedding-8B models. The $x$ -axis displays the removed words from both the instruction and the claim. The $y$ -axis represents the evaluation documents, separated into those attacking ( [POS] ) and supporting ( [NEG] ) the claim. The cell values and colours indicate the first-order delta in retrieval score when a specific word is ablated. The arguments (in-order) are given on the next page.

Figure 9: Hybrid retrieval performance combining the optimally tuned dense bi-encoder (Mixed + Augmentation) with a BM25 sparse retriever via Relative Score Fusion (RSF). The visualisations illustrate the impact of incrementally increasing the sparse weighting ( $\lambda\in[0.1,0.15,0.2]$ ). Introducing the sparse retriever acts as a topical safety net, actively improving precision on highly specialised, entity-dense corpora like ArgTumour. However, because BM25 is fundamentally stance-blind, increasing its influence directly correlates with a higher number of stance errors across all architectures. This posits a strategic trade-off: hybrid search is highly effective for domain-specific entity matching but yields diminishing returns on general-domain datasets (e.g., TFU Evaluation) where the logic of the tuned dense model is already sufficient.

Computational Argumentation

Computational argumentation studies how to represent, analyze, and apply arguments and their relationships using formal and computational methods.

See all 99 papers

Feb 26, 2026

Improving Neural Argumentative Stance Classification in Controversial Topics with Emotion-Lexicon Features

Argumentation mining comprises several subtasks, among which stance classification focuses on identifying the standpoint expressed in an argumentative text toward a specific target topic. While arguments-especially about controversial topics-often appeal to emotions, most prior work has not systematically incorporated explicit, fine-grained emotion analysis to improve performance on this task. In particular, prior research on stance classification has predominantly utilized non-argumentative texts and has been restricted to specific domains or topics, limiting generalizability. We work on five datasets from diverse domains encompassing a range of controversial topics and present an approach for expanding the Bias-Corrected NRC Emotion Lexicon using DistilBERT embeddings, which we feed into a Neural Argumentative Stance Classification model. Our method systematically expands the emotion lexicon through contextualized embeddings to identify emotionally charged terms not previously captured in the lexicon. Our expanded NRC lexicon (eNRC) improves over the baseline across all five datasets (up to +6.2 percentage points in F1 score), outperforms the original NRC on four datasets (up to +3.0), and surpasses the LLM-based approach on nearly all corpora. We provide all resources-including eNRC, the adapted corpora, and model architecture-to enable other researchers to build upon our work.

Mohammad Yeghaneh Abkenar, Weixing Wang, Manfred Stede +3

Innovations Department, Bundesdruckerei GmbH, Berlin, Germany · Department of Linguistics, University of Potsdam, Germany · Hasso-Plattner-Institut, University of Potsdam, Germany +3

Argument Mining Computational Argumentation

Ask PDF

Feb 16, 2026

TruthStance: An Annotated Dataset of Conversations on Truth Social

Argument mining and stance detection are central to understanding how opinions are formed and contested in online discourse. However, most publicly available resources focus on mainstream platforms such as Twitter and Reddit, leaving conversational structure on alt-tech platforms comparatively under-studied. We introduce TruthStance, a large-scale dataset of Truth Social conversation threads spanning 2023-2025, consisting of 24,378 posts and 523,360 comments with reply-tree structure preserved. We provide a human-annotated benchmark of 1,500 instances across argument mining and claim-based stance detection, including inter-annotator agreement, and use it to evaluate large language model (LLM) prompting strategies. Using the best-performing configuration, we release additional LLM-generated labels for 24,352 posts (argument presence) and 107,873 comments (stance to parent), enabling analysis of stance and argumentation patterns across depth, topics, and users. All code and data are released publicly.

Fathima Ameen, Danielle Brown, Manusha Malgareddy +1

North Carolina State University, Raleigh, NC, USA

Social Media Analysis Stance Detection Argument Mining

Ask PDF

Jun 14, 2026

From Argument Components to Graphs: A Multi-Agent Debate with Confidence Gating for Argument Relations

Large Language Models (LLMs) are increasingly assessed and utilized in the field of Argument Mining (AM), thanks to their strong general reasoning capabilities. However, standard training-free models often miss sophisticated details, specifically in contexts where two parts of the text have to be analyzed together. Furthermore, self-correction mechanisms tend to reinforce initial hallucinations in reasoning. Overcoming these limitations typically requires expensive, domain-specific supervised fine-tuning. Recent work has shown that a multi-agent paradigm can address such weaknesses for the component classification task through dialectical refinement with a Proponent-Opponent-Judge architecture, setting a promising direction for training-free approaches in the field. In this paper, we extend and evaluate this framework on the Argument Relation Identification and Classification (ARIC) task, reformulating it as a debate over component pairs. Besides that, we introduce a confidence gating mechanism that enables debating only on the uncertain cases and accepting the initial prediction when confidence is high. On the UKP Argument Annotated Essays v2 corpus, we demonstrate that the selective debate achieves the highest Macro F1 among all training-free methods, while debate over all samples degrades performance below that of one of the baselines. All generative approaches also outperform fine-tuned RoBERTa models on Macro F1, suggesting that the under-representation of the Attack class was more damaging to supervised fine-tuning than to inference-only models. Additionally, our framework produces human-readable debate transcripts, offering interpretability absent from both single-agent and supervised classifiers.

Jakub Bąba, Jarosław A. Chudziak

Faculty of Electronics and Information Technology, Warsaw University of Technology, Poland

Relation Extraction Argument Mining Multi-Agent Debate

Ask PDF

Abstract

Computational Argumentation

Explore similar work

Improving Neural Argumentative Stance Classification in Controversial Topics with Emotion-Lexicon Features

TruthStance: An Annotated Dataset of Conversations on Truth Social

From Argument Components to Graphs: A Multi-Agent Debate with Confidence Gating for Argument Relations