From Two Passes to One: Compact and Efficient Target-Stance Extraction

Ethan Mines; Bonnie Dorr

cs.CLSep 5, 2026

From Two Passes to One: Compact and Efficient Target-Stance Extraction

Authors: Ethan Mines, Bonnie Dorr

Organizations: University of Florida, Gainesville, FL, USA

Abstract

Target-Stance Extraction (TSE) is the task of predicting both the target (or topic) of an author's writing and the author's stance toward it. Existing approaches to TSE use a sequential pipeline of two separate neural models: one to identify the target and another to determine the stance. We present a one-pass, joint architecture that predicts both in a single forward pass, reducing trainable parameters by nearly 50% with only a 4-7 F1 point tradeoff in performance. We further demonstrate that standard target-scrubbing practices artificially suppress target prediction accuracy. Retaining explicit target mentions, as in real-world deployments, improves F1 by at least 6 points across both target classification and target generation settings. These improvements allow for significantly easier integration of TSE in downstream applications such as public opinion tracking.

Figures & tables

Figure 1: Predicting a target $\hat{t}$ and a stance $\hat{y}$ with two passes (prior approaches) versus one (our approach).

Figure 2: Comparison of Two-Pass and One-Pass TSE under the Target Generation (TG) scenario. In the two-pass setting, a BART sequence transduction model predicts the target while a separate BERT model predicts the stance. In the one-pass setting, a lightweight attention layer is appended to BART for stance prediction. In both cases, FastText embeddings map the free-form target prediction $\tilde{t}$ to a fixed target $\hat{t}$ for evaluation purposes.

Figure 3: Comparison of Two-Pass and One-Pass TSE under the Target Classification (TC) scenario. In the two-pass variant, two separate models predict the target and stance. In the one-pass variant, one BERT model with two classifier heads predicts both.

Subset	#Train / #Val / #Test	Unique Targets
SE	2160/359/1080	Atheism, Feminist Movement, Hillary Clinton, Abortion
AM	18341/2042/5109	Abortion, Cloning, Death Penalty, Gun Control, Marijuana Legalization, Minimum Wage, Nuclear Energy, School Uniforms
C19	4533/800/800	Face Masks, Fauci, Stay at Home Orders, School Closures
PS	17224/2193/2157	Joe Biden, Bernie Sanders, Donald Trump
Unrelated	8800/1120/1880	Unrelated

Table 1: Data partition sizes for SemEval-2016 Task 6 (SE), Argument Mining (AM), COVID-19 (C19), P-Stance (PS), and Unrelated components of the corpus.

Algorithm	Two-Pass	One-Pass
TC Li et al. (2023)	75.59	-
TC (Ours)	76.19	76.23
+No Hashtag	74.51	74.58
+No Scrub	84.67	83.95
+No Hashtag + No Scrub	83.79	82.27
TG Li et al. (2023)	48.31	-

Table 2:

F_{\textrm{mic\_targ}}

performance by system for both Target Classification (TC) and Target Generation (TG) scenarios. Best result for TC is bolded while the best result for TG is underlined .

Algorithm	Two-Pass	One-Pass
TC Li et al. (2023)	53.30	-
TC (Ours)	54.00	50.67
+No Hashtag	52.94	49.39
+No Scrub	60.89	59.41
+No Hashtag + No Scrub	60.73	58.41
TG Li et al. (2023)	38.92	-

Table 3: TSE F1 Scores given Predicted Targets. Best result for Target Classification (TC) is bold ; best for Target Generation (TG) is underlined .

Algorithm	Two-Pass	One-Pass
TC Li et al. (2023)	75.28	-
TC (Ours)	75.91	69.74
+No Hashtag	75.91	69.48
+No Scrub	75.27	72.96
+No Hashtag + No Scrub	75.27	75.27
TG Li et al. (2023)	79.49	-

Table 4: TSE F1 Scores given GT targets. Best result for Target Classification (TC) is bold ; best for Target Generation (TG) is underlined .

Scenario	Relative	Absolute
TC	49.89%	135,505,171
TG	48.32%	133,142,803

Table 5: Reduction of model parameter count from two-pass to one-pass systems for the Target Classification (TC) and Target Generation (TG) scenarios.

		Target
Corpus	Document	Generated	GT
SE	@DesireeAaron @ Hillary Clinton Sheets Clinton Ya got to love it # Hillary onCNN # Hillary Clinton #SemST	donald trump	Hillary Clinton
AM	Proponents of a higher minimum wage state that the current federal minimum wage of $ 7.25 per hour is too low for anyone to live on ; that a higher minimum wage will help create jobs and grow the economy ; that the declining value of the minimum wage is one of the primary causes of wage inequality between low - and middle-income workers ; and that a majority of Americans , including a slim majority of self-described conservatives , support increasing the minimum wage .	income inequality	minimum wage
C19	@Morning_ Joe Well, you Morning joe , idiots President Trump was correct on using #hydroxychloriquine for #coronavirus and # Trump did not need #Dr Fauci to tell him to use it or not. #mikabrzezinski is a Dope, who is 2nd fiddle to Joe . So get behind your Man? and stay there Mika.	bronzirus	fauci
PS	Former Vice President Joe Biden holds a community event in Muscatine, Iowa. # Joe Biden # Biden #IowaCaucus @Team Joe @ Joe Biden # Biden s #Iowa #muscatine @jules1327 @JennyUwishuknew @jmill_30 @morenoandreak @BlueTxBlue Live Link	primaries and caucuses	Joe Biden
Unrelated	Lovin’ that jazz. https://t.co/AAf5cBucla	jazz	Unrelated

Table 6: Target-scrubbed samples with high numbers of explicit target mentions and the corresponding target predictions by the one-pass TG model.

		Target			Stance
Corpus	Document	Generated	Mapped	GT	Predicted	GT
SE	My Jesus, I wish to do nothing but Thy most holy will. #Saint Philip Neri #Catholic #SemST	religion and churches	Feminist Movement	Atheism	Against	Against
AM	Abortions cause psychological damage .	mental health	abortion	abortion	Against	Against
C19	My wife and I were not allowed to work as they say we now have to wear a mask. When told we are exempt they said we need letter from doctor to prove it. Rang doctor’s and they say they will not give us a letter or refer us privately for one. #KBF	mask	face masks	face masks	Neutral	Against
PS	How about @BetoORourke & @KamalaHarris ticket? Not sure if @JoeBiden has been around too long???	rourke & amp kamala harris	Joe Biden	Joe Biden	Against	Against
Unrelated	@HellNaw4Life @iamDOAS Hispanics & Latinos are already white	hispanic americans	Donald Trump	Unrelated	Neutral	Neutral

Table 7: Randomly selected samples without target scrubbing and corresponding predictions by the one-pass Target Generation (TG) model.

Stance Detection

Stance detection identifies whether people support, oppose, or remain neutral toward a target, claim, or position in text and conversation.

See all 36 papers

StanceEval 2026: The Second Stance Detection Shared Task
Oct 2, 2026Rasha Albalawi, Nuha Albadi, Hamzah Luqman +6Stance Detection Cross-Domain Generalization
Information and Computer Science Department, KFUPM, Saudi Arabia · SDAIA–KFUPM Joint Research Center for Artificial Intelligence, Saudi Arabia · HUMAIN, Saudi Arabia
PAST-TIDE: Prototype-Anchored Statement Tuning with Topic-Invariant Normalization for Stance Detection
Jul 6, 2026Md. Shakhoyat Rahman Shujon, MD Jahid Hasan Jim, Md. Milon Islam +2Stance Detection Arabic NLP
Department of Computer Science and Engineering, Khulna University of Engineering & Technology · Department of Electrical and Computer Engineering, University of Waterloo · Department of Machine Learning, Mohamed bin Zayed University of Artificial Intelligence
TTLab at StanceEval-2026: A Cloze-Style Prompting Approach for Arabic-Language Stance Detection (CLASP-Ar)
Sep 24, 2026Bhuvanesh Verma, Ali Abusaleh, Alexander MehlerStance Detection Arabic NLP
Text Technology Lab (TTLab), Goethe University Frankfurt

Oct 2, 2026J/K move · Enter open · S save

StanceEval 2026: The Second Stance Detection Shared Task

Rasha Albalawi, Nuha Albadi, Hamzah Luqman, Asma Yamani+5

Information and Computer Science Department, KFUPM, Saudi Arabia · SDAIA–KFUPM Joint Research Center for Artificial Intelligence, Saudi Arabia · HUMAIN, Saudi Arabia

Stance Detection Cross-Domain Generalization Social Media Analysis

Open paper AskPDF arXiv

StanceEval 2026 is the second edition of the StanceEval shared task series on stance detection in Arabic social media text. Stance detection aims to identify a writer's stance toward a given topic. Given a tweet and a target, participating systems must determine whether the writer's stance is Favor, Against, or None. This edition focuses on cross-target generalization across two distinct evaluation tracks: Track 1 evaluates thematically related cross-target transfer (testing on Women Driving, related to Women Empowerment from training data), while Track 2 evaluates cross-domain transfer to completely unseen targets (E-Cars and Trimester System). The shared task attracted 80 registered teams from 12 countries. During the evaluation phase, 30 unique teams submitted entries, with 21 teams officially ranked in Track 1 and 13 in Track 2 following validation filtering, and 20 teams submitting system-description papers. Participating teams employed diverse methodologies, including fine-tuned pretrained language models, prompt-based and retrieval-augmented large language models (LLMs), fine-tuned LLMs, and hybrid cascades. Top systems achieved impressive

F_{avg2}

scores of 0.8994 on Track 1 and 0.9400 on Track 2, substantially outperforming the strongest baselines (0.7366 and 0.7475, respectively), where

F_{avg2}

denotes the macro-averaged F1 score over the Favor and Against classes. Counterintuitively, performance on the unseen targets was higher than on the related target, a disparity could be driven by extreme target polarization, class imbalance, and dialectal or sarcastic nuance across topics.

Abstract

Stance Detection

Explore similar work

StanceEval 2026: The Second Stance Detection Shared Task

PAST-TIDE: Prototype-Anchored Statement Tuning with Topic-Invariant Normalization for Stance Detection

TTLab at StanceEval-2026: A Cloze-Style Prompting Approach for Arabic-Language Stance Detection (CLASP-Ar)