The Overlooked Repetitive Lengthening Form in Sentiment Analysis

Lei Wang; Eduard Dragut

cs.CLApr 1, 2026

The Overlooked Repetitive Lengthening Form in Sentiment Analysis

Authors: Lei Wang, Eduard Dragut

Organizations: Temple University, Philadelphia, PA, USA

Abstract

Individuals engaging in online communication frequently express personal opinions with informal styles (e.g., memes and emojis). While Language Models (LMs) with informal communications have been widely discussed, a unique and emphatic style, the Repetitive Lengthening Form (RLF), has been overlooked for years. In this paper, we explore answers to two research questions: 1) Is RLF important for sentiment analysis (SA)? 2) Can LMs understand RLF? Inspired by previous linguistic research, we curate Lengthening, the first multi-domain dataset with 850k samples focused on RLF for SA. Moreover, we introduce Explainable Instruction Tuning (ExpInstruct), a two-stage instruction tuning framework aimed to improve both performance and explainability of LLMs for RLF. We further propose a novel unified approach to quantify LMs' understanding of informal expressions. We show that RLF sentences are expressive expressions and can serve as signatures of document-level sentiment. Additionally, RLF has potential value for online content analysis. Our results show that fine-tuned Pre-trained Language Models (PLMs) can surpass zero-shot GPT-4 in performance but not in explanation for RLF. Finally, we show ExpInstruct can improve the open-sourced LLMs to match zero-shot GPT-4 in performance and explainability for RLF with limited samples. Code and sample data are available at https://github.com/Tom-Owl/OverlookedRLF

Figures & tables

Figure 1: An overview of our work for RLF. (1) We introduce Lengthening in Section 3 . (2) We propose the ExpInstruct framework and describe prompt details in Section 4.1 . (3) Experiments details are in Section 5 .

Domain	Dataset	Original Tag	#Document	RLF Ratio(%)	#Samples	Label Distribution (1/0)	#Unique RLF Word	#Unique Root Word
Books	Amazon Review	1-5 stars	51,312k	5.41	255k	92/8(%)	11,813	1,592
Electronics	Amazon Review	1-5 stars	20,994k	4.32	212k	74/26(%)	11,367	1,541
Restaurants	Yelp	1-5 stars	6,990k	11.28	315k	74/26(%)	11,642	1,587
Social Media	Twitter	binary	1,600k	13.36	44k	46/54(%)	5,337	1,149
Hotels	TripAdvisor	1-5 stars	879k	10.01	24k	76/24(%)	4,558	1,250
ALL	Lengthening	binary	850k	100	850k	78/22(%)	19,610	1,677

Table 1: Summary statistics for our dataset. Lengthening is the first large-scale dataset featuring RLF for SA task, grounded in 4 public datasets with an average of 5.8% documents containing RLF. More details of the generation process of our dataset are described in Section 3

Domain	Lengthening Style	POS	Lengthening Word	Root Word	Label	Example Sentence
Books	Punctuation	Noun	book!!!!!	book!	1	Do yourself a favour a read this book!!!!!
Electronics	Letter	Verb	loooove	love	1	I loooove my new phone case.
Restaurants	Punctuation	Adj	amazing!!!!!	amazing!	1	We are from Seattle and this coffee is amazing!!!!!
Twitter	Letter	Adv	SOOOO	SO	0	SOOOO bummed i’m going to miss sam’s party tonight.
Hotel	Punctuation	Noun	year………….	year…	0	I am looking to go back next year………….

Table 2: Samples from our Lengthening dataset.

Figure 2: Prompt Design and Template for ExpInstruct. (a) Prompt with CoT for word-level explainability. (b) Simple Prompt for SA. (c) Prompt Template for Instruction tuning

Figure 3: Comparing normalized WIS for an RLF sentence from zero-shot GPT-4 and fine-tuned RoBERTa.

Metric	Acc(%)		F1(%)
Backbone Model	RLF	w/o RLF	RLF	w/o RLF
RoBERTa (Large)	85.94 $\pm$ 0.03	84.40 $\pm$ 0.04	90.67 $\pm$ 0.04	89.56 $\pm$ 0.03
GPT-2 (Medium)	79.56 $\pm$ 0.07	77.56 $\pm$ 0.08	85.76 $\pm$ 0.08	84.22 $\pm$ 0.06
T5 (Base)	83.22 $\pm$ 0.06	81.93 $\pm$ 0.04	88.71 $\pm$ 0.07	87.85 $\pm$ 0.03
LLaMA2 (13B)	76.25 $\pm$ 1.76	70.41 $\pm$ 0.57	84.30 $\pm$ 1.28	82.64 $\pm$ 0.39
GPT-4	86.26 $\pm$ 0.93	86.20 $\pm$ 1.45	90.12 $\pm$ 0.75	90.08 $\pm$ 1.14

Table 3: Overall zero-shot accuracy and F1 score for sentences with RLF and without RLF words (w/o RLF). Bold denotes the best results in a column and underline highlights the second best results.

\pm

indicates standard deviation score.

Panel of Figure 4 — Figure 4: Comparison of accuracy between the RLF and w/o RLF groups using zero-shot and fine-tuned models by sentence length. The lines represent average values, and the error bar indicate the standard deviation for each length group across 3 runs. Both results across 3 fine-tuned models show a convergence between the RLF and w/o RLF groups when the sentence character length is around 80.

Backbone Model	RoBERTa (Large)				GPT-2 (Medium)				T5 (Base)
Metric	Acc(%)		F1(%)		Acc(%)		F1(%)		Acc(%)		F1(%)
Domain / Group	RLF	w/o RLF	RLF	w/o RLF	RLF	w/o RLF	RLF	w/o RLF	RLF	w/o RLF	RLF	w/o RLF
Books	87.36 $\pm$ 0.17	86.90 $\pm$ 0.16	92.76 $\pm$ 0.10	92.46 $\pm$ 0.10	80.65 $\pm$ 0.07	81.21 $\pm$ 0.18	88.45 $\pm$ 0.06	88.84 $\pm$ 0.11	86.09 $\pm$ 0.17	87.07 $\pm$ 0.08	92.01 $\pm$ 0.10	92.63 $\pm$ 0.05
Restaurants	88.46 $\pm$ 0.07	85.93 $\pm$ 0.09	92.07 $\pm$ 0.07	90.24 $\pm$ 0.06	82.67 $\pm$ 0.20	79.04 $\pm$ 0.02	87.60 $\pm$ 0.18	84.70 $\pm$ 0.02	85.26 $\pm$ 0.23	82.79 $\pm$ 0.14	89.70 $\pm$ 0.19	87.95 $\pm$ 0.09
Electronics	84.56 $\pm$ 0.08	82.54 $\pm$ 0.19	88.86 $\pm$ 0.10	87.27 $\pm$ 0.16	76.35 $\pm$ 0.09	73.28 $\pm$ 0.17	81.72 $\pm$ 0.16	78.99 $\pm$ 0.21	80.53 $\pm$ 0.12	77.84 $\pm$ 0.05	85.65 $\pm$ 0.02	83.54 $\pm$ 0.09
Twitter	66.30 $\pm$ 0.36	66.87 $\pm$ 0.40	68.64 $\pm$ 0.35	68.41 $\pm$ 0.43	66.84 $\pm$ 0.13	65.71 $\pm$ 0.57	66.11 $\pm$ 0.14	65.53 $\pm$ 0.98	65.55 $\pm$ 0.35	64.56 $\pm$ 0.05	66.36 $\pm$ 0.50	65.22 $\pm$ 0.19

Table 4: Zero-shot accuracy and F1 score for sentences with/without RLF words in different domains. We report RoBERTa (Large), GPT-2 (Medium) and T5 (Base) because limited test samples for GPT-4 and LLaMA2-13B. Bold and underline indicate best and second best results.

\pm

indicates standard deviation score.

Lengthening Style	Punctuation Repetitive			Letter Repetitive			Overall
Backbone Model	$S_{exp}$	Acc(%)	F1(%)	$S_{exp}$	Acc(%)	F1(%)	$S_{exp}$	Acc(%)	F1(%)
Zero-shot
RoBERTa (Large)	0.24 $\pm$ .001	86.65 $\pm$ 0.01	91.19 $\pm$ 0.01	0.10 $\pm$ .001	83.61 $\pm$ 0.16	88.93 $\pm$ 0.17	0.21 $\pm$ .001	85.94 $\pm$ 0.03	90.67 $\pm$ 0.04
GPT-2 (Medium)	0.49 $\pm$ .001	80.55 $\pm$ 0.11	86.57 $\pm$ 0.08	0.07 $\pm$ .001	76.34 $\pm$ 0.11	82.98 $\pm$ 0.15	0.39 $\pm$ .002	79.56 $\pm$ 0.07	85.76 $\pm$ 0.08
T5 (Base)	0.20 $\pm$ .001	83.97 $\pm$ 0.04	89.30 $\pm$ 0.04	0.11 $\pm$ .001	80.75 $\pm$ 0.10	86.74 $\pm$ 0.16	0.18 $\pm$ .001	83.22 $\pm$ 0.06	88.71 $\pm$ 0.07
LLaMA2 (13B)	0.18 $\pm$ .004	77.25 $\pm$ 2.62	85.23 $\pm$ 1.89	0.25 $\pm$ .003	73.54 $\pm$ 3.86	81.46 $\pm$ 2.88	0.20 $\pm$ .004	76.25 $\pm$ 1.76	84.30 $\pm$ 1.28

Table 5: Comparision performances between zero-shot and finetuned models with two lengthening styles. In each column, the best result is highlighted in bold, and the second best result is underlined.

	PP	PN	NP	NN
RLF	127	9	11	53
w/o RLF	120	16	13	51

Table 6: Confusion matrices for sample distribution. We categorized data with combinations of document and sentence labels (e.g., PP represents Positive document with Positive sentence).

Backbone Model	IAA	Reliability	$S_{exp}$
RoBERTa (Large)	0.41	0.57 $\pm$ .159	0.22 $\pm$ .010
GPT-2 (Medium)	0.44	0.59 $\pm$ .110	0.29 $\pm$ .005
T5 (Base)	0.42	0.48 $\pm$ .093	0.23 $\pm$ .002
GPT-4	0.43	0.79 $\pm$ .055	0.38 $\pm$ .005
ExpInstruct	0.34	0.67 $\pm$ .105	0.35 $\pm$ .019

Table 7: Detail results of Explanation Reliability for zero-shot GPT-4 and the 4 fine-tuned models (RoBERTa, GPT-2, T5, and ExpInstruct). The correlation coefficient between the reliability score and

S_{exp}

is 0.91.

Metric	Acc(%)		F1(%)
Backbone Model	Zero-shot	Fine-tuned	Zero-shot	Fine-tuned
RoBERTa (Large)	93.99 $\pm$ 0.13	95.99 $\pm$ 0.51 $\uparrow$	95.82 $\pm$ 0.13	97.18 $\pm$ 0.36 $\uparrow$
GPT-2 (Medium)	89.79 $\pm$ 0.38	94.14 $\pm$ 0.16 $\uparrow$	92.72 $\pm$ 0.35	95.94 $\pm$ 0.11 $\uparrow$
T5 (Base)	92.61 $\pm$ 0.64	93.37 $\pm$ 0.06 $\uparrow$	94.84 $\pm$ 0.47	95.38 $\pm$ 0.04 $\uparrow$
LLaMA2 (13B)	71.06 $\pm$ 0.42	94.18 $\pm$ 0.63 $\uparrow$	83.04 $\pm$ 0.34	95.96 $\pm$ 0.37 $\uparrow$

Table 8: Document-level gains for LMs fine-tuned with Lengthening . We observe average improvement in accuracy as 7.6% and F1 score as 4.5% among the models.

	$S_{exp}$	Acc(%)	F1(%)
# Training Samples
0	0.20 $\pm$ .004	76.25 $\pm$ 1.76	84.30 $\pm$ 1.28
500	0.33 $\pm$ .005	82.09 $\pm$ 0.89	86.94 $\pm$ 0.35
1,000	0.35 $\pm$ .015	85.36 $\pm$ 0.18	89.58 $\pm$ 0.26
ExpInstruct (1,600)	0.35 $\pm$ .019	87.20 $\pm$ 0.69	90.96 $\pm$ 0.51
Instruction Strategy

Table 9: Results of ablation study for effects of sample size and instruction strategy.

Backbone Model	$S_{exp}$	Acc(%)	F1(%)
Fine-tuned with subset & Test on subset
RoBERTa (Large)	0.19 $\pm$ .005	84.70 $\pm$ 0.83	89.18 $\pm$ 0.59
GPT-2 (Medium)	0.38 $\pm$ .016	79.00 $\pm$ 0.76	84.32 $\pm$ 0.59
T5 (Base)	0.12 $\pm$ .003	79.52 $\pm$ 0.76	84.97 $\pm$ 0.65
Zero-shot with OOD (randomly sample 3k)
GPT-4	0.35 $\pm$ .008	87.55 $\pm$ 0.92	91.61 $\pm$ 0.53

Table 10: Comparision performances on the subset, PLMs finetuned on Lengthening and test on subset

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Model	#Dataset (k)	#train (k)	#val (k)	#test (k)	data split	experiment
PLMs	850	595.2	84.8	170	7 : 1 : 2	3 runs
ExpInstruct	3	1.6	0.4	1.0	custom	3-fold

Appendix

Table 11: Details of dataset split.

Lengthening Style	POS Tag	Ratio (%)	Examples of normalized forms
Letter Repetitive	ADV	14.41	’so+’, ’wa+y’, ’way+’, ’reall+y’, ’wa+y+’
Letter Repetitive	INTJ	6.67	’hmm+’, ’ah+’, ’oh+’, ’um+’, ’aw+’
Letter Repetitive	others	3.20	’all+’, ’to+’, ’go+d’, ’bu+t’, ’al+’
Letter Repetitive	VERB	3.19	’lo+ve’, ’love+’, ’lo+ved’, ’recomm+end’, ’lo+ves’
Letter Repetitive	ADJ	3.06	’lo+ng’, ’hu+ge’, ’slo+w’, ’long+’, ’litt+le’
Letter Repetitive	NOUN	1.71	’boo+k’, ’go+d’, ’a+s’, ’wa+y’, ’boo+’

Appendix

Table 12: Examples of normalized forms by lengthening style and POS tag.

Figure 5: Our customized user interface for human evaluation. Annotators are asked to do two tasks: annotation for sentiment labels and explanation reliability.

Sentiment Analysis

Sentiment analysis detects and measures opinions, emotions, and evaluative stance in text and multimodal data.

See all 127 papers

LFQA-HP-1M: A Large-Scale Human Preference Dataset for Long-Form Question Answering
Feb 27, 2026Rafid Ishrak Jahan, Fahmid Shahriar Iqbal, Sagnik Ray ChoudhuryLLM Evaluation LLM-as-a-Judge
University of North Texas, Denton, Texas, USA
Usage-Modulated Sentiment Representations in Large Language Models
Oct 4, 2026Hongfei Du, Jiacheng Shi, Yanfu Zhang +2Causal Interventions in Language Models Representation Geometry in Language Models
Department of Computer Science, William & Mary, Williamsburg, VA, USA
Task Complexity Matters: An Empirical Study of Reasoning in LLMs for Sentiment Analysis
Feb 27, 2026Donghao Huang, Zhaoxia WangEfficient Language Model Reasoning LLM Evaluation
School of Computing and Information Systems, Singapore Management University, 80 Stamford Rd, Singapore 178902, Singapore · Research and Development, Mastercard, 4250 Fairfax Dr, Arlington, VA 22203, USA

Feb 27, 2026J/K move · Enter open · S save

LFQA-HP-1M: A Large-Scale Human Preference Dataset for Long-Form Question Answering

Rafid Ishrak Jahan, Fahmid Shahriar Iqbal, Sagnik Ray Choudhury

University of North Texas, Denton, Texas, USA

LLM Evaluation LLM-as-a-Judge Bias in LLM-as-a-Judge

Open paper AskPDF arXiv

Long-form question answering (LFQA) demands nuanced evaluation of multi-sentence explanatory responses, yet existing metrics often fail to reflect human judgment. We present LFQA-HP-1M, a large-scale dataset comprising 1.3M human pairwise preference annotations for LFQA. We propose nine rubrics for answer quality evaluation, and show that simple linear models based on these features perform comparably to state-of-the-art LLM evaluators. We further examine transitivity consistency, positional bias, and verbosity biases in LLM evaluators and demonstrate their vulnerability to adversarial perturbations. Overall, this work provides one of the largest public LFQA preference datasets and a rubric-driven framework for transparent and reliable evaluation.

Abstract

Sentiment Analysis

Explore similar work

LFQA-HP-1M: A Large-Scale Human Preference Dataset for Long-Form Question Answering

Usage-Modulated Sentiment Representations in Large Language Models

Task Complexity Matters: An Empirical Study of Reasoning in LLMs for Sentiment Analysis