Structured Sentiment Analysis Using Sequence Labeling as Dependency Graph Parsing

Muhammad Imran; Ana Ezquerro; Carlos Gómez-Rodríguez; Anders Søgaard; David Vilares

cs.CLOct 8, 2026

Structured Sentiment Analysis Using Sequence Labeling as Dependency Graph Parsing

Authors: Muhammad Imran, Ana Ezquerro, Carlos Gómez-Rodríguez, Anders Søgaard, David Vilares

Organizations: Universidade da Coruña, CITIC, Departamento de Ciencias de la Computación y Tecnologías de la Información, Campus de Elviña s/n, 15071, A Coruña, Spain · Graz University of Technology, Institute in Machine Learning and Neural Computation, Rechbauerstraße 12, 8010, Graz, Austria · University of Copenhagen, Department of Computer Science, Lyngbyvej 2, DK-2100, Copenhagen, Denmark

Abstract

This study addresses the problem of structured sentiment analysis, whose goal is to obtain a fine-grained sentiment graph where the nodes represent spans of sentiment holders, targets, and expressions, while the arcs define the relationships among them. Our proposed approach casts the task as dependency graph parsing, but departs from traditional parsing methods by solving it through sequence labeling. To do so, we leverage recent advances in linearized graph encodings that allow each word in the input to be assigned a label, effectively capturing the structure of the dependency graph. We conducted experiments on seven datasets spanning five languages (English, Spanish, Norwegian, Basque, and Catalan), showing performance competitive with leading, more complex single-model approaches.

Figures & tables

Figure 1: Head-first transformation of a sentiment graph to a dependency graph [ 6 ] . The sentiment graph has two opinion tuples:

O_{1}=(\emptyset

, the hotel , in great condition , +

)

and

O_{2}=(

my experience , the service , ruined , -

)

Panel of Figure 2 — Figure 2: Absolute ( A ), relative ( R ) and bracketing ( B ) encodings for an example extracted from [ 7 ] . Note that the <b> token is needed to identify the roots of the graph.

Panel of Figure 3 — Figure 3: Graph example with two crossing arcs in the same direction: $(0\to 6)$ crosses $(3\to 7)$ . For the bracketing ( B ) encoding, $(3\to 7)$ is encoded with a different set of brackets ( /* , >* ). The $6k$ bit encodings require $k=2$ , so each plane is encoded with an independent sequence of bits.

Panel of Figure 4 — Figure 4: $4k$ ( B4 $k$ ) and $6k$ -bit ( B6 $k$ ) encoding example from the graph displayed in Figure 3 . Note that the arcs of the original graph need to be distributed in two different subsets that are independently encoded. Artificial arcs (green dashed lines) are generated in the $4k$ -bit encoding to satisfy the first assumption of the algorithm (each node needs one and only one parent). Those arcs are ignored for B6 1 and B6 2 . Note that in this case the subsets of the $4k$ and $6k$ -bit encoding are the same, but this might not happen for other graphs.

Figure 5: Visualization of the system at training time. Each module is represented with a different color: purple for the sentiment-to-dependency transformation, green for the graph linearization and blue for the neural network. Hypothetical errors are displayed in red for completeness. The input is a sentiment graph ( 1 ) that is converted into a dependency graph ( 2 ) through Barnes et al. [6] . The input sentence is fed to $\mathcal{M}_{\theta}^{e}$ and $\mathcal{M}_{\theta}^{\ell}$ to obtain the hidden embeddings ( 3 ) and the predicted labels ( 4 ). The arcs from 2 are used to pair embeddings and feeding them to $\mathcal{M}_{\theta}^{r}$ ( 5 ). The network is optimized with the cross-entropy loss of the predicted labels and predicted relations.

Figure 6: Visualization of the system at inference time. Same color notation as in Figure 5 . The input is a sentence ( 1 ). The embeddings and predicted labels are obtained by the trained system, as depicted in Figure 5 . The decoding process is applied to the predicted labels to recover an unlabeled dependency graph. The $r$ component of each predicted arc is obtained by gathering the corresponding embeddings of 2 and feed them to $\mathcal{M}_{\theta}^{r}$ . The predicted dependency graph ( 5 ) is transformed into a sentiment graph ( 6 ) through the inverse transformation of [ 6 ] .

	#sents	#holders	#targets	#expr	$n$	$\|O\|$	$\|A\|/n$
DS en	2253	69	869	1525	19.99	0.36	0.07
	232	8	117	186	18.09	0.42	0.09
	318	12	155	251	20.40	0.41	0.08
MPQA en	5581	2925	8625	3147	24.37	0.28	0.10
	1979	852	2499	1024	23.91	0.27	0.08
	2025	894	2696	958	23.85	0.24	0.08

Table 1: General statistics of the datasets used in our experimental study. Subscripts denote the language denoted by its ISO-639 code. Different splits in each subrow: train, development and test. Columns

n

|O|

and

|A|/n

denote the average sentence length, number of opinion tuples and ratio of number of arcs and nodes, respectively.

	Br.			$6k$ -bit			$\Omega$
	1	2	3	1	2	3	$\Omega$
DS en	99.0	99.8	99.8	92.9	99.8	99.8	99.8
	99.0	99.8	99.8	92.5	99.8	99.8	99.8
MPQA en	92.4	95.5	95.9	83.1	94.2	95.7	95.9
	90.4	96.5	96.5	82.3	95.0	96.5	96.5
MB ca	99.1	99.8	99.8	86.1	97.9	99.4	99.8

Table 2: Coverage (SF 1 ) of each encoding in the test split.

\Omega

represents the SF 1 of the original sentiment to dependency transformation. The first and second subrows for each dataset are for the head-first and head-final approach, respectively.

		A	R	Br.			$6k$ -bit			#rels
		A	R	1	2	3	1	2	3	#rels
head-first	DS en	117	134	37	45	45	15	30	37	6
	MPQA en	528	615	170	380	403	22	96	140	13
	MB ca	303	274	96	138	138	19	55	72	4
	MB eu	197	193	59	91	91	18	48	61	4
	NoReC no	1047	1371	276	722	778	24	147	243	14

Table 3: Size of the label space (

|\mathcal{L}|

) for each encoding and representation; and number of unique arc labels ( #rels ). Same notation as in Table 4 .

	h.first									h.final
	A	R	Br.			$6k$ -bit			Biaf.	A	R	Br.			$6k$ -bit			Biaf.
	A	R	1	2	3	1	2	3	Biaf.	A	R	1	2	3	1	2	3	Biaf.
DS en	37.9	19.8	44.0	41.4	41.4	43.8	46.1	43.3	35.0	42.1	18.6	36.5	39.8	40.9	41.2	38.9	41.2	44.3
MPQA en	17.6	18.6	21.3	17.1	28.5	32.6	28.8	32.3	29.8	16.8	13.1	27.2	17.8	28.5	31.5	34.8	31.6	29.8
MB ca	68.3	57.1	61.6	69.2	70.8	73.0	77.0	71.8	61.1	51.1	47.6	56.8	64.6	62.7	62.8	65.0	67.0	69.3
MB eu	71.3	57.1	60.5	66.4	69.5	72.5	68.6	71.5	62.1	61.4	54.3	55.3	64.6	60.6	65.5	63.0	65.2	70.9

Table 4: SF 1 score in the test set for the absolute ( A ), relative ( R ), bracketing ( Br ) and

6k

-bit encodings. Subcolumns in the bracketing and bit-based encodings denote the value of the hyperparameter

k

. The best result is highlighted in bold. The average is included in the last row (

\mu

	DS en	MPQA en	MB ca	MB eu	NoReC no	OpeNER en	OpeNER es	$\mu$
* [ 22 ]	49.4	44.7	72.8	73.9	52.9	76.0	72.2	63.1
[ 8 ]	48.5	41.6	72.8	73.9	52.4	76.3	74.2	62.8
[ 27 ]	46.3	40.2	70.9	71.5	53.3	75.6	73.2	61.6
[ 4 ]	41.0	37.5	68.1	72.3	50.4	74.7	73.5	59.6
[ 1 ]	41.4	34.9	65.3	68.0	46.2	69.8	69.2	56.4
[ 52 ]	49.0	35.1	68.4	68.6	49.6	67.6	62.3	56.1

Table 5: SF 1 comparison of different studies and our best-performing encoding (

6k

-bit with

k=2

). Systems proposed in the SemEval-2022 Task 10 [ 7 ] are marked with an asterisk. Best results highlighted in bold.

Figure 7: F 1 -score ( $y$ -axis) per displacement ( $x$ -axis) in the head-first (left) and head-final (right) representations, alongside the rug plot of the ground truth displacements. The biaffine baseline (DM) and the encodings are denoted with different acronyms: absolute (A), relative (R), bracketing (B), and $6k$ -bit (B6) encodings, using subscripts for the value of the hyperparameter $k$ , when applicable.

Dataset	Model	Spans (F 1 )			T.SF 1	Dep.		Sent.
		Holder	Target	Exp	T.SF 1	UF 1	LF 1	NSF 1	SF 1
MB ca	[ 37 ] OT	48.0	72.5	68.9	-	-	-	65.7	63.3
	[ 37 ] LE	60.8	70.8	72.5	-	-	-	64.5	62.1
	[ 37 ] NC	56.1	69.8	70.5	-	-	-	63.5	61.7
	[ 21 ]	36.8	-	75.5	73.4	-	-	66.4	63.2
	[ 32 ]	46.2	74.2	71.0	60.9		64.5	-	59.3

Table 6: Fine-grained SSA results comparison against state of the art systems and Biaffine baseline: Spans (F 1 ) column is the token-level F 1 -score for the holder, target and expression; T.SF 1 is the targeted F 1 score; UF 1 and LF 1 are the unlabeled and labeled F 1 scores for dependency graph parsing; and S F 1 and NSF 1 are the sentiment F 1 score with or without the sentiment polarity. Our best-performing encoding

6k

-bit ( B6 3 ) where subscript denotes the value of the hyperparemeter

k

. Dash (-) for results not reported in the original paper. Best result highlighted in bold.

Figure 8: Performance (SF 1 , y-axis) vs inference speed (tokens per second, x-axis) on the head-first representation. The Pareto front is displayed with dashed lines. The biaffine baseline (Biaff) and the encodings are denoted with different acronyms: absolute (A), relative (R), bracketing (B), and $6k$ -bit (B6), using subscripts for the value of the hyperparameter $k$ , where applicable.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

	Spans (F 1 )			T.SF 1	Dep.		Sent.
	Holder	Target	Exp	T.SF 1	UF 1	LF 1	NSF 1	SF 1
[ 32 ]	65.8	71.0	76.7	59.6		66.1	-	62.7
[ 37 ] NC	58.9	63.5	73.9	-	-	-	59.8	58.6
[ 37 ] LE	57.6	64.9	72.5	-	-	-	60.0	58.8
[ 37 ] OT	64.2	67.4	73.2	-	-	-	62.5	61.3
[ 21 ]	61.4	-	70.0	74.5	-	-	63.8	62.6

Appendix

Table A2 : Fine-grained results on the MultiBooked (Basque) dataset. Same notation as in Table A .

	Spans (F 1 )			T.SF 1	Dep.		Sent.
	Holder	Target	Exp	T.SF 1	UF 1	LF 1	NSF 1	SF 1
[ 32 ]	63.6	55.3	56.1	31.5		40.0	-	32.7
[ 37 ] NC	60.3	51.8	54.2	-	-	-	42.7	39.3
[ 37 ] LE	64.0	52.3	56.1	-	-	-	43.7	40.4
[ 37 ] OT	65.1	58.3	60.7	-	-	-	47.8	41.6
[ 21 ]	62.5	-	61.4	61.3	-	-	46.7	41.9

Appendix

Table A3 : Fine-grained results on the NoReC dataset.

	Spans (F 1 )			T.F 1	Dep.		Sent.
	Holder	Target	Exp.	T.F 1	UF 1	LF 1	NSF 1	SF 1
[ 32 ]	50.0	44.8	43.7	30.7	-	34.9	-	27.4
[ 37 ] NC	31.4	35.0	35.1	-	-	-	24.8	22.9
[ 37 ] LE	32.5	38.0	36.2	-	-	-	28.8	27.3
[ 37 ] OT	42.2	40.6	39.3	-	-	-	33.2	31.2
[ 21 ]	47.0	-	44.7	45.0	-	-	37.8	34.1

Appendix

Table A4 : Fine-grained results on the Darmstadt Universities dataset. Same notation as in Table A .

	Spans (F 1 )			T.SF 1	Dep.		Sent.
	Holder	Target	Exp	T.SF 1	UF 1	LF 1	NSF 1	SF 1
[ 32 ]	47.9	50.7	47.8	33.7		37.6	-	20.4
[ 37 ] NC	58.4	60.3	55.8	-	-	-	38.7	28.3
[ 37 ] LE	53.6	53.4	53.4	-	-	-	33.8	27.0
[ 37 ] OT	55.7	64.0	53.5	-	-	-	45.1	34.1
[ 21 ]	46.5	-	51.8	47.6	-	-	29.3	24.1

Appendix

Table A5 : Fine-grained results on the MPQA dataset.

		Spans (F 1 )			T.SF 1	Dep.		Sent.
		Holder	Target	Exp	T.SF 1	UF 1	LF 1	NSF 1	SF 1
h.first	A	66.2	77.7	76.2	53.0	50.0	49.0	41.8	41.2
	R	70.4	81.9	74.3	69.4	63.4	61.4	53.8	52.5
	B 1	73.1	80.9	75.3	73.8	66.8	65.2	51.7	51.3
	B 2	68.8	82.3	76.9	75.6	68.6	66.7	56.1	55.0
	B 3	65.7	80.8	75.6	75.8	67.4	65.6	54.0	53.0

Appendix

Table A6 : Fine-grained results on the OpeNER (English) dataset.

		Spans (F 1 )			T.SF 1	Dep.		Sent.
		Holder	Target	Exp	T.SF 1	UF 1	LF 1	NSF 1	SF 1
h.first	A	68.4	74.7	75.2	48.9	45.8	44.5	39.0	38.5
	R	75.9	79.7	71.3	63.7	58.5	56.4	47.0	45.6
	B 1	67.5	78.1	73.1	68.6	61.8	60.4	49.0	47.9
	B 2	70.6	79.7	77.0	71.5	64.8	62.9	57.7	56.5
	B 3	69.9	79.1	75.2	70.3	64.5	62.9	56.4	55.4

Appendix

Table A7 : Fine-grained results on the OpeNER (Spanish) dataset.

Encoding	Training		Inference
Encoding	Token/s	Sentences/s	Token/s	Sentences/s
A	3792.25	224.78	4454.28	268.80
R	3818.01	225.91	4491.96	270.72
B 1	3831.82	226.45	4409.86	266.07
B 2	3796.78	224.57	4502.59	271.32
B 3	3828.62	226.31	4486.74	270.80

Appendix

Table A8 : Model efficiency in terms of training and inference throughput (tokens/s, sentences/s)

Explore similar work

GHI: Graphormer over Conditioned Hypergraph Incidence for Aspect-Based Sentiment Analysis
May 21, 2026Yu Du, Wenlong Zhu, Xingze Li +3Hypergraph Neural Networks Aspect-Based Sentiment Analysis
Qiqihar University
Revisiting Semantic Role Labeling: Efficient Structured Inference with Dependency-Informed Analysis
May 4, 2026Sangpil Youm, Leah Jones, Bonnie J. DorrStructured Prediction
University of Florida, Gainesville, Florida
GiLT: Augmenting Transformer Language Models with Dependency Graphs
May 15, 2026Tianyu Huang, Yida Zhao, Chuyan Zhou +1Transformer Language Modeling
School of Information Science and Technology, ShanghaiTech University · Shanghai Engineering Research Center of Intelligent Vision and Imaging

May 21, 2026, cs.CLJ/K move · Enter open · S save

GHI: Graphormer over Conditioned Hypergraph Incidence for Aspect-Based Sentiment Analysis

Yu Du, Wenlong Zhu, Xingze Li, Chenglong Cao+2

Qiqihar University

Hypergraph Neural Networks Aspect-Based Sentiment Analysis Graph Transformer Hypergraph Learning

Open paper AskPDF arXiv

Aspect-based sentiment analysis (ABSA) requires models to bind sentiment evidence to the correct aspect, making it a natural testbed for fine-grained structural reasoning. We introduce GHI, a Graphormer-over-Conditioned-Hypergraph-Incidence framework that is designed as an incidence-based structural reasoning layer built on a bipartite topology. GHI represents diverse linguistic and semantic evidence as token--hyperedge incidence relations, allowing different structural signals to be incorporated through a unified interface. Extensive experiments on six standard ABSA benchmarks show that GHI outperforms all baselines on the SemEval domains, and multi-seed evaluations show stable improvements over strong DeBERTa. Further experiments show that with only 247M parameters, GHI approaches the performance of 11B Flan-T5 based methods on the ISE benchmark. Moreover, it demonstrates strong robustness on the challenging ARTS datasets, maintaining highly competitive performance where traditional models degrade. These results demonstrate that compact structural reasoning remains a valuable alternative to scale-driven approaches for fine-grained tasks.

Abstract

Explore similar work

GHI: Graphormer over Conditioned Hypergraph Incidence for Aspect-Based Sentiment Analysis

Revisiting Semantic Role Labeling: Efficient Structured Inference with Dependency-Informed Analysis

GiLT: Augmenting Transformer Language Models with Dependency Graphs