Polarity Classification with Multitask Learning Syntax Augmentation

Muhammad Imran; Carlos Gómez-Rodríguez

cs.CLNov 5, 2026

Polarity Classification with Multitask Learning Syntax Augmentation

Authors: Muhammad Imran, Carlos Gómez-Rodríguez

Organizations: Universidade da Coruña, CITIC, Depto. de Ciencias de la Computación y Tecnologías de la Información, Campus de Elviña s/n, 15071 A Coruña, Spain

Abstract

Sentiment Analysis (SA) is a fundamental NLP task concerned with the identification and categorization of subjective information in text. While lexicon-based methods offer transparency, they often struggle with linguistic phenomena such as negation and intensification. Syntactic parsing has been integrated into SA pipelines to address this, but conventional parsers introduce significant computational overhead. Recent work shows that reformulating dependency parsing as a Sequence Labeling (SEL) task can improve efficiency. We propose and compare two strategies for integrating syntactic knowledge into a neural sentiment analysis model under unified Multi-Task Learning (MTL). The first, Standard-MTL, jointly learns dependency parsing (via sequence labeling) and sentiment polarity classification using a shared encoder with two task-specific heads. The second, Contrastive-MTL, provides implicit syntactic guidance through a contrastive auxiliary objective and a syntax-biased attention pooling mechanism, without requiring explicit label prediction. Both rely on syntactic annotations during training and eliminate need for hand-crafted compositional rules or external parsing at inference. We evaluate eleven datasets spanning seven languages---the MARC corpus (English, Spanish, French, German) and seven datasets from SemEval-2022 Task 10 (English, Spanish, Norwegian, Basque, Catalan)---using four established SEL encodings. With the relative encoding, Contrastive-MTL outperforms a syntax-naive baseline on 10 of 11 datasets and surpasses the external rule-based pipeline (SELSP) that parses at inference on all six datasets where the latter applies, while being 1.5--9 $\times$ faster at evaluation. Standard-MTL is more dataset- and encoding-dependent. These results show implicit syntactic supervision during training can match or exceed explicit parsing at inference without computational overhead.

Figures & tables

Review	Polarity
Clean , renewed rooms.	Positive
Food very basic.	Negative
What can we say.	No Annotation

Table 1: Polarity expression and corresponding polarity from the English training dataset in ( Agerri et al., 2013 )

Dataset	Data	Total	Discarded	Remaining
	Split	Rows	Rows	Rows
OpeNER en	Train	1744	344	1400
	Dev	249	51	198
	Test	499	92	407
OpeNER es	Train	1438	186	1252
	Dev	206	32	174

Table 2: Dataset statistics by split: total rows, empty labels, and remaining usable rows

Figure 1: Overall architecture of the proposed Multi-Task Learning Framework: Joint Dependency Parsing (Standard-MTL) and Contrastive Syntax-Guided Attention (Contrastive-MTL)

Hyperparameter	Value
Encoder	XLM-RoBERTa-large
Encoder init.	Pre-trained
Head init.	Random
Optimizer	AdamW
Enc. learning rate	$2\times 10^{-5}$
Head learning rate	$1\times 10^{-4}$

Table 3: Hyperparameters used to train the models

Figure 2: $\Delta$ F1 relative to the syntax-naive baseline: (a) Standard-MTL and (b) Contrastive-MTL.

Figure 3: Relative-encoding F1 for Baseline, SELSP, Standard-MTL and Contrastive-MTL across all eleven datasets. SELSP is available only for the six English and Spanish datasets; the white region with N/A marks the five datasets for which SELSP is not applicable.

Figure 4: Encoding-induced

\Delta

F1 relative to the syntax-naive baseline, for Standard-MTL and Contrastive-MTL.

Figure 5: Inference time comparison (log scale) for Baseline, SELSP, Standard-MTL and Contrastive-MTL.

Dataset	Encoding	Standard-MTL				Contrastive-MTL
		Acc	Prec	Rec	F1	Acc	Prec	Rec	F1
Darmstadt_unis	Baseline	0.822	0.546	0.553	0.549	0.822	0.546	0.553	0.549
	SELSP	0.414	0.581	0.414	0.399	0.414	0.581	0.414	0.399
	Brk7Bits	0.843	0.574	0.562	0.567	0.865	0.575	0.582	0.579
	BrkBased	0.867	0.576	0.584	0.580	0.911	0.608	0.612	0.610
	Absolute	0.867	0.576	0.581	0.578	0.867	0.580	0.579	0.579

Table 4: Performance comparison of explicit versus implicit syntactic augmentation. Standard Multi-Task Learning (Standard-MTL) and Contrastive Multi-Task Learning (Contrastive-MTL) are compared against a syntax-naive baseline and an external rule-based pipeline (SELSP) across all datasets and encoding schemes. SELSP results are only reported for the English and Spanish datasets, as the system does not support the remaining languages (Catalan, Basque, Norwegian, French, and German).

Dataset	Encoding	Contrastive-only MTL				Syntax-only MTL
		Acc	Prec	Rec	F1	Acc	Prec	Rec	F1
Darmstadt_unis	Baseline	0.822	0.546	0.553	0.549	0.822	0.546	0.553	0.549
	Brk7Bits	0.899	0.597	0.607	0.602	0.843	0.577	0.554	0.558
	BrkBased	0.878	0.595	0.586	0.589	0.844	0.565	0.564	0.563
	Absolute	0.856	0.578	0.568	0.570	0.867	0.578	0.581	0.579
	Relative	0.876	0.592	0.581	0.586	0.867	0.578	0.581	0.579

Table 5: Ablation study isolating the contributions of Contrastive Learning and Syntax-Guided Attention: Performance of Contrastive-only MTL (contrastive loss only) Versus Syntax-only MTL (syntax bias only)

Dataset	Method	Standard-MTL		Contrastive-MTL
		Train (s)	Eval (s)	Train (s)	Eval (s)
Darmstadt_unis	Baseline	206.061	0.678	206.061	0.678
Darmstadt_unis	SELSP	-	5.104	-	5.104
	Ours	325.681	0.686	282.294	0.674
MPQA	Baseline	486.676	1.930	486.676	1.930
MPQA	SELSP	-	17.056	-	17.056

Table 6: Computational cost comparison between explicit (Standard MTL) and implicit (Contrastive MTL) syntactic augmentation strategies, reporting training and evaluation times in seconds computed on NVIDIA RTX 6000 GPU. For each MTL variant, we report the mean over the four encodings (Brk7Bits, BrkBased, Absolute, Relative). SELSP is an external rule-based pipeline and is therefore not trained; its evaluation time is only reported for the English and Spanish datasets.

Figure 6: Inference-time attention over UD relations on OpeNER en (Contrastive-MTL, Relative encoding,

n\geq 20

). Left panel: mean softmax attention weight per relation, with

\pm 1

std across tokens; relations sorted by mean. Right panel: the learned training-time relation bias (Section 7.2 ), which is discarded at inference. Sentiment-relevant relations advmod , amod , and neg receive high attention; function-word relations det and poss receive low attention.

Relation	$n$	Attention	Bias
advmod	320	0.0762	$+0.673$
amod	375	0.0662	$-0.060$
dep	40	0.0580	$+0.660$
prep	31	0.0565	$-0.034$
cop	265	0.0537	$-0.107$
neg	63	0.0536	$+0.291$

Table 7: Inference-time attention and training-time learned bias per UD relation on OpeNER en (Contrastive-MTL, Relative encoding,

n\geq 20

Lexical relations (predictable from token identity)
Relation	$n$	Baseline	Contrastive-MTL	Contrastive-only MTL	Syntax-only MTL
det	586	0.804	0.988	0.988	0.930
punct	76	0.790	0.974	0.961	0.263
amod	375	0.627	0.880	0.899	0.307
nn	220	0.218	0.832	0.832	0.077
Context-dependent relations (require the sentence)

Table 8: Per-relation linear-probe accuracy on OpeNER en (Relative encoding. Lexical relations are largely predictable from the token itself; context-dependent relations require the surrounding sentence. Contrastive-MTL uses the attention bias and both contrastive losses; Contrastive-only MTL drops the bias; Syntax-only MTL drops the contrastive losses. Best per row in bold; relations with

n<20

omitted.

Sentiment Analysis

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

See all 133 papers

Sep 24, 2026

Multi-Task Learning by using Contextualized Word Representations for Syntactic Parsing of a Morphologically Rich Language

We address the challenge of syntactic parsing for Urdu, a morphologically rich language, and present state-of-the-art results for both constituency and dependency parsing. This paper offers four major contributions: 1) the conversion of the CLE-UTB phrase structure treebank into a dependency treebank by developing language-specific head-word and phrase-to-dependency label mapping rules; 2) a novel sequence labeling scheme that transforms the parsing task into a unified representation; 3) the training of contextualized word representations on a large 220 million tokens Urdu corpus collected from the web; and 4) development of parsing framework using two learning paradigms, single-task and multi-task learning. Several post-processing rules are applied to improve the quality of the automatically converted dependency structure treebank. The proposed sequence labeling scheme enables the use of a shared architecture that learns the syntactic structures from both grammatical structures simultaneously and hence improves generalization. Experiments show that the multi-task learning setup significantly enhances parsing performance, achieving an F1 score of 91.39 for constituency parsing (an improvement of 3.29 points) and a labeled attachment score of 85.69 for dependency parsing (an improvement of 1.49 points). These results demonstrate that learning cross-task representations provides measurable benefits and advances the state of syntactic parsing for Urdu.

Toqeer Ehsan, Miriam Butt, Sarmad Hussain +2

Reliable Intelligence Team, VTT Technical Research Centre of Finland Ltd., 02150 Espoo, Finland. · Department of Linguistics, University of Konstanz, 78464 Konstanz, Germany. · Center for Language Engineering, Al-Khawarizmi Institute of Computer Science, University of Engineering and Technology, 54890 Lahore, Pakistan. +2

Multi-Task Learning Low-Resource Language Processing Dependency Parsing Constituency Parsing

Ask PDF

Oct 8, 2026

Structured Sentiment Analysis Using Sequence Labeling as Dependency Graph Parsing

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.

Muhammad Imran, Ana Ezquerro, Carlos Gómez-Rodríguez +2

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

Sentiment Analysis Aspect-Based Sentiment Analysis

Ask PDF

May 6, 2026

PSK at SemEval-2026 Task 9: Multilingual Polarization Detection Using Ensemble Gemma Models with Synthetic Data Augmentation

We present our system for SemEval-2026 Task 9: Multilingual Polarization Detection, a binary classification task spanning 22 languages. Our approach fine-tunes separate Gemma~3 models (12B and 27B parameters) per language using Low-Rank Adaptation (LoRA), augmented with synthetic data generated by a large language model (LLM). We employ three synthetic data strategies (direct generation, paraphrasing, and contrastive pair creation) using GPT-4o-mini, with a multi-stage quality filtering pipeline including embedding-based deduplication. We find that per-language threshold tuning on the development set yields 2 to 4% F1 improvements without retraining. We also use weighted ensembles of 12B and 27B model predictions with per-language strategy selection. Our final system achieves a mean macro-F1 of 0.811 across all 22 languages, ranking 2nd overall of the participating teams, with 1st place finishes in 3 languages and top-3 in 8 languages. We also find that alternative architectures (XLM-RoBERTa, Qwen3) that showed strong development set performance suffered 30 to 50% F1 drops on the test set, highlighting the importance of generalization.

Srikar Kashyap Pulipaka

Independent Researcher

Topic pending

Ask PDF

Abstract

Sentiment Analysis

Explore similar work

Multi-Task Learning by using Contextualized Word Representations for Syntactic Parsing of a Morphologically Rich Language

Structured Sentiment Analysis Using Sequence Labeling as Dependency Graph Parsing

PSK at SemEval-2026 Task 9: Multilingual Polarization Detection Using Ensemble Gemma Models with Synthetic Data Augmentation