Advancing Cross-Lingual Aspect-Based Sentiment Analysis with LLMs
and Constrained Decoding for Sequence-to-Sequence Models
Jakub
ˇ
Sm
´
ıd
1,2 a
, Pavel P
ˇ
rib
´
a
ˇ
n
1 b
and Pavel Kr
´
al
2 c
1
Department of Computer Science and Engineering, University of West Bohemia in Pilsen,
Univerzitn
´
ı, Pilsen, Czech Republic
2
NTIS - New Technologies for the Information Society, University of West Bohemia in Pilsen,
Univerzitn
´
ı, Pilsen, Czech Republic
Keywords:
Cross-Lingual Aspect-Based Sentiment Analysis, Aspect-Based Sentiment Analysis, Large Language
Models, Transformers, Constrained Decoding.
Abstract:
Aspect-based sentiment analysis (ABSA) has made significant strides, yet challenges remain for low-resource
languages due to the predominant focus on English. Current cross-lingual ABSA studies often centre on
simpler tasks and rely heavily on external translation tools. In this paper, we present a novel sequence-to-
sequence method for compound ABSA tasks that eliminates the need for such tools. Our approach, which
uses constrained decoding, improves cross-lingual ABSA performance by up to 10%. This method broadens
the scope of cross-lingual ABSA, enabling it to handle more complex tasks and providing a practical, efficient
alternative to translation-dependent techniques. Furthermore, we compare our approach with large language
models (LLMs) and show that while fine-tuned multilingual LLMs can achieve comparable results, English-
centric LLMs struggle with these tasks.
1 INTRODUCTION
Sentiment analysis aims to understand and quantify
opinions expressed in text, playing a critical role in
applications like customer feedback analysis, social
media monitoring, and market research. Within this
field, aspect-based sentiment analysis (ABSA) fo-
cuses on extracting fine-grained sentiment elements
from text (Zhang et al., 2022). These elements in-
clude aspect term (a), aspect category (c), and senti-
ment polarity (p). For example, in the review “Tasty
soup”, these elements are “soup”, “food quality”,
and “positive”, respectively. Implicitly referenced as-
pect terms, as in “Delicious”, are frequently labelled
as “NULL”.
Initially, ABSA research focused on individ-
ual sentiment elements, e.g. aspect term extrac-
tion and aspect category detection (Pontiki et al.,
2014). Recent studies have shifted towards compound
tasks involving multiple elements, such as end-to-
end ABSA (E2E-ABSA), aspect category term ex-
traction (ACTE), and target aspect category detection
a
https://orcid.org/0000-0002-4492-5481
b
https://orcid.org/0000-0002-8744-8726
c
https://orcid.org/0000-0002-3096-675X
(TASD) (Wan et al., 2020). Table 1 shows the output
formats of these ABSA tasks.
Table 1: Output format for selected ABSA tasks for an input
review: “Tasty soup but pricey tea”.
Task Output Example output
E2E-ABSA {(a, p)} {(“soup”, POS), (“tea”, NEG)}
ACTE {(a, c)} {(“soup”, food), (“tea”, drinks)}
TASD {(a, c, p)} {(“soup”, food, POS), (“tea”, drinks, NEG)}
While ABSA research traditionally focuses on En-
glish, real-world applications demand multilingual
capabilities. However, annotating multilingual data
is costly and time-intensive. Although multilingual
pre-trained models have become standard for cross-
lingual transfer in natural language processing (NLP)
tasks (Hu et al., 2020), applying them to cross-
lingual ABSA presents challenges due to language-
specific knowledge requirements. These models are
usually fine-tuned on source language data and di-
rectly applied to target language data. However, they
might lack language-specific knowledge for ABSA
tasks involving user-generated texts with abbrevia-
tions, slang, and language-dependent aspects. A pos-
sible solution is using translated target language data
with projected labels, but its effectiveness depends on
Šmíd, J., P
ˇ
ribá
ˇ
n, P. and Král, P.
Advancing Cross-Lingual Aspect-Based Sentiment Analysis with LLMs and Constrained Decoding for Sequence-to-Sequence Models.
DOI: 10.5220/0013349400003890
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 17th International Conference on Agents and Artificial Intelligence (ICAART 2025) - Volume 2, pages 757-766
ISBN: 978-989-758-737-5; ISSN: 2184-433X
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
757

the quality of the translation and alignment.
Modern monolingual ABSA approaches use pre-
trained sequence-to-sequence models, framing com-
pound tasks as text generation problems. In contrast,
cross-lingual ABSA research remains limited, focus-
ing mainly on simple tasks and E2E-ABSA, with no
studies employing the sequence-to-sequence methods
that excel in monolingual ABSA.
Recent advancements in large language mod-
els (LLMs), such as GPT-4o (OpenAI, 2024) and
LLaMA 3(AI@Meta, 2024), have achieved remark-
able results across NLP tasks. However, fine-
tuned models outperform LLMs on compound ABSA
tasks (Zhang et al., 2024). Fine-tuned LLaMA-based
models lead in English ABSA (
ˇ
Sm
´
ıd et al., 2024), but
their cross-lingual performance remains unexplored.
The main motivation of this paper is the limited
research on compound cross-lingual ABSA tasks, the
absence of sequence-to-sequence approaches widely
used in monolingual ABSA, and the reliance on ex-
ternal translation tools in related work, which adds
complexity and potential error to the process. To ad-
dress these shortcomings in existing works, we in-
troduce a novel sequence-to-sequence method that
achieves favourable results for compound ABSA
tasks in cross-lingual settings without relying on ex-
ternal translation tools. Additionally, we explore the
capabilities of several LLMs for cross-lingual ABSA,
as their performance on this specific task has not been
thoroughly investigated.
Our main contributions are as follows: 1) We in-
troduce the first sequence-to-sequence approach for
compound cross-lingual ABSA tasks, which does not
rely on external translation tools. 2) We significantly
improve zero-shot cross-lingual ABSA performance
using constrained decoding. 3) We compare our
method to LLMs, specifically GPT-4o mini and fine-
tuned LLaMA 3 and LLaMA 3.1, showing that only
fine-tuned multilingual LLaMA 3.1 achieves compa-
rable results to our approach. 4) We conduct exper-
iments on benchmark datasets in five languages and
three compound ABSA tasks, achieving new state-of-
the-art results in both cross-lingual and monolingual
settings. To the best of our knowledge, we are the first
to examine two compound cross-lingual ABSA tasks
and the cross-lingual capabilities of LLMs for ABSA.
2 RELATED WORK
Cross-lingual ABSA research focuses on three main
tasks: aspect term extraction (Klinger and Cimi-
ano, 2015; Wang and Pan, 2018; Jebbara and Cimi-
ano, 2019), aspect sentiment classification (Lambert,
2015; Barnes et al., 2016; Akhtar et al., 2018), and
E2E-ABSA (Li et al., 2020; Zhang et al., 2021b;
Lin et al., 2023; Lin et al., 2024). Of these tasks,
only E2E-ABSA is compound task, i.e. it focuses
on extracting more than one sentiment element si-
multaneously. Early methods relied on translation
and word alignment tools like FastAlign (Dyer et al.,
2013), with quality improvements through instance
selection (Klinger and Cimiano, 2015) or constrained
translation (Lambert, 2015). Others used cross-
lingual embeddings trained on bilingual corpora for
language-independent ABSA (Lambert, 2015; Barnes
et al., 2016; Akhtar et al., 2018; Wang and Pan, 2018;
Jebbara and Cimiano, 2019). Recent work focus
on multilingual Transformer-based (Vaswani et al.,
2017) encoder-only models, such as mBERT (De-
vlin et al., 2019) and XLM-R (Conneau et al.,
2020) combined with machine translation, with ad-
ditional enhancements from parameter warm-up (Li
et al., 2020), distillation on unlabelled target language
data (Zhang et al., 2021b), contrastive learning (Lin
et al., 2023), and dynamic loss weighting (Lin et al.,
2024).
The latest monolingual ABSA research focuses
primarily on text generation, exploring converting la-
bels to natural language (Zhang et al., 2021c; Zhang
et al., 2021a), multi-tasking (Gao et al., 2022), gener-
ating tuples as paths of a tree (Mao et al., 2022), ele-
ment ordering (Hu et al., 2022b; Gou et al., 2023), and
tagging-assisted generation (Xianlong et al., 2023).
Research shows that fine-tuned models outper-
form non-fine-tuned LLMs in compound ABSA
tasks (Gou et al., 2023; Zhang et al., 2024), whereas
fine-tuned LLaMA-based models achieve state-of-
the-art results in English ABSA (
ˇ
Sm
´
ıd et al., 2024).
3 METHODOLOGY
This section presents our approach to addressing the
triplet task (TASD), which can be easily modified for
tuple tasks with minor adjustments. Figure 1 depicts
the proposed approach.
3.1 Problem Definition
Given an input sentence, the aim is to predict all sen-
timent tuples T = (a, c, p), each composed of an as-
pect term (a), aspect category (c), and sentiment po-
larity (p). Following prior works (Zhang et al., 2021a;
Gou et al., 2023), we convert elements (a, c, p) into
natural language (e
a
, e
c
, e
p
). For instance, we trans-
late the “neutral” sentiment polarity into “ok” and the
“NULL” aspect term into “it”, as shown in Figure 1.
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
758

Source
Language
Dataset
Target
Language
Dataset
The place is small and cramped but the food is
fantastic.
(a, c, p): (place, AMBIENCE#GENERAL, negative)
(a, c, p): (food, FOOD#QUALITY, positive)
Example
The place is small and
cramped but the food is
fantastic. | [A] [C] [P]
Input
Model
[A] place [C] ambience
general [P] bad [;] [A]
food [C] food quality [P]
great
Output
la verdad es que ha estado genial.El lugar muy
acogedor de un gusto impecable.
(a, c, p): (NULL, RESTAURANT#GENERAL, positive)
(a, c, p): (lugar, AMBIENCE#GENERAL, positive)
Example
[A] it [C] restaurant
general [P] great [;] [A]
lugar[C] ambience
general [P] great
Output
la verdad es que ha estado
genial.El lugar muy
acogedor de un gusto
impecable. | [A] [C] [P]
Input
Train Predict Source Target
Constrained
Decoding
Figure 1: Overview of the proposed framework, which includes converting input labels to natural language phrases, fine-
tuning on source language data, and making predictions on target language data using constrained decoding for enhancement.
3.2 Input and Output Building
To build our model’s inputs and outputs, we employ
element markers to represent each sentiment element:
[A] for e
a
, [C] for e
c
, and [P] for e
p
. These markers
prefix each element, forming the target sequence to-
gether. We also append these markers to the input
sequence to guide the model for correct output. We
follow the priority order e
a
> e
c
> e
p
recommended
in prior research (Gou et al., 2023). For example, we
create the following input-output pair:
Input (x): They offer a tasty soup | [A] [C] [P]
Output (y): [A] soup [C] food quality [P] great
For sentences with multiple sentiment tuples, we
use the symbol [;] to concatenate their target schemes
into the final target sequence. Different examples are
depicted in Figure 1.
3.3 Constrained Decoding
To prevent the fine-tuned model from generating as-
pect terms in the source language instead of the tar-
get language, we have developed scheme-guided con-
strained decoding (CD) (Cao et al., 2021), which en-
sures that generated elements match their respective
vocabulary sets by incorporating target schema infor-
mation. This approach is beneficial in few-shot mono-
lingual settings (Gou et al., 2023).
Constrained decoding dynamically adjusts candi-
date token lists based on the current state, enhancing
control and accuracy in the generation process. For
example, if the current token is ’[’, the next token
should be chosen from special terms: A, C, P, and
;. Additionally, it tracks previously generated output
and current terms, guiding the decoding of subsequent
tokens based on Table 2. Appendix 6 shows the algo-
rithm in more detail.
Table 2: Candidate lists of tokens. <eos> indicates the end
of a sequence, and “. . .” denotes arbitrary text.
Generated output Candidate tokens
[
. . . [A / [C / [P / [; ]
. . . [A] Input sentence, “it”
. . . [C] All categories
. . . [P] great, ok, bad
. . . [A] . . . Input sentence, “it”, [
. . . [C] . . . All categories, [
. . . [P] . . . great, ok, bad, <eos> [
. . . [A] . . . [ C
. . . [C] . . . [ P
. . . [P] . . . [ ;
. . . [;] [
. . . [;] [ A
3.4 Training
We fine-tune a pre-trained sequence-to-sequence
model with provided input-output pairs. Sequence-to-
sequence models consist of two components: the en-
coder, which transforms input sequence x into a con-
textualized sequence e, and the decoder, which mod-
els the conditional probability distribution P
Θ
(y|e) of
the target sequence y based on the encoded input e,
where Θ represents the model’s parameters. At each
decoding step i, the decoder generates the output y
i
using previous outputs y
0
, . . . , y
i−1
and the encoded
input e. During fine-tuning, we update all model pa-
rameters and minimize the log-likelihood as
L = −
n
∑
i=1
log p
Θ
(y
i
|e, y
<i
), (1)
where n is the length of the target sequence y.
Advancing Cross-Lingual Aspect-Based Sentiment Analysis with LLMs and Constrained Decoding for Sequence-to-Sequence Models
759

Table 3: Dataset statistics for each language. POS, NEG and NEU denote the number of positive, negative and neutral
examples, respectively.
En Es Fr Nl Ru Tr
Train
Sentences 1,800 1,863 1,559 1,549 3,289 1,108
Triplets 2,266 2,455 2,276 1,676 3,697 1,386
Categories 12 12 12 12 12 12
POS/NEG/NEU 1,503/672/91 1,736/607/112 1,045/1,092/139 969/584/124 2,805/641/250 746/521/119
NULL aspects 569 700 694 513 821 135
Dev
Sentences 200 207 174 173 366 124
Triplets 241 265 254 184 392 149
Categories 11 11 12 11 12 10
POS/NEG/NEU 154/77/10 189/67/8 115/120/15 94/62/28 298/68/26 74/65/10
NULL aspects 58 83 66 64 109 15
Test
Sentences 676 881 694 575 1,209 144
Triplets 859 1,072 954 613 1,300 159
Categories 12 12 13 13 12 11
POS/NEG/NEU 611/204/44 750/274/48 441/434/79 369/211/33 870/321/103 104/49/6
NULL aspects 209 341 236 219 325 0
4 EXPERIMENTS
We report results with a 95% confidence interval from
5 runs with different seeds. The primary evaluation
metric is the micro F1-score, the standard metric in
ABSA research. We consider a predicted sentiment
tuple correct only if all its elements exactly match the
gold tuple.
4.1 Tasks and Dataset
We evaluate our method on the E2E-ABSA, ACTE,
and TASD tasks (see Table 1 for task details).
We perform experiments on the standard
SemEval-2016 dataset (Pontiki et al., 2016) with
restaurant reviews in English (en), Spanish (es),
French (fr), Dutch (nl), Russian (ru), and Turkish
(tr), with provided training and test sets. We split the
training data into a 9:1 ratio to create a validation
set. We consider English as the source language and
other languages as the target ones. Table 3 shows the
data statistics for each language.
4.2 Prompts for LLMs
Figure 2 shows the prompt for LLMs for the TASD
task, including one example for few-shot settings for
Spanish. This prompt can be adapted for various
tasks by excluding the unnecessary sentiment element
for the specific task, such as the sentiment polarity
for the ACTE task. For few-shot prompts, examples
are drawn from the first 10 examples of the training
dataset in the respective language.
4.3 Compared Methods
We compare our method with constrained decoding
against models without it. For E2E-ABSA—the only
task with available related work—we evaluate against
approaches using decoder-only models. One such
method (Li et al., 2020) operates in a true zero-shot
setting, training only on source language data with-
out machine translation. Other methods integrate ma-
chine translation with additional enhancements, in-
cluding alignment-free projection and aspect code-
switching for interchanging aspect terms between lan-
guages with distillation on unlabelled target language
data (Zhang et al., 2021b), contrastive learning (Lin
et al., 2023), and dynamically weighted loss com-
bined with anti-decoupling to improve semantic infor-
mation utilization and address class imbalances (Lin
et al., 2024).
It is important to note that these methods use
slightly different task definitions and datasets. Specif-
ically, previous work employs decoder-only models,
excludes implicit aspect terms (“NULL”), and re-
stricts each aspect term to a single sentiment polar-
ity. For example, prior research (Zhang et al., 2021b)
reports 612 tuples in the English test set after filter-
ing “NULL” aspect terms and merging sentiment po-
larities for each aspect term, whereas we report 859
tuples. In contrast, our approach predicts “NULL”
aspect terms and allows multiple sentiment polarities
per aspect term, making the task inherently more chal-
lenging.
4.4 Experimental Details
We use the large mT5 (Xue et al., 2021), selected
based on related work for English (Zhang et al.,
2021c; Zhang et al., 2021a; Gao et al., 2022; Gou
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
760

According to the following sentiment elements definition:
- The "aspect term" refers to a specific feature, attribute, or aspect of a product or service on which a user can express an opinion.
Explicit aspect terms appear explicitly as a substring of the given text. The aspect term might be "null" for the implicit aspect.
- The "aspect category" refers to the category that aspect belongs to, and the available categories include: "ambience general",
"drinks prices", "drinks quality", "drinks style_options", "food prices", "food quality", "food style_options", "location general",
"restaurant general", "restaurant miscellaneous", "restaurant prices", "service general".
- The "sentiment polarity" refers to the degree of positivity, negativity or neutrality expressed in the opinion towards a particular
aspect or feature of a product or service, and the available polarities include: "positive", "negative" and "neutral". "neutral" means
mildly positive or mildly negative. Triplets with objective sentiment polarity should be ignored.
Please carefully follow the instructions. Ensure that aspect terms are recognized as exact matches in the review or are "null" for
implicit aspects. Ensure that aspect categories are from the available categories. Ensure that sentiment polarities are from the
available polarities.
Recognize all sentiment elements with their corresponding aspect terms, aspect categories, and sentiment polarity in the given
input text (review). Provide your response in the format of a Python list of tuples: 'Sentiment elements: [("aspect term", "aspect
category", "sentiment polarity"), ...]'. Note that ", ..." indicates that there might be more tuples in the list if applicable and must not
occur in the answer. Ensure there is no additional text in the response.
Input: """Comida abundante, buena relación calidad-precio si pides entrante+segundo se puede cenar por
unos 12 euros"""
Sentiment elements: [("Comida", "food style_options", "positive"), ("null", "food prices", "positive")]
Input: """La comida estuvo muy sabrosa."""
Output: Sentiment elements: [("comida", "food quality", "positive")]
Figure 2: Prompt for the TASD task with example input, expected output in a green box, and one demonstration in Spanish
enclosed in a dashed box. The demonstrations are used solely in few-shot scenarios.
et al., 2023) employing monolingual T5 (Raffel et al.,
2020), and the large mBART (Tang et al., 2020) to
evaluate our method across different architectures.
Both models are from the HuggingFace Transformers
library
1
(Wolf et al., 2020). We fine-tune the models
for all experiments over 20 epochs with a batch size
of 16 and employ greedy search for decoding. For
mT5, we use a learning rate of 1e-4 and the Adafac-
tor optimizer (Shazeer and Stern, 2018). For mBART,
we use a learning rate of 1e-5 and the AdamW opti-
mizer (Loshchilov and Hutter, 2017). These settings
were chosen based on consistent performance on val-
idation data across all languages and tasks.
We evaluate GPT-4o mini (OpenAI, 2024) us-
ing zero- and few-shot prompts. Additionally, we
fine-tune the 8B versions of LLaMA 3 (AI@Meta,
2024) and LLaMA 3.1 (Dubey et al., 2024), em-
ploying QLoRA (Dettmers et al., 2023) with 4-bit
NormalFloat quantization, a batch size of 16, a con-
stant learning rate of 2e-4, AdamW optimizer, LoRA
adapters (Hu et al., 2022a) on all linear Transformer
block layers, and LoRA r = 64 and α = 16. Utilizing
the zero-shot prompt shown in Figure 2, i.e. without
the demonstrations, we fine-tune the model for up to 5
epochs, selecting the best-performing model based on
validation loss. All experiments are conducted using
an NVIDIA A40 with 48 GB GPU.
1
https://github.com/huggingface/transformers
5 RESULTS
Table 4 presents the results. Some key observations
include the following:
1) Constrained decoding significantly improves
cross-lingual ABSA by up to 10% over baseline mod-
els, effectively mitigating the issue where the model
predicts aspect terms in English instead of the target
language (e.g. “place” instead of “lugar”). The im-
provement is most noticeable in Spanish and Russian.
Constrained decoding is unnecessary in monolingual
experiments as the problem does not occur.
2) MT5 generally outperforms mBART and bene-
fits more from constrained decoding.
3) GPT-4o mini performs relatively well for the
E2E-ABSA task, which excludes aspect categories.
However, its performance is notably lower on the
ACTE and TASD tasks, indicating that identifying
aspect categories is challenging for the model. In-
corporating few-shot prompts boosts performance by
10–20% in most cases.
4) Our approach consistently outperforms GPT-4o
mini across all tasks and languages. Notably, we
achieve approximately a 20% improvement on the
TASD task across all languages in zero-shot cross-
lingual settings. Even when GPT-4o mini is enhanced
with 10-shot prompts, our results remain significantly
superior. The biggest difference is generally for Rus-
Advancing Cross-Lingual Aspect-Based Sentiment Analysis with LLMs and Constrained Decoding for Sequence-to-Sequence Models
761

Table 4: F1 scores for zero-shot cross-lingual ABSA with English as the source language and other languages as target
languages compared to monolingual results and GPT-4o mini. The compared works have different models and E2E-ABSA
definitions. Bold results indicate significant improvements using constrained decoding (CD). Underlined results are the best
absolute results for each language and task in both monolingual and cross-lingual settings.
E2E-ABSA ACTE TASD
Es Fr Nl Ru Tr Es Fr Nl Ru Tr Es Fr Nl Ru Tr
Monolingual
mT5
w/o CD 74.4
±0.6
69.9
±0.5
71.6
±1.0
72.4
±0.2
60.1
±1.7
70.4
±0.7
63.7
±0.8
68.8
±0.5
73.2
±0.5
59.1
±0.5
65.8
±0.4
59.0
±0.6
62.9
±1.4
67.0
±0.9
54.1
±3.0
w/ CD 75.3
±0.6
69.8
±1.4
67.0
±0.4
72.2
±0.4
60.7
±1.1
69.9
±0.4
64.9
±0.5
62.9
±0.5
72.8
±1.0
60.4
±2.1
66.2
±0.5
58.9
±1.1
57.6
±0.5
66.4
±0.4
53.9
±1.5
mBART
w/o CD 73.0
±0.5
66.4
±1.1
68.9
±1.2
68.7
±1.6
56.0
±2.7
66.4
±1.6
61.1
±1.6
64.1
±1.2
70.9
±0.6
56.8
±2.2
62.9
±1.2
54.8
±0.9
57.6
±0.9
62.6
±0.7
49.3
±3.1
w/ CD 71.9
±1.3
64.0
±1.7
61.6
±1.0
66.2
±1.1
54.4
±2.3
66.8
±1.5
58.2
±1.2
58.0
±1.2
67.4
±0.3
55.3
±1.5
61.5
±1.4
52.4
±0.6
52.1
±1.0
60.1
±1.9
47.6
±2.7
GPT-4o mini
0-shot 55.7 52.5 45.8 48.3 47.9 34.8 31.7 27.9 31.7 27.4 32.4 31.8 29.1 31.0 32.2
10-shot 63.8 55.2 58.9 58.9 51.0 56.2 46.1 49.5 45.8 47.5 50.9 40.9 46.0 42.1 44.3
LLaMA 3 70.0
±2.0
63.1
±1.8
66.0
±1.2
60.7
±2.3
48.6
±2.0
59.8
±2.9
54.9
±1.2
58.1
±4.3
62.1
±1.7
46.3
±2.9
57.2
±1.7
48.2
±2.4
55.4
±2.8
53.7
±2.9
39.1
±3.2
LLaMA 3.1 77.4
±0.5
71.2
±0.5
74.0
±0.5
73.7
±0.3
59.2
±1.6
70.9
±0.6
66.8
±2.3
70.6
±0.8
75.3
±0.9
60.7
±0.5
65.8
±1.2
62.0
±0.6
65.7
±1.1
68.4
±1.0
58.6
±1.0
Cross-lingual
mT5
w/o CD 59.2
±0.5
57.8
±1.2
57.1
±0.9
56.4
±2.1
44.4
±1.4
52.5
±1.0
55.8
±0.7
52.3
±1.3
55.0
±2.7
41.4
±1.4
48.3
±0.5
50.4
±1.4
47.7
±1.1
48.6
±2.0
39.1
±3.6
w/ CD 69.3
±1.0
61.1
±1.2
60.8
±0.3
63.7
±1.3
48.9
±1.4
62.8
±1.4
57.5
±0.3
54.1
±0.2
60.4
±0.9
49.0
±0.9
57.6
±0.6
50.4
±0.8
50.4
±1.3
54.9
±2.0
43.8
±0.8
mBART
w/o CD 61.1
±2.6
49.4
±3.8
51.6
±2.7
57.1
±1.4
31.6
±3.9
52.5
±1.4
49.3
±1.5
44.5
±1.4
53.8
±1.5
31.1
±2.1
47.6
±1.9
39.6
±0.8
39.1
±0.9
48.5
±1.1
23.5
±2.6
w/ CD 61.7
±2.7
49.2
±4.1
50.1
±3.5
57.8
±1.8
30.3
±3.0
54.8
±0.4
49.2
±0.6
46.9
±0.9
55.9
±0.2
34.7
±1.1
51.1
±1.2
39.9
±0.6
38.9
±0.9
50.5
±0.7
27.3
±1.1
LLaMA 3 47.2
±5.0
41.7
±3.0
43.3
±3.5
53.1
±2.2
26.6
±9.7
44.8
±8.5
38.1
±3.5
40.0
±6.0
50.6
±4.1
29.9
±7.1
39.5
±7.5
31.5
±3.2
29.8
±5.7
46.5
±2.6
22.8
±7.2
LLaMA 3.1 73.4
±0.7
68.1
±0.5
64.2
±0.8
58.8
±1.0
48.1
±4.1
64.9
±1.1
60.9
±3.0
55.7
±1.5
59.2
±1.6
47.7
±2.1
61.3
±0.6
57.3
±1.2
54.8
±0.8
53.1
±1.2
41.4
±0.7
(Li et al., 2020) 67.1 56.4 59.0 56.8 46.2 - - - - - - - - - -
(Zhang et al., 2021b) 69.2 61.0 63.7 62.0 - - - - - - - - - - -
(Lin et al., 2023) 61.6 49.5 51.0 50.8 - - - - - - - - - - -
(Lin et al., 2024) 69.6 60.7 61.3 62.3 - - - - - - - - - - -
sian and French.
5) Our method consistently exceeds the perfor-
mance of fine-tuned LLaMA 3 across all tasks and
language combinations, likely due to LLaMA 3 being
primarily pre-trained for English.
6) LLaMA 3.1 extends language support to include
Spanish and French but lacks compatibility with Rus-
sian, Dutch and Turkish. It achieves the best results
in most tasks and languages in monolingual settings,
although mT5 performs similarly when accounting
for confidence intervals. In cross-lingual scenarios,
mT5 with constrained decoding usually outperforms
LLaMA 3.1 for Russian and Turkish and demon-
strates similar or slightly inferior performance for
other languages. While LLaMA 3.1 does not offi-
cially support Dutch, it likely benefits from linguistic
similarities with its supported languages.
7) Results for Turkish are consistently worse than
for other languages, which might be because it is the
the only language not from the Indo-European family.
As mentioned in Section 4, comparing our E2E-
ABSA results with prior research is challenging due
to methodological differences. Previous works ex-
clude implicit aspect terms, limit one sentiment po-
larity per aspect term, and use decoder-only models.
Moreover, prior approaches rely on external trans-
lation tools and can be affected by the translation
quality, whereas our method avoids external tools en-
tirely. In contrast, our approach predicts implicit as-
pect terms, allows multiple sentiment polarities per
aspect term, and avoids external tools entirely. De-
spite these challenges, our method with constrained
decoding achieves comparable results and proves. We
find constrained decoding to be more practical than
relying on external translation tools.
No prior research exists for ABSA tasks beyond
E2E-ABSA in cross-lingual settings, which serves as
one of the primary motivations for this paper.
5.1 Inference and Training Speed
Table 5 summarizes the average training time per
epoch and inference time per example (both abso-
lute in seconds and relative) for various models on the
TASD task, with English as the source language and
Spanish as the target language.
Table 5: Average absolute and relative training time per
epoch and inference time per example for different mod-
els on the TASD task, with English as the source language
and Spanish as the target language.
Model
Training Time Per Epoch Inference Time Per Example
Absolute [s] Relative Absolute [s] Relative
mT5 210 1.00 0.24 1.00
LLaMA 3.1 924 4.40 1.04 4.33
The mT5 model is the reference, with a relative
training time of 1.00. The LLaMA 3.1 model is sig-
nificantly slower, requiring 4.40 times the training
time of mT5 and a much higher inference time, 15.32
times that of mT5. This comparison indicates that
while larger models like LLaMA 3.1 may offer per-
formance gains, they come with substantial computa-
tional costs during both training and inference.
5.2 Recommendations
In summary, constrained decoding improves cross-
lingual results significantly. Overall, we recom-
mend using the mT5 model with constrained decod-
ing in most scenarios, while LLaMA 3.1 is prefer-
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
762

able for languages it specifically supports. The mT5
model consistently outperforms the English-specific
LLaMA 3 across all tasks and languages. Moreover,
mT5 delivers excellent results compared to the mul-
tilingual LLaMA 3.1 despite having approximately
seven times fewer parameters while also requiring
less training time per epoch and offering faster in-
ference. Additionally, fine-tuning LLaMA models
on consumer GPUs demands specialized techniques,
which can be a limiting factor. However, LLaMA 3.1
may be the preferred choice if hardware resources are
sufficient and training time and inference are not ma-
jor concerns.
5.3 Error Analysis
We perform an error analysis to understand the chal-
lenges of sentiment prediction better. Specifically, we
manually examine the predictions for the first 100 test
samples of the Spanish dataset in the TASD task, us-
ing the best-performing runs of LLaMA 3.1 and mT5
with and without constrained decoding. Figure 3 de-
picts the results of the error analysis.
aspect term category
polarity
0
10
20
30
40
50
60
Sentiment element
Number of errors
LLaMA 3.1
mT5 w/ CD
mT5 w/o CD
Figure 3: Number of error types for LLaMA 3.1 and mT5
with and without constrained decoding (CD) on the Spanish
target language and the TASD task.
The primary source of errors lies in aspect term
prediction, where the model often misses some aspect
terms, generates additional ones, or produces incom-
plete terms instead of full ones. Without constrained
decoding, the mT5 model may also generate text that
does not appear in the original review, alter its for-
mat, or use the source language instead of the target
language (e.g. “service” instead of “servicio”). Con-
strained decoding addresses these issues effectively,
reducing the number of errors.
Further challenges arise from less frequent as-
pect categories and inconsistent annotations, partic-
ularly for categories like “restaurant general” and
“restaurant miscellaneous”, which impact perfor-
mance. Another common error is the confusion be-
tween “restaurant prices” and “food prices” cate-
gories. Additionally, some categories appear in only
one or a few languages, such as “food general”,
found exclusively in the Dutch test set, limiting the
classifier’s ability to learn from other source lan-
guages.
Sentiment polarity prediction is generally less
challenging than other sentiment elements, with er-
rors primarily occurring in misclassifying the “neu-
tral” polarity.
6 CONCLUSION
This paper addresses three compound cross-lingual
ABSA tasks using sequence-to-sequence models and
constrained decoding without needing external trans-
lation tools. Through extensive experiments in
six languages, we emphasize the effectiveness of
constrained decoding in enhancing zero-shot cross-
lingual ABSA. The proposed approach offers a prac-
tical alternative to external translation tools, demon-
strating robustness and effectiveness across various
language pairs and models, and opens up new possi-
bilities for advanced cross-lingual ABSA. Addition-
ally, we compare our method to modern LLMs, find-
ing that older multilingual models outperform fine-
tuned English-centric and closed-source LLMs. We
show that fine-tuning multilingual LLMs boosts per-
formance significantly, surpassing smaller models in
supported languages.
Future research could explore multi-task learning,
where a single model is trained simultaneously on
multiple tasks, enabling a unified approach to han-
dling diverse ABSA challenges. Additional experi-
ments could examine various source-target language
pair combinations to assess cross-lingual adaptabil-
ity further. Finally, investigating tasks involving a
fourth sentiment element (opinion terms) would be
valuable, though current efforts are constrained by the
limited availability of annotated data, particularly in
languages other than English.
ACKNOWLEDGEMENTS
This work was created with the partial support of
the project R&D of Technologies for Advanced Dig-
italization in the Pilsen Metropolitan Area (Dig-
iTech) No. CZ.02.01.01/00/23 021/0008436 and by
the Grant No. SGS-2022-016 Advanced methods
Advancing Cross-Lingual Aspect-Based Sentiment Analysis with LLMs and Constrained Decoding for Sequence-to-Sequence Models
763

of data processing and analysis. Computational re-
sources were provided by the e-INFRA CZ project
(ID:90254), supported by the Ministry of Education,
Youth and Sports of the Czech Republic.
REFERENCES
AI@Meta (2024). Llama 3 model card.
Akhtar, M. S., Sawant, P., Sen, S., Ekbal, A., and Bhat-
tacharyya, P. (2018). Solving data sparsity for as-
pect based sentiment analysis using cross-linguality
and multi-linguality. In Walker, M., Ji, H., and Stent,
A., editors, Proceedings of the 2018 Conference of
the North American Chapter of the Association for
Computational Linguistics: Human Language Tech-
nologies, Volume 1 (Long Papers), pages 572–582,
New Orleans, Louisiana. Association for Computa-
tional Linguistics.
Barnes, J., Lambert, P., and Badia, T. (2016). Exploring dis-
tributional representations and machine translation for
aspect-based cross-lingual sentiment classification. In
Matsumoto, Y. and Prasad, R., editors, Proceedings of
COLING 2016, the 26th International Conference on
Computational Linguistics: Technical Papers, pages
1613–1623, Osaka, Japan. The COLING 2016 Orga-
nizing Committee.
Cao, N. D., Izacard, G., Riedel, S., and Petroni, F. (2021).
Autoregressive entity retrieval. In 9th International
Conference on Learning Representations, ICLR 2021,
Virtual Event, Austria, May 3-7, 2021. OpenRe-
view.net.
Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V.,
Wenzek, G., Guzm
´
an, F., Grave, E., Ott, M., Zettle-
moyer, L., and Stoyanov, V. (2020). Unsupervised
cross-lingual representation learning at scale. In Juraf-
sky, D., Chai, J., Schluter, N., and Tetreault, J., editors,
Proceedings of the 58th Annual Meeting of the As-
sociation for Computational Linguistics, pages 8440–
8451, Online. Association for Computational Linguis-
tics.
Dettmers, T., Pagnoni, A., Holtzman, A., and Zettlemoyer,
L. (2023). Qlora: Efficient finetuning of quantized
llms. In Oh, A., Naumann, T., Globerson, A., Saenko,
K., Hardt, M., and Levine, S., editors, Advances in
Neural Information Processing Systems, volume 36,
pages 10088–10115. Curran Associates, Inc.
Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K.
(2019). BERT: Pre-training of deep bidirectional
transformers for language understanding. In Burstein,
J., Doran, C., and Solorio, T., editors, Proceedings
of the 2019 Conference of the North American Chap-
ter of the Association for Computational Linguistics:
Human Language Technologies, Volume 1 (Long and
Short Papers), pages 4171–4186, Minneapolis, Min-
nesota. Association for Computational Linguistics.
Dubey, A., Jauhri, A., Pandey, A., et al. (2024). The llama
3 herd of models.
Dyer, C., Chahuneau, V., and Smith, N. A. (2013). A
simple, fast, and effective reparameterization of IBM
model 2. In Vanderwende, L., Daum
´
e III, H., and
Kirchhoff, K., editors, Proceedings of the 2013 Con-
ference of the North American Chapter of the Associa-
tion for Computational Linguistics: Human Language
Technologies, pages 644–648, Atlanta, Georgia. Asso-
ciation for Computational Linguistics.
Gao, T., Fang, J., Liu, H., Liu, Z., Liu, C., Liu, P., Bao,
Y., and Yan, W. (2022). LEGO-ABSA: A prompt-
based task assemblable unified generative framework
for multi-task aspect-based sentiment analysis. In Cal-
zolari, N., Huang, C.-R., Kim, H., Pustejovsky, J.,
Wanner, L., Choi, K.-S., Ryu, P.-M., Chen, H.-H.,
Donatelli, L., Ji, H., Kurohashi, S., Paggio, P., Xue,
N., Kim, S., Hahm, Y., He, Z., Lee, T. K., Santus,
E., Bond, F., and Na, S.-H., editors, Proceedings of
the 29th International Conference on Computational
Linguistics, pages 7002–7012, Gyeongju, Republic
of Korea. International Committee on Computational
Linguistics.
Gou, Z., Guo, Q., and Yang, Y. (2023). MvP: Multi-view
prompting improves aspect sentiment tuple predic-
tion. In Rogers, A., Boyd-Graber, J., and Okazaki, N.,
editors, Proceedings of the 61st Annual Meeting of the
Association for Computational Linguistics (Volume 1:
Long Papers), pages 4380–4397, Toronto, Canada.
Association for Computational Linguistics.
Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang,
S., Wang, L., and Chen, W. (2022a). LoRA: Low-rank
adaptation of large language models. In International
Conference on Learning Representations.
Hu, J., Ruder, S., Siddhant, A., Neubig, G., Firat, O., and
Johnson, M. (2020). XTREME: A massively mul-
tilingual multi-task benchmark for evaluating cross-
lingual generalisation. In III, H. D. and Singh, A.,
editors, Proceedings of the 37th International Con-
ference on Machine Learning, volume 119 of Pro-
ceedings of Machine Learning Research, pages 4411–
4421. PMLR.
Hu, M., Wu, Y., Gao, H., Bai, Y., and Zhao, S.
(2022b). Improving aspect sentiment quad prediction
via template-order data augmentation. In Proceed-
ings of the 2022 Conference on Empirical Methods in
Natural Language Processing, pages 7889–7900, Abu
Dhabi, United Arab Emirates. Association for Com-
putational Linguistics.
Jebbara, S. and Cimiano, P. (2019). Zero-shot cross-lingual
opinion target extraction. In Burstein, J., Doran, C.,
and Solorio, T., editors, Proceedings of the 2019 Con-
ference of the North American Chapter of the Asso-
ciation for Computational Linguistics: Human Lan-
guage Technologies, Volume 1 (Long and Short Pa-
pers), pages 2486–2495, Minneapolis, Minnesota. As-
sociation for Computational Linguistics.
Klinger, R. and Cimiano, P. (2015). Instance selection im-
proves cross-lingual model training for fine-grained
sentiment analysis. In Proceedings of the Nine-
teenth Conference on Computational Natural Lan-
guage Learning, pages 153–163, Beijing, China. As-
sociation for Computational Linguistics.
Lambert, P. (2015). Aspect-level cross-lingual sentiment
classification with constrained SMT. In Zong, C. and
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
764

Strube, M., editors, Proceedings of the 53rd Annual
Meeting of the Association for Computational Lin-
guistics and the 7th International Joint Conference on
Natural Language Processing (Volume 2: Short Pa-
pers), pages 781–787, Beijing, China. Association for
Computational Linguistics.
Li, X., Bing, L., Zhang, W., Li, Z., and Lam, W. (2020). Un-
supervised cross-lingual adaptation for sequence tag-
ging and beyond. arXiv preprint arXiv:2010.12405.
Lin, N., Fu, Y., Lin, X., Zhou, D., Yang, A., and Jiang,
S. (2023). Cl-xabsa: Contrastive learning for cross-
lingual aspect-based sentiment analysis. IEEE/ACM
Transactions on Audio, Speech, and Language Pro-
cessing, 31:2935–2946.
Lin, N., Zeng, M., Liao, X., Liu, W., Yang, A., and Zhou,
D. (2024). Addressing class-imbalance challenges
in cross-lingual aspect-based sentiment analysis: Dy-
namic weighted loss and anti-decoupling. Expert Sys-
tems with Applications, 257:125059.
Loshchilov, I. and Hutter, F. (2017). Decoupled weight de-
cay regularization. arXiv preprint arXiv:1711.05101.
Mao, Y., Shen, Y., Yang, J., Zhu, X., and Cai, L. (2022).
Seq2Path: Generating sentiment tuples as paths of a
tree. In Muresan, S., Nakov, P., and Villavicencio,
A., editors, Findings of the Association for Compu-
tational Linguistics: ACL 2022, pages 2215–2225,
Dublin, Ireland. Association for Computational Lin-
guistics.
OpenAI (2024). GPT-4o. Accessed November 2024.
Pontiki, M., Galanis, D., Papageorgiou, H., Androutsopou-
los, I., Manandhar, S., AL-Smadi, M., Al-Ayyoub, M.,
Zhao, Y., Qin, B., De Clercq, O., Hoste, V., Apid-
ianaki, M., Tannier, X., Loukachevitch, N., Kotel-
nikov, E., Bel, N., Jim
´
enez-Zafra, S. M., and Eryi
˘
git,
G. (2016). SemEval-2016 task 5: Aspect based sen-
timent analysis. In Bethard, S., Carpuat, M., Cer,
D., Jurgens, D., Nakov, P., and Zesch, T., editors,
Proceedings of the 10th International Workshop on
Semantic Evaluation (SemEval-2016), pages 19–30,
San Diego, California. Association for Computational
Linguistics.
Pontiki, M., Galanis, D., Pavlopoulos, J., Papageorgiou,
H., Androutsopoulos, I., and Manandhar, S. (2014).
SemEval-2014 task 4: Aspect based sentiment analy-
sis. In Nakov, P. and Zesch, T., editors, Proceedings
of the 8th International Workshop on Semantic Evalu-
ation (SemEval 2014), pages 27–35, Dublin, Ireland.
Association for Computational Linguistics.
Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S.,
Matena, M., Zhou, Y., Li, W., and Liu, P. J. (2020).
Exploring the limits of transfer learning with a uni-
fied text-to-text transformer. The Journal of Machine
Learning Research, 21(1):5485–5551.
Shazeer, N. and Stern, M. (2018). Adafactor: Adaptive
learning rates with sublinear memory cost. In Dy, J. G.
and Krause, A., editors, Proceedings of the 35th In-
ternational Conference on Machine Learning, ICML
2018, Stockholmsm
¨
assan, Stockholm, Sweden, July
10-15, 2018, volume 80 of Proceedings of Machine
Learning Research, pages 4603–4611. PMLR.
ˇ
Sm
´
ıd, J., Priban, P., and Kral, P. (2024). LLaMA-
based models for aspect-based sentiment analysis. In
De Clercq, O., Barriere, V., Barnes, J., Klinger, R.,
Sedoc, J., and Tafreshi, S., editors, Proceedings of the
14th Workshop on Computational Approaches to Sub-
jectivity, Sentiment, & Social Media Analysis, pages
63–70, Bangkok, Thailand. Association for Computa-
tional Linguistics.
Tang, Y., Tran, C., Li, X., Chen, P.-J., Goyal, N., Chaud-
hary, V., Gu, J., and Fan, A. (2020). Multilin-
gual translation with extensible multilingual pretrain-
ing and finetuning. arXiv preprint arXiv:2008.00401.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones,
L., Gomez, A. N., Kaiser, L. u., and Polosukhin,
I. (2017). Attention is all you need. In Guyon,
I., Luxburg, U. V., Bengio, S., Wallach, H., Fer-
gus, R., Vishwanathan, S., and Garnett, R., editors,
Advances in Neural Information Processing Systems,
volume 30, page 6000–6010. Curran Associates, Inc.
Wan, H., Yang, Y., Du, J., Liu, Y., Qi, K., and Pan,
J. Z. (2020). Target-aspect-sentiment joint detec-
tion for aspect-based sentiment analysis. Proceed-
ings of the AAAI Conference on Artificial Intelligence,
34(05):9122–9129.
Wang, W. and Pan, S. J. (2018). Transition-based adver-
sarial network for cross-lingual aspect extraction. In
IJCAI, pages 4475–4481.
Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C.,
Moi, A., Cistac, P., Rault, T., Louf, R., Funtowicz,
M., Davison, J., Shleifer, S., von Platen, P., Ma, C.,
Jernite, Y., Plu, J., Xu, C., Le Scao, T., Gugger, S.,
Drame, M., Lhoest, Q., and Rush, A. (2020). Trans-
formers: State-of-the-art natural language processing.
In Liu, Q. and Schlangen, D., editors, Proceedings of
the 2020 Conference on Empirical Methods in Nat-
ural Language Processing: System Demonstrations,
pages 38–45, Online. Association for Computational
Linguistics.
Xianlong, L., Yang, M., and Wang, Y. (2023). Tagging-
assisted generation model with encoder and decoder
supervision for aspect sentiment triplet extraction. In
Bouamor, H., Pino, J., and Bali, K., editors, Proceed-
ings of the 2023 Conference on Empirical Methods in
Natural Language Processing, pages 2078–2093, Sin-
gapore. Association for Computational Linguistics.
Xue, L., Constant, N., Roberts, A., Kale, M., Al-Rfou,
R., Siddhant, A., Barua, A., and Raffel, C. (2021).
mT5: A massively multilingual pre-trained text-to-
text transformer. In Toutanova, K., Rumshisky,
A., Zettlemoyer, L., Hakkani-Tur, D., Beltagy, I.,
Bethard, S., Cotterell, R., Chakraborty, T., and Zhou,
Y., editors, Proceedings of the 2021 Conference of the
North American Chapter of the Association for Com-
putational Linguistics: Human Language Technolo-
gies, pages 483–498, Online. Association for Compu-
tational Linguistics.
Zhang, W., Deng, Y., Li, X., Yuan, Y., Bing, L., and Lam,
W. (2021a). Aspect sentiment quad prediction as para-
phrase generation. In Moens, M.-F., Huang, X., Spe-
cia, L., and Yih, S. W.-t., editors, Proceedings of
the 2021 Conference on Empirical Methods in Nat-
Advancing Cross-Lingual Aspect-Based Sentiment Analysis with LLMs and Constrained Decoding for Sequence-to-Sequence Models
765

ural Language Processing, pages 9209–9219, Online
and Punta Cana, Dominican Republic. Association for
Computational Linguistics.
Zhang, W., Deng, Y., Liu, B., Pan, S., and Bing, L.
(2024). Sentiment analysis in the era of large lan-
guage models: A reality check. In Duh, K., Gomez,
H., and Bethard, S., editors, Findings of the Associ-
ation for Computational Linguistics: NAACL 2024,
pages 3881–3906, Mexico City, Mexico. Association
for Computational Linguistics.
Zhang, W., He, R., Peng, H., Bing, L., and Lam, W.
(2021b). Cross-lingual aspect-based sentiment analy-
sis with aspect term code-switching. In Moens, M.-F.,
Huang, X., Specia, L., and Yih, S. W.-t., editors, Pro-
ceedings of the 2021 Conference on Empirical Meth-
ods in Natural Language Processing, pages 9220–
9230, Online and Punta Cana, Dominican Republic.
Association for Computational Linguistics.
Zhang, W., Li, X., Deng, Y., Bing, L., and Lam, W. (2021c).
Towards generative aspect-based sentiment analysis.
In Zong, C., Xia, F., Li, W., and Navigli, R., editors,
Proceedings of the 59th Annual Meeting of the As-
sociation for Computational Linguistics and the 11th
International Joint Conference on Natural Language
Processing (Volume 2: Short Papers), pages 504–510,
Online. Association for Computational Linguistics.
Zhang, W., Li, X., Deng, Y., Bing, L., and Lam, W.
(2022). A survey on aspect-based sentiment analysis:
tasks, methods, and challenges. IEEE Transactions on
Knowledge and Data Engineering.
APPENDIX
Constrained Decoding
Algorithm 1 shows the pseudo-code of proposed con-
strained decoding algorithm.
Data: Generated sequence, Input sentence
tokens, Special token map
Result: Candidate tokens for the next step
Get positions of “[” and “]” in the generated
sequence;
if no “[” tokens generated then
return “[”;
end
Count “[” and “]” tokens and find last “[”;
Get last generated token;
if fewer “]” than “[” and last generated
token is special then
return “]”;
end
if last generated token is “[” then
if last special token is “;” or none then
return “A”;
end
if last special token is “A” then
return “C”;
end
if last special token is “C” then
return “P”;
end
if last special token is “P” then
return “;”;
end
end
if last special token is “;” then
return “[”;
end
Initialize result as an empty list;
if last special token is “A” then
Add input sentence tokens and “it” to
result;
end
if last special token is “C” then
Add category tokens to result;
end
if last special token is “P” then
Add sentiment tokens to result;
end
if last generated token is not “]” then
Add “]” to result;
if last special token is “P” then
Add “⟨eos⟩” to result;
end
end
return result;
Algorithm 1: Proposed constrained decoding for the TASD
task.
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
766