Organizations: Xinjiang Engineering Technology Research Center for Smart Education and Application, School of Computer Science and Technology, Xinjiang Normal University, Urumqi, 830054, Xinjiang, China. · Key Laboratory of Aerospace Information Security and Trusted Computing, School of Cyber Science and Engineering, Wuhan University, Wuhan, 430000,Hubei, China. · Hainan Bielefeld University of Applied Sciences, Central Avenue and East 13, Danzhou, 578001, Hainan, China. · Journalism & Communication College, Xinjiang University, No. 666 Shengli Road, Urumqi, 830046, Xinjiang, China.
In recent years, aspect-based sentiment analysis (ABSA) has made rapid progress and shown strong practical value. However, existing research and benchmarks are largely concentrated on high-resource languages, leaving fine-grained sentiment extraction in low-resource languages under-explored. To address this gap, we constructed the first Low-resource languages Aspect-based Sentiment Quadruple dataset, named LASQ, which includes two low-resource languages: Uzbek and Uyghur. Secondly, it includes a fine-grained target-aspect-opinion-sentiment quadruple extraction task. To facilitate future research, we designed a grid-tagging model that integrates syntactic knowledge. This model incorporates part-of-speech (POS) and dependency knowledge into the model through our designed Syntax Knowledge Embedding Module (SKEM), thereby alleviating the lexical sparsity problem caused by agglutinative languages. Experiments on LASQ demonstrate consistent gains over competitive baselines, validating both the dataset's utility and the effectiveness of the proposed modeling approach.
Figures & tables
Target
Aspect
Opinion
Sentiment
Multilingual
Low-resource
ASTE
✗
✓
✓
✓
✗
✗
MAMS
✗
✓
✓
✓
✗
✗
ACOS
✓
✓
✓
✓
✗
✗
DiaASQ
✓
✓
✓
✓
✓
✗
Uzbek
✗
✓
✗
✓
✗
✓
Vietnamese
✗
✓
✗
✓
✗
✓
Table 1: Comparison between our dataset and existing ABSA datasets.
Figure 1: Input and output examples of our dataset. The text in parentheses represents the corresponding English translation.
Figure 2: LASQ dataset construction flowchart.
Figure 3: An example of annotation on the brat platform and its saved ANN format annotation result (bottom right corner of the figure). The example in the image is in Uzbek and means "This game interface design is very user-friendly."
# Doc.
# Sent.
# Token
# Entity
# Relation
# Quad.
T
A
O
TA
TO
AO
UZ
Train
2,451
6,146
101,061
2,471
7,075
8,697
7,884
8,712
9,348
9,384
Dev
306
780
12,705
307
841
1,031
943
1,031
1,097
1,100
Test
307
744
12,561
307
955
1,130
1,058
1,130
1,239
1,244
Total
3,064
7,670
126,327
3,085
8,871
10,858
9,885
10,873
11,684
11,728
UY
Train
2,422
8,539
90,184
2,427
6,767
8,248
7,577
8,250
9,202
9,231
Table 2: Statistics of our dataset. T/A/O represents target, aspect and opinion, and TA/TO/AO represents target-aspect pair, target-opinion pair and aspect-opinion pair. (Doc.: Document, Sent.: Sentence, Quad.: Quadruple).
Figure 4: (a) Our grid-tagging system. (b) Grid-tagging system in grid. The example in the image is in Uzbek and means "This game interface design is very user-friendly."
Figure 5: The overall architecture of our model. ⨁ and ⨂ represent matrix concatenation and matrix multiplication. The example in the image is in Uzbek and means "This game interface design is very user-friendly."
Hyperparameter
Value
Learning Rate (LLMs)
1e-5
Learning Rate (Others)
1e-3
Epochs
15
Batch Size
1
hp
20
hs
512
Table 3: Detailed hyperparameter settings.
Entity (F1)
Relation (F1)
Quadruple
T
A
O
TA
TO
AO
P
R
F1
UZ
EC-ACOS
64.06
23.32
44.75
16.41
30.92
15.16
16.94
8.72
11.51
Span-ASTE
65.21
24.29
46.81
17.92
32.17
16.83
18.96
10.31
13.36
ParaPhrase
63.16
22.86
47.87
16.91
33.50
15.89
22.27
8.96
12.78
Meta-WP
75.62
25.35
54.51
20.57
40.40
19.29
25.52
11.13
15.50
MRM-STS
73.52
26.52
50.57
21.05
39.70
20.26
18.66
12.82
15.90
Table 4: Comparison of our method with baseline models on our LASQ dataset. T/A/O represents target, aspect and opinion, and TA/TO/AO represents target-aspect pair, target-opinion pair and aspect-opinion pair. Bold represents the best results for each column under different language settings.The values in parentheses represent the variance of the results under three random numbers. ZS represents the zero-shot settings, and ICL represents the in-context learning settings.
Entity (F1)
Relation (F1)
Quadruple
T
A
O
TA
TO
AO
P
R
F1
UZ
Our Model
75.90
28.44
54.86
22.27
40.79
20.45
22.71
13.82
17.18
w/o POS
76.43
26.56
54.43
21.59
40.60
19.44
24.02
12.54
16.48
w/o Dependency
75.95
25.04
54.42
19.77
40.37
18.67
22.53
12.55
16.12
w/o SKEM
77.44
26.12
55.19
21.21
42.41
20.54
21.16
12.73
15.90
Random POS
76.32
24.55
53.77
19.96
40.67
18.65
24.00
12.005
16.04
Table 5: Ablation experiment.
Figure 6: Our method performs error analysis on quadruple extraction.
Case 1 (Gmail) : Gmail interfeysi qulay, lekin sinxronlash ba’zan sekin.
Table 6: Case study on Uzbek reviews about Google applications: effect of syntactic knowledge.
Training
Inference
(doc/s)
(doc/s)
UZ
EC-ACOS
5.60
14.35
Span-ASTE
5.15
14.79
ParaPhrase
10.36
9.74
Meta-WP
21.31
76.50
MRM-STS
5.06
13.30
Table 7: Efficiency analysis of our model and baseline models. “doc/s” represents the number of sentences processed per second.
Figure 7: We present the results of a more in-depth analysis of our LASQ dataset. Figure (a) shows the statistics for document length; Figure (b) shows the statistics for quadruples; Figure (c) shows the proportion of cross-sentence relations; and Figure (d) shows the proportion of entities with different morphological complexities.