LLM-based universal information extraction (UIE) methods often rely on additional information beyond the original training data, which increases training complexity yet often yields limited gains. To address this, we propose ProUIE, a Macro-to-Micro progressive learning approach that improves UIE without introducing any external information. ProUIE consists of three stages: (i) macro-level Complete Modeling (CM), which learns NER, RE, and EE along their intrinsic difficulty order on the full training data to build a unified extraction foundation, (ii) meso-level Streamlined Alignment (SA), which operates on sampled data with simplified target formats, streamlining and regularizing structured outputs to make them more concise and controllable, and (iii) micro-level Deep Exploration (DE), which applies GRPO with stepwise fine-grained rewards (SFR) over structural units to guide exploration and improve performance. Experiments on 36 public datasets show that ProUIE consistently improves unified extraction, outperforming strong instruction-tuned baselines on average for NER and RE while using a smaller backbone, and it further demonstrates clear gains in large-scale production-oriented information extraction.
Figures & tables
Figure 1: Comparison of UIE methods. (a) Traditional single-task methods train separate models for different extraction targets. (b) Generally LLM-based UIE unifies tasks but often introduces auxiliary signals beyond standard annotations. (c) ProUIE improves UIE without introducing any external information through a Macro-to-Micro progressive learning pipeline.
Figure 2: The framework of ProUIE.
Dataset
BERT-base
InstructUIE (11B)
GollIE (7B)
GollIE (13B)
GollIE (34B)
KnowCoder (7B)
UniversalNER (7B)
ProUIE (4B)
ACE2005
87.30
79.94
88.10
89.40
89.60
86.10
86.69
86.15
AnatEM
85.82
88.52
–
–
–
86.40
88.65
89.09
bc2gm
80.90
80.69
–
–
–
82.00
82.42
83.87
bc4chemd
86.72
87.62
–
–
–
–
89.21
91.97
bc5cdr
85.28
89.02
87.50
87.90
88.40
89.30
89.34
88.93
broad twitter
58.61
80.27
–
–
–
78.30
81.25
79.27
Table 1: NER results (F1, %). Highest in each row is bold ; second best is underlined .
Dataset
UIE
USM
InstructUIE(11B)
KnowCoder(7B)
ProUIE(4B)
ADE_corpus
–
–
82.31
84.30
84.30
CoNLL2004
76.00
78.84
78.48
73.30
73.17
GIDS
–
–
81.98
78.00
77.96
kbp37
–
–
36.14
73.20
69.45
NYT
–
–
90.47
93.70
96.41
NYT11 HRL
–
–
56.06
–
52.68
Table 2: RE results (F1, %). Highest in each row is bold; second best is underlined.
Dataset
BERT-base
UIE
USM
InstructUIE
ProUIE
ACE2005
72.50
73.36
72.41
77.13
70.65
CASIE
68.98
69.33
71.73
67.80
54.17
PHEE
–
–
–
70.14
67.87
Table 3: EE Trigger Identification results (F1, %).
Dataset
BERT-base
UIE
USM
InstructUIE
ProUIE
ACE2005
59.90
54.79
55.83
72.94
52.58
CASIE
60.37
61.30
63.26
63.53
60.75
PHEE
–
–
–
62.91
71.86
Table 4: EE Argument Extraction results (F1, %).
Method
NER
RE
EE
ProUIE (CM+SA+DE)
79.44
71.17
61.73
w/o SA (CM+DE)
78.19
68.68
57.92
w/o DE (CM+SA)
76.22
68.26
58.50
CM only
76.27
70.56
58.07
Table 5: Stage-wise ablation study of ProUIE.
Figure 3: Production-scale mobile query understanding: slot-level exact-match accuracy on six representative fields.
Task
Setting
P50
P70
P99
min
mean
max
min
mean
max
min
mean
max
RE
CM
58
70
75
76
83
85
100
114
161
CM+SA
10
13
20
14
18
35
39
74
126
EE
CM
45
81
150
62
105
170
115
171
199
CM+SA
16
28
34
18
37
42
48
90
99
Table 6: Token-length statistics of generated outputs.
Reward Granularity
NER (Avg)
RE (Avg)
EE (Avg)
Coarse-only
77.43
65.88
57.14
Fine-only
77.86
67.44
57.84
Coarse-to-fine SFR
78.25
67.69
58.40
Table 7: Coarse-to-fine reward ablation in DE.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Task
Dataset
∣Y∣
#Train
#Val
#Test
NER
ACE2005
7
15,000
971
1,060
broad_twitter_corpus
3
5,334
2,000
2,001
CoNLL2003
4
15,000
3,250
3,453
multiNERD
16
134,144
10,000
10,000
Ontonotes
18
30,000
8,528
8,262
polyglot-NER
3
393,982
10,000
10,000
Appendix
Table 8: Detailed dataset statistics. ∣Y∣ denotes the number of labels.
Figure 4: Prompt template for NER fine-tuning in CM stage.
Figure 5: Prompt template for RE fine-tuning in CM stage.
Figure 6: Prompt template for EE fine-tuning in CM stage.
Figure 7: Concise RE output for SA stage.
Figure 8: Concise EE output for SA stage.
Stage
NER
RE
EE
#Examples
CM (Phase 1)
✓
–
–
845,646
CM (Phase 2, 2:8)
✓
✓
–
224,745
CM (Phase 3, 3:3:4)
✓
✓
✓
24,979
SA (1:1)
–
✓
✓
18,000
DE (Phase 1)
✓
–
–
80,000
DE (Phase 2)
–
✓
–
50,000
Appendix
Table 9: Stage-wise data composition used in ProUIE.
Figure 9: A RE case study showing progressive refinement from CM to SA to DE.
Slot
Definition
body
The main subject of the query (i.e., what the user is searching for or focusing on).
action
The action expressed in the query.
environment
The environment or scenario described in the query.
location
The address/location constraint mentioned in the query.
number
The quantity associated with a body in the query.
source
The search source indicated in the query (e.g., where to search from).
Appendix
Table 10: Definitions of representative slots in the production-oriented evaluation.