Frame-Based Approach for Reference Metadata Extraction - Technologies and Applications of Artificial Intelligence - page 160

Tabl e 2. Average field error rates achieved by FBA and CRFs in the training set.
(5076 styles)

TrainingSet, 10,000 Author Title Journal Volume Issue Page Year Average

FBA 0.60 0.84 0.78 3.06 4.09 3.56 1.43 2.05

CRFs 0.20 0.30 2.60 17.30 8.40 15.20 6.10 7.15

Tabl e 3. Average field error rates achieved by FBA and CRFs in EndNote style set.
(5076 styles)

EndNoteSet, 20,000 Author Title Journal Volume Issue Page Year Average

2.29

FBA 0.82 1.13 0.94 3.34 3.96 4.51 1.35

CRFs 1.20 1.80 3.20 16.70 9.10 14.00 6.60 7.51

Tabl e 4. Average field error rates achieved by FBA, CRFs, and Chen et al. in BibPro
set. (6 styles)

BibProSet, 10,000 Author Title Journal Volume Issue Page Year Average

FBA 0.52 1.52 1.78 2.73 2.92 1.95 1.89 1.90

CRFs 2.50 4.00 2.30 4.70 5.00 1.80 18.50 5.54

Chen et al. 2012 1.77 5.07 7.80 2.41 6.62 1.90 0.90 3.78

Tabl e 5. Average field error rates achieved by FBA, CRFs, and Cortez et al. in FluxHS
set

FluxHS, 2,000 Author Title Journal Volume Issue Page Year Average

FBA 7.47 10.33 5.66 2.06 - 1.05 0.15 3.81

CRFs 10.30 14.20 3.20 11.20 - 10.70 10.50 8.58

1.00 5.60 7.00 2.30 - 0.50 0.40 2.40 Flux (Cortez et al. 2007)

Tabl e 6. Average field error rates achieved by FBA and CRFs in free style set

FreeStyleSet, 1,500 Author Title Journal Volume Issue Page Year Average

FBA 0.91 0.82 2.69 2.09 1.79 3.28 1.09 1.81

CRFs 9.90 28.90 14.60 19.70 29.40 22.10 14.10 19.80

similar to those of the training and the EndNote set, and achieves better results
over CRFs and the BibPro system of [2]. The overall field error rates of FBA,
CRFs, and BibPro system are 1.90%, 5.54%, 3.78%, respectively.
In Table 5, we compared the unsupervised knowledge-based approach pro-
posed by [4] (denoted as Flux) as well as the baseline CRFs in the FluxHS
dataset. Note that this dataset does not contain issues of a reference. The overall
field error rates achieved by FBA, CRFs and Flux are 3.81%, 8.58% and 2.40%,
respectively. The performance of FBA remains better than that of CRFs, but a
little worse than that of the Flux system. One possible explanation is that this
dataset is in the health science domain, and there are many special terms used