Please enable JavaScript.
Coggle requires JavaScript to display documents.
DocStruct Research (integration (processing by page (paragraphs by page,…
DocStruct Research
features work
new feats to add
algorithm for line spacing feature
tag specific features
result of previous span
some features to out layer
add basic float features
features to delete
text pattern
first letter class
last letter class
chars map reducing
first word of span
features adaptation
pointer with new features
horizontal position features to test
test each feature separately
features from all pages (Titles)
model
joint model (tag/paragraph)
possibility to change tag
choose priority on model
joint tagset
completely joint
partially joint
Tags model
books + sciPapers tagsets
GRU model
model tuning
dynamic learning rate
stabilize with CRF loss
weighted cost to sciEnPDF
pointer improvement for titles
titles tagset (TitleBeginning...)
convolutional
sciPapers model
classification model
object detection model
convolutional + feats
convolutional + character level
convolutionla + txt features
join convolutional and GRU models
paragraphs model
feature last symbol of previous span
language model
get basic model
Tables model experiments
integration
measure speed
processing by page
paragraphs by page
add to proto
get model for tags
using tables tag from Mark
memory estimations
tagset joining to use on any document
analyse
dataset
classes problems
keywords
authors/institutions/footnote
analyse results by classes
Info tag analyse
correction
weird correction
ideas to find bugs in datasets
manual correction of visible errors
annotation
automatic titles dataset
additional annotation
additional manual Titles corpora
genre specific fields classification
patents
add to corpora chinese patents
regression
regression tool tuning
evaluate regression
convert my output to json format
philosophic questions
using of classes in product
apply model to random documents
find part of span to use it in metadata
language independent models
spans sorting
started
finished
not started
blocked