search

Orange-OpenSource / conllueditor

ConllEditor is a tool to edit dependency syntax trees in CoNLL-U format.
BSD 3-Clause "New" or "Revised" License
54 stars 17 forks source link
conll-u ud universaldependencies

readme

Editor for Treebanks in CoNLL-U format and Front-End for dependency parser servers

This Software is a tool which facilitates the editing of syntactic relations and morphological features of files in CoNLL-U format (http://universaldependencies.org/format.html). It uses a Java-based server and a HTML/CSS/Javascript based front-end. The editor loads the CoNLL-U file and saves changes to disk (and performs a git commit if the file is under git version control).

The editor provides the following functionalities:

Current version: 2.25.4 (see change history)

ConlluEditor can also be used as front-end to display the results of dependency parsing in the same way as the editor.

For more information see section Parser Front-End

In order to compare two files (e.g. a gold file with a predicted file) ConlluEditor provides

For more information see section File Comparison

Installation

Requirements

In order two compile the server, you also need

Docker

Alternatively, a recent version of Docker can be used to run the docker image (see section docker below)

Github Release

To avoid compilation, you can download the latest release, unzip the conllueditor-a.b.c.zip into a directory and start the server as described below.

License

Compilation

<script type="text/javascript" src="https://ajax.googleapis.com/ajax/libs/jquery/3.3.1/jquery.min.js">
./gui/lib/bootstrap-4.1.3/js/bootstrap.min.js
./gui/lib/bootstrap-4.1.3/js/bootstrap.bundle.js.map
./gui/lib/bootstrap-4.1.3/js/bootstrap.js.map
./gui/lib/bootstrap-4.1.3/js/bootstrap.bundle.min.js.map
./gui/lib/bootstrap-4.1.3/js/bootstrap.bundle.js
./gui/lib/bootstrap-4.1.3/js/bootstrap.js
./gui/lib/bootstrap-4.1.3/js/bootstrap.bundle.min.js
./gui/lib/bootstrap-4.1.3/js/bootstrap.min.js.map
./gui/lib/bootstrap-4.1.3/css/bootstrap-grid.css.map
./gui/lib/bootstrap-4.1.3/css/bootstrap.min.css.map
./gui/lib/bootstrap-4.1.3/css/bootstrap-grid.min.css.map
./gui/lib/bootstrap-4.1.3/css/bootstrap.css.map
./gui/lib/bootstrap-4.1.3/css/bootstrap-reboot.min.css
./gui/lib/bootstrap-4.1.3/css/bootstrap-reboot.css.map
./gui/lib/bootstrap-4.1.3/css/bootstrap-grid.min.css
./gui/lib/bootstrap-4.1.3/css/bootstrap-grid.css
./gui/lib/bootstrap-4.1.3/css/bootstrap.min.css
./gui/lib/bootstrap-4.1.3/css/bootstrap-reboot.css
./gui/lib/bootstrap-4.1.3/css/bootstrap.css
./gui/lib/bootstrap-4.1.3/css/bootstrap-reboot.min.css.map
./gui/lib/jquery-3.3.1.min.js
./gui/lib/popper.min.js
./gui/lib/popper.min.js.map
./gui/lib/jquery-ui-1.12.1
./gui/lib/jquery-ui-1.12.1/jquery-ui.min.css
./gui/lib/jquery-ui-1.12.1/LICENSE.txt
./gui/lib/jquery-ui-1.12.1/jquery-ui.theme.min.css
./gui/lib/jquery-ui-1.12.1/jquery-ui.structure.css
./gui/lib/jquery-ui-1.12.1/images
./gui/lib/jquery-ui-1.12.1/images/ui-icons_cc0000_256x240.png
./gui/lib/jquery-ui-1.12.1/images/ui-icons_444444_256x240.png
./gui/lib/jquery-ui-1.12.1/images/ui-icons_555555_256x240.png
./gui/lib/jquery-ui-1.12.1/images/ui-icons_777620_256x240.png
./gui/lib/jquery-ui-1.12.1/images/ui-icons_777777_256x240.png
./gui/lib/jquery-ui-1.12.1/images/ui-icons_ffffff_256x240.png
./gui/lib/jquery-ui-1.12.1/jquery-ui.structure.min.css
./gui/lib/jquery-ui-1.12.1/jquery-ui.css
./gui/lib/jquery-ui-1.12.1/jquery-ui.theme.css
./gui/lib/jquery-ui-1.12.1/AUTHORS.txt
./gui/lib/jquery-ui-1.12.1/index.html
./gui/lib/jquery-ui-1.12.1/jquery-ui.js
./gui/lib/jquery-ui-1.12.1/package.json
./gui/lib/jquery-ui-1.12.1/external
./gui/lib/jquery-ui-1.12.1/external/jquery
./gui/lib/jquery-ui-1.12.1/external/jquery/jquery.js
./gui/lib/jquery-ui-1.12.1/jquery-ui.min.js

Update ConlluEditor

to get bug corrections/new features, run

git pull
rm -rf target/test-classes
mvn install

Starting the server

On smaller machines, the memory management of the java VM (-Xmx... option) may have to be modified in bin/conlluedit.sh. The current value (-Xmx4g) is largely sufficient to load larger treebanks with up to 1,5M tokens.

ConlluEditor comes with a simple HTTP server:

bin/conlluedit.sh --rootdir  /path/to/ConlluEditor/gui treebank.conllu 8888

or (the option -r calculates the rootdir from the position of conlluedit.sh)

bin/conlluedit.sh -r treebank.conllu 8888

Point your navigator to http://localhost:8888 .

Using docker

If you prefer a docker image, you can use the following

Run the image in a docker container from the directory where your .conllu-file (and other files like lists of UPOS, XPOS etc) reside and replace

docker run --rm -t -d \
  --name conllueditor -p 8888:5555 \
    --user 1000:1000 \
    -v </absolute/path/to/datadir>:/data \
    --env filename=<yourfile>.conllu \
    jheinecke/conllueditor:latest

Note that the option -d causes the docker container to run in the background and error messages will not appear on screen unless you run docker logs conllueditor. If you want to see any messages the server outputs to the terminal, omit the -d.

Other parameters (shown below in section other options) can be given with --env <optionname>=value e.g. --env UPOS=uposfile.txt, --env UPOS=cpos.ud, --env XPOS=xpos.txt, --env language=cy, --env include_unused=1, --env deprels=deprels.ud, --env features=feat_val.ud, --env shortcuts=hortcuts.json, or --env compare=file.conllu, . However all files given, must reside in the </absolute/path/to/datadir> directory.

When finished, stop the docker container (it is removed automatically by the --rm option given):

docker stop conllueditor

Other options

If the .conllu file contains major tree errors (like cycles, no token with head 0 or head ids beyond the end of the sentence, warnings are writte to screen. Such errors may occur if an automatic pre-annotation tools do not work correctly. UD parsers like Udpipe do not produce invalid files. In this case please correct the errors using a text editor before loading the file into ConlluEditor.

Editing

More help on editing can be found by clicking the Help button.

If the server has been (re)started reload the page in your navigator.

Load the first sentence by clicking on read sentence: clicking on a word and then clicking on the head-word creates a dependency relation. An edit window opens to enter the relation a name. Existing relations can be renamed by clicking on their name. Clicking twice on a word deletes its eventual dependency relation and makes it root. To edit form, lemma etc. CTRL-click or double click on the word. For more help use the Help button. To delete a word, click ont it and hit the delete key.

Display modes

The sentence is shown as a dependency tree or as a flat graph (“dependency hedge”), morphological features can be shown or hidden with the features button, information of the MISC column can be shown with the misc button. multiword tokens (having n-m ids) are marked by a grey line spanning the multiword expression. If any UPOS/XPOS/deprel is not in the validation lists (specified with the --UPOS etc. options) it is shown in red.

Edit screen (tree graph)

The select button tree graph allows to toggle between a dependency tree layout to a more horizontal representation or a tabular view.

Choose view

The button fixed width displays the tree/hedge width a (configurable) fixed word width.

Edit screen (flat graph)

The table view is still experimental, all green fields can be edited: The shortcuts also work in table view, click on the ID of the word to modify via shortcuts (UPOS, XPOS, deprel, features) and type the short cut sequence. Cotrl-click on an ID does not unmakr the token after applying a shortcut. This allows to apply several shortcuts on the same word. Columns can be made larger or narrower by using the buttons + and -.

Edit screen (flat graph)

Editing words and dependency relations

Word editing window (CTRL-click on the word). If UPOS/XPOS/deprels are given to the server, autocompleting is proposed

Edit screen (flat graph)

(Basic) Dependency relation window (click on deprel)

Edit screen (flat graph)

For languages which are written from the right to the left like Arabic or Hebrew, activate the display right to left-check button

Edit screen (Arabic)

Empty Nodes

Empty nodes (having n.1 ids) are shown in a dashed box:

Empty nodes

The flat graph mode also displays enhanced dependencies. In this mode enhanced dependencies can be added/modified/deleted (activate edit enhanced dependencies). if the button show basic in enhanced is active, all enhanced dependency relations which are also a basic dependency, are displayed too.

Empty nodes

See section Enhanced Dependencies for more information

Split or merge adjacent words (editing tokenisation)

In order to split a word or join two (adjacent) words, use the modify button: the command split <wordid> inserts a new word to the right of <wordid>. This new word can then be edit, with a CTRL-click. The command join <wordid> merges the word with <wordid> with the following. This joined word gets the dependency relation of the word closer to root.

Whole sentences can be split with the sentsplit <wordid> command. The current sentences can be concatenated with the following sentence with sentjoin command.

In order to create a multiword token, use the compose <wordid> <length> command. Click on the multiword token bar (at the bottom of the dependency tree/graph to open a dialogue which allows to edit or delete the token (i.e. the n-m line).

All operations which change the tokenisation of the sentence will create a incoherent # text and forms warning. This is because the è# text = .... metadata must be coherent with the concatenation of forms (taken into accountSpacesAfter/SpaceAfterfields in the MISC column. Unless earlier versions, the# text ...is no longer updated automatically, but must be adapted manually using theedit metadata` button.

Commands to be used with modify button:

Show CoNLL-U, LaTeX and SD-parse format and file statistics

The buttons CoNLL-U, LaTeX and SD-parse open a window which contains the current sentence in the corresponding format. LaTeX output includes MWE units as well as enhanced dependencies. Enhanced dependencies which are identical too basic dependencies are shown in the LaTeX output if the button show basic in enhanced is activated. The download downloads the current image as a svg-file. A click on the filename on the top of the screens opens a windows with some file statistics:

Empty nodes

Searching

The search fields at the top of the screen can be used to search for forms, lemmas, UPOS, XPOS, deprels (or sequences of these), comments and sentences ids.

Other search modes can be chosen with the search select bar (top rop right)

Complex search and search and replace

This opens a search and search-and-display field. The search fields provides a simple language to find sentences with one or several nodes (see Mass Editing)

For instance

Upos:NOUN and Deprel:obj and head(Feat:Tense:Pres)

N.B. Feat:Tense=Pres or Feat:Tense:Pres can be used. However, since = is a hotkey to start validation on the current sentence : is easier to enter.

searches for a sentence which contains a word with Upos NOUN, Deprel obj and a head which has the feature Tense=Pres (click the search button)

Complex search

Adding a replace expression (as in Mass Editing) like

xpos:this(Upos)+"xpos" Feat:"Special:Present"

and click search & replace. This replaces all matching tokens in the first sentence where at least one token is matching the search condition:

Search and replace

In order to undo this replacement, click the undo button ( undo), to continue replacing, click search & replace again

Grew Match search

ConlluEditor implements a subset of search as defined in [Grew Match](http://universal.grew.fr/?corpus=UD_English-ParTUT@2.11">Gre Match). The subset includes

Some examples can be found in grewtests.txt

Matching subtrees

legacy, will be redrawn in a future version This mode allows you to input a tree (using _ as wildcards and regular expressions). Clicking the loop symbol searches for a sentence which match the subtree. The subtree must be a valid Conllu(plus) sentence with a single root. The _ character matches any value in the sentence. Columns are interpreted as regular expressions. (N.B. in this mode the ?-key is no longer a hot-key to display shortcuts) E.g.

# global.columns = ID   LEMMA   UPOS    FEATS   HEAD    DEPREL
1   _   AD.*    _   3   _
2   _   DET Gender=Fem  3   _
3   _   NOUN    _   0   _

matches any sentences wich contains a NOUN which as at least two dependants: A AD.* (i.e. ADP or ADJ) and a DET which has the feature Gender=Fem. Currently the word order is ignored.

Subtree search

The subtree can be entered using the sd-parse format

#sdparse
dans un contexte
case(contexte, dans)
det(contexte, un)

Subtree search

If two words have identical forms, position information must added

#sdparse
the cat sees the other cat
nsubj(sees, cat-2)
obj(sees, cat-6)

The sdparse format may contain _ as wildcards (in this case position information is obligatory)

#sdparse
_ _ _
case(_-3, _-1)
det(_-3, _-2)

Subtree search

In order facilitate the edition of the subtree, you can enter the Id of a word in the tree and click the import subtree? Doing so enters a partial tree of the word the current sentence and all its direct and indirect dependents.

Import subtree from current sentence

If the first line in the subtree window is a valid CoNLL-U Plus column definition, the imported subtree contains only the indicated columns. For example, if the subtree window contains (as a first line):

# global.columns = ID UPOS HEAD DEPREL

import subtree imports

Import subtree with selected columns

Mass Editing

A simple languages to modify tokens if a (complex condition is met) see Mass Editing

Metadata editing

The CoNLL-U format provides some special comment lines to indicate whether the current sentence is the beginning of a new document, new paragraph, the sentence itself, as well as its sentence id, translations (mostly into English) or transliterations. Clicking on edit metadata opens the Metadata dialogue. For translations, the translations must be prefixed with the language code as shown in the screen shot. If the words of the current sentence, contains transliteration information (MISC column, Translit field), the sentence transliteration can be initialized by clicking on init

Metadata edit

Enhanced Dependencies

Enhanced dependencies ([http://universaldependencies.org/format.html#syntactic-annotation]) in graphic mode can only be edited in flat mode. If the button edit enhanced dependencies is activated clicking on words creates enhanced dependency relations. Click on the dependency label to modify it or to delete the enhanced dependency relation. Alternatively, enhanced dependencies can be edited manually via the word edit menu.

Other annotation

A subset of the CoNLL-U Plus is supported. The CoNLL-U Plus file must have the standard 10 CoNLL-U columns and the additional columns defined in the first line

# global.columns = ID FORM LEMMA UPOS XPOS FEATS HEAD DEPREL DEPS MISC SEM:NE SEM:COREF

Valid CoNLL-U Plus with missing standard columns are currently rejected.

The additional columns are displayed under the graph/hedge and can be edited. The values in these columns are not interpreted (for instance BIO markings), they are just shown.

extra columns in tree view

extra columns in graph view

editing extra columns

CoNLL-U / CoNLLu Plus conversion

A simple conversion program is provided to convert any CoNLL-U (Plus) file into another. It takes as arguments the input file name and a comma separated list of output columns. Specified output columns absent in the input file will be replaced with _. If the output column list is absent, a standard CoNLL-U file is produced.

for instance

bin/conlluconvert.sh src/test/resources/test.conllup ID,FORM,DEPREL,HEAD,SEM:NE

generates

# global.columns = ID FORM DEPREL HEAD SEM:NE
# sent_id = fr-ud-dev_00001
# sentence 0
# text = Aviator, un film sur la vie de Howard Hughes.
1   Aviator root    0   B:Work
2   ,   punct   1   _
3   un  det 4   _
4   film    appos   1   _
5   sur case    7   _
6   la  det 7   _
7   vie nmod    4   _
8   de  case    9   _
9   Howard  nmod    7   B:Person
10  Hughes  flat:name   9   I:Person
...

No semantic/plausibility check is performed. E.g. bin/conlluconvert.sh <inputfile> FORM,DEPREL,HEAD will happily delete the ID column from the output file, so the HEAD column does not make much sense anymore.

Shortcuts

ConlluEdit uses a file gui/shortcuts.json which defines shortcuts to accelerate editing: These single letter keys change the UPOS/XPOS/deplabel/feature/misc of the active word to the defined value. To activate a word, click once on the word. Shortcuts can be single letters or a sequence of multiple letters:

{
    "upos": {
        "AV": "ADV",
        "C": "CCONJ",
        "D": "DET",
        "P": "PROPN",
        "I": "INTJ",
        "A": "ADJ",
        "N": "NOUN",
        "AP": "ADP",
        "R": "PRON",
        "S": "SCONJ",
        "T": "PART",
        "U": "NUM",
        "V": "VERB",
        "X": "AUX",
        ".": "PUNCT"
    },
   "feats": {
       ":ns": "Number=Sing",
       ":np": "Number=Plur",
       ":gf": "Gender=Fem",
       ...
    },
    "misc": {
       ":sa": "SpacesAfter=\\n",
       ":san": "SpaceAfter=No"
    },
    "deplabel": {
        "a": "amod",
        "av": "advmod",
        "c": "case",
        ....
    }
}

There is a timeout, so multi-letter shortcuts must be typed with maximally 700ms intervals (change this value with the --shortcutTimeout <milliseconds> option). This makes it possible to define shortcuts with different length and an identical sequence at the beginning (like above A and AV).

A personalised list (same format as gui/shortcuts.json) can be used with the --shortcuts option.

There is a list of predefined shortcuts which cannot be altered:

Multiuser/save/git

The ConlluEditor can be used by multiple annotators at the time, provided that no sentence is edited by more than one person at a time. To be on the safe side, start a server for every annotator on a different port/machine. After each modification the edited file is saved:

Validation

The ConlluEditor is able to load run a validation script on the current sentence. The programme and its arguments must be configured in a text file

script: /path/to/UniversalDependencies/tools/validate.py --lang cy --max-err 0 --level 5 {FILE}

{FILE} will be replaced with a file which contains the sentence to be validated in CoNLL-U format. This configuration file must be given to the server with the option --validator <filename>. The validation button will launch the validator on the current sentence.

Validation rules

In order to check somme language dependend validity (like adjective - noun agreement) a set of conditions can be specified (using the syntax as in Mass Editing) can be used.

E.g. if the current word has a nsubj Deprel, is a noun ou proper noun, has verb as head which is finite, than the verb must have the feature Person:3

Deprel:nsubj and (Upos:NOUN or Upos:PROPN) and head(Upos:VERB) and head(Feat:VerbForm=Fin) == head(Feat:Person:3)

More examples can be found in src/test/resources/validrules.txt

Run the validation:

./bin/validate.sh  src/test/resources/validrules.txt  src/test/resources/test.conllu

Server API (used by the GUI)

The GUI uses the following API when editing

upos (in the first example) can be replaced with form, lemma, xpos, head, deprel, feats, misc -for the latter two the value must be a comma separated list of Key=Value. In all cases the new version of the edited sentence is returned (in json). However if the sentid or id are invalid, an error messages (in json format) is returned.

For searching, the following API is implemented

To search backwards use cmd=findlemma true regex. Other columns can be found with findupos, findxpos, finddeprel, findfeat, findcomment (searches comments preceding a sentence) and findsentid (searches # sent_id). findform can be used for search forms. In difference to the other find-commands, findform searches for strings (no regex) in the sentence (across several words)

Parser Front-End

In order to display the CoNLL-U output of taggers/dependency parser servers, there the front-end provides a graphical user interface and sends the sentenced to be parsed to the back-end server such as UDPipe. You can either use a local instance of the UDPipe-Server, or the instance provided by the authors of UDPipe (http://lindat.mff.cuni.cz/services/udpipe/).

First edit the configuration file src/test/resources/udpipeserver.conf:

# configuration to  make a request to a UDPipe server

# the url to send raw text to in order to get a CoNLL-U result (HTTP POST)
parse: http://lindat.mff.cuni.cz/services/udpipe/api/process

# headers eventually needed for the server to respond correctly
parseheaders: Accept:application.json, ...

# the URL providing some information on the parser (HTTP GET)
info: http://lindat.mff.cuni.cz/services/udpipe/api/models

# the name of the field to contain the text to be parsed
txt: data

# other fields which need to be present in the call to the parser
other: model=english-ewt-ud-2.5-191206,tokenizer=,tagger=,parser=

# if the parser response is in json, give the path to the CoNLL-U result
jsonpath: result

Second, run the Front-End server:

./bin/parserclient.sh src/test/resources/udpipeserver.conf 3434

Open your browser on the given port: http://localhost:3434

Parser Front End Main view

The dependency tree can also be displays as a hedge

Parser Front End Main view flat

The parser info buttons provides information given by the parser server

Parser Front End Main Info

If the tokenizer segments the input in several sentences, all we be displayed.

The buttons CoNLL-U, LaTeX and SD-parse open a window which contains the current sentence in the corresponding format. LaTeX output includes MWE units as well as enhanced dependencies. The download downloads the current image as a svg-file.

File Comparison

ConlluEditor permits you to visualise the differences of two CoNLL-U files, as for instance a gold file and the output of a parsing tool. In order use the comparison mode start ConlluEditor with an option to indicate the gold file

./bin/conlluedit.sh --rootdir  /path/to/ConlluEditor/gui --compare gold.conllu predicted.conllu 8888

The dependency tree of the gold file is shown in grey underneath the tree of the file being edited. The results of some evaluation metrics for the current sentence are presented at the bottom of the header. Individual words of the edited file, which differ from the gold file are shown with a red border. If the mouse hovers over such a word, the corresponding lines from the edited file and the gold file are shown underneath the evaluation scores. The comparison mode works in the flat view too.

CoNLL-U File Comparison

Transliteration

The MISC columns provides a key Translit= for a transliteration of the word form into tha roman alphabet (including diacritics). The python script bin/transliterate.py will read any CoNLL-U file and add a Tranlist=<value> field to the MISC column if not already present. The transliteration rules are specified in bin/translit.json.

usage:

transliterate.py [-h] [--outfile OUTFILE] --infile INFILE --language LANGUAGE [--raw] [--overwrite] [--sentence]

arguments:

For the time being the following scripts/languages are covered

Other alphabets with a straight forward transliteration can be added easily added (please send me you updated translit.json file!). In some cases the transliteration can be erroneous, since this script does not take into account complex context depending transliterations. It is thought as a help for an initial version to be manually validated

Find similar Sentences

When editing large treebanks it is likely that verys similar or even identical sentences are in the train/validation or test corpora There is a tool which helps to identify these sentences.

The following command shows all sentences which appear more than once in any of the given files

./bin/findsimilar.sh options --group1 file1.conllu file2.conllu ...

To find overlapping sentences which are in a test-corpus and in dev- or train-corpus, use

./bin/findsimilar.sh options --group1 cy_ccg-ud-dev.conllu cy_ccg-ud-train.conllu --group2 cy_ccg-ud-test.conllu

options:

For Form the Levenshtein distance is calculated on characters, whereas for all other columns, the Levenshtein distance is calculate on a token basis.

Since every sentence will be compared once with every other sentence, this will take some time for CoNLL-U files wich many sentences and only a single thread and maximal distances != 0

Known bugs

Todo list

Reference

@inproceedings{heinecke2019,
  author = {Heinecke, Johannes},
  title = {{ConlluEditor: a fully graphical editor for Universal Dependencies treebank files}},
  year = {2019},
  booktitle = {{Universal Dependencies Workshop 2019}},
  address = {Paris},
  url = {https://github.com/Orange-OpenSource/conllueditor/},
  howpublished = {\url{https://syntaxfest.github.io/syntaxfest19/proceedings/papers/paper_55.pdf}},
}