1. Introduction
Plaid UD is an editor for building Universal Dependencies treebanks. It runs in a web browser and covers tokenization, annotation, and the dependency tree in one program.
Key features include:
-
Tokenization. Split a text into sentences, tokens, and multi-word tokens, by hand or with a tokenizer.
-
Annotation. Edit UPOS, XPOS, lemmas, and features in a grid, and draw the dependency tree and the enhanced graph above and below the words.
-
AI assistance. A connected parser fills in a first draft of a whole document, and an assistant answers questions about the corpus and proposes changes for you to approve. Machine output is marked until a person confirms it.
-
Search and bulk edit tools. Search every document with Grew patterns, count the results, and apply Grew rules to many sentences at once.
-
Collaboration. A project is shared with its members, each of whom may have reader, writer, or maintainer privileges, and all of them work on the same data on the server.
-
Audited edits. Every change is recorded with its author, and any document can be viewed or restored as it was at an earlier time.
-
CoNLL-U import and export. Bring in existing treebanks and export documents or whole projects. Empty nodes and the MISC column are not supported.
2. Getting Started
Plaid UD is served by your Plaid server. There is nothing to install.
Open the editor in your browser at the /ud/ path on that server.
If you are running Plaid locally with the default settings, that is:
http://localhost:8080/ud/
Sign in with the email address and password your project administrator gave you. Your session stays signed in until the token expires or you sign out. Your name at the top right opens your account, with your Profile and Sign out.
The editor has three levels you move between:
-
Projects: the list of all projects you can access. This is your home screen after signing in.
-
Documents: the texts inside one project. Beside the document list sit the project’s own tabs: Search, Guidelines, Assistant and Import and export for everyone, and Validation, Activity and Settings for maintainers.
-
A document’s tabs: Text Editor, Annotate, Comments, Export and Details, switched at the top of an open document.
The Assistant tab appears only when an assistant is running. The assistant can also be opened in a panel beside any screen (see The Panel).
3. Projects
A project in Plaid UD holds a collection of documents that share one annotation scheme. Universal Dependencies describes a text at three levels, and the editor follows the same model:
| Level | What it is |
|---|---|
Sentences |
Spans that tile the whole text, one per sentence. |
Tokens |
The orthographic words you see in the text, the UD tokens. A token may be a multi-word token (MWT), e.g. Spanish del standing for de + el. |
Words |
The UD syntactic words, the numbered rows of a CoNLL-U file. All annotation lives here: part-of-speech, lemma, features, and the dependency relation. An ordinary token is a single word. A multi-word token contains several. |
For most tokens there is exactly one word, so the two coincide and you can ignore the distinction. The split only matters for multi-word tokens, where one written token expands into several annotated words.
3.1. Creating and Opening Projects
From the Projects screen you can sort and search the list, and create a new project with New UD project. A new project is ready for Universal Dependencies annotation right away. Click it to open its documents.
4. Documents
Open a project to see its documents. As with projects, you can sort and search the list and create new documents. A new document starts empty. You give it a name, then add text in the Text Editor (next section) or by importing.
4.1. Importing CoNLL-U
If you already have annotated data, you can import it instead of typing it in.
Open the project’s Import and export tab, drop .conllu files on it, one or many, and press Import.
Each # newdoc boundary inside a file becomes its own document.
A project’s export zip can be dropped as it is.
A file too large for the server is not imported, and its row says so.
|
Warning
|
Import brings in tokens, sentence metadata, the dependency tree, and the enhanced graph from the DEPS column, but two CoNLL-U features are not stored and will be dropped on import: empty nodes and the MISC column. If you need those preserved, Plaid UD is not the right tool for that file. |
An enhanced relation whose head is an empty node is dropped along with the node. A relation with no head, or whose head is not a row of the sentence, is also dropped, and the import says how many.
5. Tokenizing Text
The Text Editor tab is where raw text becomes tokens. In a project not set up for UD it is read-only, as Annotate is, and a maintainer can set the project up from there.
5.1. Entering and Tokenizing Text
Paste or type your text into the editor, one sentence per line, and press Save.
Ctrl/Cmd+Z undoes typing and Ctrl/Cmd+Shift+Z redoes it.
If someone else saved the text after you started editing, your save adds your changes to theirs.
If you both changed the same passage, the save is refused with Changed elsewhere in the same passage., and Discard changes shows the saved text so you can make the edit again.
There are two ways to turn saved text into tokens:
-
Tokenize splits the text no token covers yet into tokens. A text with no sentences is first split into sentences, one per line. A text with sentences keeps them, with their tokens and annotations. It uses the project’s tokenizer locale, or else its language (see project settings). Each word gets its form as a machine-made lemma, and you add the other annotations and the tree yourself in the Annotate tab.
-
Parse does the whole job at once. Where your project has a parsing service connected, such as the bundled Stanza parser, it fills in sentences, tokens, words, every annotation, and the dependency tree. This gives you a first draft to correct by hand.
Each button opens a dialog naming the method it will use and its options (see Services). On a document that already has tokens, Tokenize keeps its sentences and tokens and adds tokens only for the text no token covers, such as punctuation left untokenized in another app. Clear tokens removes every token, along with every annotation on them. It asks first, with counts. Tokenize and Parse also ask first when their sentence breaks would delete annotations that cross them.
Tokens follow later edits to the text. A change inside a token keeps its annotations. Typing a space inside a token splits it, and its annotations stay on the part holding more of its letters. The other part is a new token with no annotations. A token all of whose letters you replace, by selecting it and typing or by pasting over it, is deleted with its annotations. Retyping a token’s own text is no change, and the token keeps its annotations. When you select two or more tokens and type or paste over them, letters you type again at either end are kept only as whole tokens: select came and and paste c, and both tokens go. Deleting the space between two tokens does not merge them, and a token is deleted only when all its text is. A sentence whose exact text you replace keeps its fields. A paste or a retype that covers a whole sentence and more text deletes it with its fields. A save that deletes tokens, or a sentence with fields, asks first, with counts.
A document of more than 15,000 words shows Long document at the top of this tab and the Annotate tab. Editing slows down past that length, so split such a document into shorter ones.
5.2. Working with Tokens
After tokenizing, the text is shown with each token as a clickable chip. A green border marks the first token of each sentence, and a teal chip marks a multi-word token.
Click any token to toggle whether a sentence starts there. Removing a boundary joins the sentence to the one above it. The joined sentence keeps the fields of the one above, and the fields of the one below are deleted, after a question that names them. Splitting a sentence removes the dependency relations that would cross the new boundary, and asks first if other annotations would go too.
Hover over a token to open its panel, where you can toggle Start of sentence, edit the token’s words, or delete the token. Splitting a token into several words makes a multi-word token, and merging them makes it an ordinary token again. Changing the forms but not the number of words keeps their annotations. Changing the number replaces the words, and asks first if annotations would be lost. You can also select a span of text to make a new token from it.
6. Annotation
The Annotate tab is where the real work happens. It shows each sentence as an interactive dependency tree above a grid of editable cells.
Sentences come 25 to a page, with a pager above and below them. A document reopens on the page you left it on, and a link to a particular sentence, such as a search result, turns to its page first. Sentence numbers belong to the document, so sentence 40 is sentence 40 on any page.
6.1. Tag and Lemma Editing
Each word has cells for:
| Column | Meaning |
|---|---|
Lemma |
The word’s dictionary form. An empty cell shows the word’s form in grey, which is not stored until you type it. |
XPOS |
Language-specific part-of-speech tag. |
UPOS |
Universal part-of-speech tag. |
Features |
Morphological features, as |
Click a row’s label to hide or show the row. The Features row starts hidden. Click a cell to edit it. The UPOS, XPOS and Features cells, and the dependency-relation labels in the tree, offer autocomplete from your project’s controlled vocabularies. By default any value is accepted, but a maintainer can restrict a list to its own values (see Suggestions or Rules).
Move around the grid with Tab, Enter, Escape and the arrow keys. Arrow Up out of the top row jumps into the dependency tree. The full list is in Keyboard Shortcuts.
6.2. Tree Editing
Above the grid, each sentence’s dependency tree is drawn as arcs between words.
-
Draw an arc: press and drag from one word to another. The word you start on becomes the head, and the word you release on becomes its dependent.
-
Mark the root: drag from the band across the top of the tree, where a root’s arc starts, down to a word. The band is a faint grey, and darker under the pointer.
-
Rename a relation: click an arc’s label to edit its dependency relation (
nsubj,obj, …), again with autocomplete. A label left empty stays as it was. -
Delete a relation: click its label, then the bin to the left of the box that opens, or press Shift+Backspace before typing anything.
-
Keyboard: from the grid, Ctrl/Cmd+D jumps focus into the tree’s labels. Tab and Shift+Tab walk between them. See Keyboard Shortcuts for the rest.
A word has one head, and the tree has no cycle. Drawing a relation to a word that already has a head moves it to the new head, keeping its label. A relation that would close a cycle is refused.
6.3. Enhanced Dependencies
Every sentence has an enhanced dependency graph beside its tree. The enhanced graph starts out equal to the tree, and you record only where it differs. The tree is drawn above the words and the enhanced relations below them.
-
Add a relation: hold Ctrl or Cmd as you start to draw an arc, and let go on a word or anywhere beneath one. That word keeps its head in the tree and gains a second one in the enhanced graph, as for a subject shared by two coordinated verbs.
-
Relabel a relation in the enhanced graph: draw over a relation of the tree with Ctrl or Cmd held, and type the new label, such as
nmod:offornmod. The tree’s relation is shown faded. Deleting the new label puts it back. -
Leave a relation out of the enhanced graph: Ctrl/Cmd+click a relation of the tree, or press Ctrl/Cmd+E on its label. The relation is shown faded. Do the same again to put it back.
Changes to the tree carry over to the enhanced graph.
A word given a new head, by hand or by a rewrite rule, keeps its relabel there: an obl relabelled obl:on moves to the new head as obl:on.
A second head the enhanced graph gives the word stays where it is.
On export, DEPS lists every relation of the enhanced graph for each word.
Empty nodes are not supported, so an elided predicate cannot be added, and gapping stays as the tree’s orphan relation.
6.4. Metadata
Plaid UD keeps notes about documents and about sentences, and edits them in two places. Notes about a document, such as its source, genre or licence, are on its Details tab. They stay in Plaid and are not written to the CoNLL-U export. Notes about a sentence are behind Edit metadata, in the row of actions below its grid.
Every sentence carries sent_id, CoNLL-U’s own identifier for it.
Other fields are a maintainer’s choice, under Settings → UD settings. A free translation is conventionally text_en, naming its language.
Fields save as you leave them, and clearing one removes it.
On export, sentence fields are written above their sentence, as # sent_id = … or # text_en = ….
On import, a comment line under # newdoc id, such as # source = …, becomes a field of the document’s first sentence, not of the document.
A sentence with no sent_id, such as one made by a split, is given one on export: the document’s name with an underscore for each space or slash, then its number, as in Field_notes-3.
6.5. Text Direction
A document’s Details tab also holds Text direction, which says which way its tokens are laid out. It is set to Automatic, which reads the document’s own text, so a treebank in Arabic, Hebrew or another right-to-left script is laid out right to left. Set it by hand for a document the text cannot speak for, such as a right-to-left language written in a Latin transliteration.
Direction affects the layout of the tokens, not the values in them, so a UPOS tag or an English lemma still reads left to right. Arrow keys follow the words: in a right-to-left sentence the left arrow moves forward.
6.6. Earlier Values in This Project
In a Lemma, XPOS or Features cell, Alt+Down lists the values this project has already used for a word like this one, with how many times each. The lemma cell looks up the word’s form. The XPOS and Features cells look up its lemma. Pick a value with the arrows and Enter, or click it. In Features the pick is added to the word’s features. Escape goes back to the cell.
6.7. Moving Between the Text and the Annotations
The Text Editor and the Annotate tab each open the other at the sentence you are on. In Annotate, Edit text under a sentence, or Alt+click on any word in it, opens the Text Editor at that sentence. In the Text Editor, Alt+click on a token opens Annotate at its sentence. The sentence you arrive at is briefly outlined.
6.8. Machine vs. Human Annotations
Annotations nobody has checked yet are marked, so you can see at a glance what is left to review. A violet, dashed mark means an automatic parser produced it. An amber, dashed mark means a member whose work the project reviews entered it (see Sharing and Permissions). Hovering a cell or a relation names its state and who produced it. When the parser gave its confidence, hovering also shows the percentage, and a value below the project’s confidence threshold (50% unless the project sets another) gets a heavier, darker mark (a thicker arc in the tree).
Any edit you make to a marked cell or relation confirms it and clears the mark, even typing the same value again. Moving through a cell with Tab or an arrow key, or leaving it with Escape, confirms nothing. If your own work is reviewed, the mark stays on what you enter or accept until someone else confirms it.
If a prediction is already correct, you don’t have to retype it:
-
Ctrl/Cmd+Enter accepts everything proposed on the word and moves to the next word.
-
Ctrl/Cmd+Backspace discards the parser’s unconfirmed work on the word and moves to the next word.
-
Ctrl/Cmd+Shift+↓ and Ctrl/Cmd+Shift+↑ jump to the next and previous word that still needs a look.
-
Accept predictions and Discard predictions, under each sentence, do the same for the whole sentence.
Gestures and marks, above the first sentence, opens a legend of these marks and every gesture in the editor.
6.9. Saving
A cell is saved when you leave it, and a relation when you draw or change it. There is no save button. A save the server has not answered after three seconds reads Saving… until it lands. If the server cannot be reached, saves are retried until it answers, and the page reads Can’t reach the server, retrying meanwhile, or Offline, retrying when the computer has no network. While a script or a service has the document locked, the page reads In use elsewhere, waiting, and your edits are saved once it is done.
Two people can have one document open, but neither sees the other’s edits as they are made. Values in cells of different words are all saved. When someone else changed the same cell first, the cell shows their value with yours under it: Yours: X · Enter to keep yours. Enter saves yours, and Escape or typing another value drops it. Other conflicting edits are refused with a message, and the page shows the latest version so you can redo your edit.
7. Export
The Export tab shows the current document as CoNLL-U, with Copy and Download.
A token followed by no space is marked SpaceAfter=No in the MISC column, the only value written there.
Text that no token covers, such as punctuation in a text tokenized in another app, is in # text but has no row. Tokenize in the Text Editor gives it its tokens.
A sentence that holds no tokens is left out of the file, and the tab counts such sentences. Tokenize gives them tokens too.
To export a whole project, use its Import and export tab, which packages every document into one ZIP of .conllu files.
8. Guidelines
The Guidelines tab is the project’s own annotation manual, where the conventions the annotators have agreed on are written down. The controlled value lists in the settings say which tags a column takes. Guidelines are for what a list cannot say, such as which analysis this treebank uses for a construction, or how a hard case was decided.
Everyone who can open the project can read the guidelines, and anyone who can annotate can write them. Each guideline has a title and a body, written in a rich-text editor or, with Markdown, as Markdown source. A short title naming the subject, with the rule itself in the body’s first line, makes a guideline easy to find for people and for the Assistant. If someone else saves a guideline while you have it open, your save is refused, and saving again overwrites theirs.
The Assistant reads the guidelines when it answers, and a pinned guideline is always sent to it in full. It can also draft a guideline, or change one, and you approve or discard the change like any other it proposes. Every change to a guideline is recorded in the project’s Activity. Guidelines are not part of a project export.
9. Assistant
The Assistant tab is a chat with a language model that knows your project. Ask it which words still have no lemma, whether any dependency relations look wrong, or to summarize the parts of speech across the corpus. It reads the documents, the search results, the project’s guidelines and the history to answer.
A line above each answer sums up the work behind it, such as "read 87 sentences in 40 documents · 2 searches · 52 steps". Opened, it lists each step the assistant took, and a step opens to what the assistant was given and what it got back.
When you ask the assistant to change something, it does not touch the data. It proposes a plan, which you approve with Approve as unverified or Approve as verified, or Discard. Every change is listed under its document and links to that sentence in the annotation editor, so a plan can be checked before it is approved. A change to something a person made or accepted is marked Accepted and counted separately. A change over a whole document or the whole project, such as confirming everything waiting in a document or replacing a lemma everywhere, is one row with its counts, and Show where lists the documents and a few examples.
Approved changes are written under your account. Approve as unverified records every value the assistant writes as its proposal, unconfirmed, as a service’s proposals are, to be confirmed later in the editor. Approve as verified records them as the assistant’s work accepted by you. A confirmation in the plan is made under either button. Tick Record as human-made to record them as your own work instead, and the card then offers a single Approve. A plan of only confirmations, deletions and other changes that write no value offers a single Approve as well. If your work is reviewed in this project, the second button reads Approve as your contribution, and what you approve with it waits for a reviewer, as anything you enter does. Once a plan is applied, its card says how it was approved, such as Approved as unverified. Reading needs Reader access, and applying a plan needs Writer.
A plan is Out of date once a sentence it changes has been edited, and can then only be discarded. Ask again to have it planned on the current text. A change over a whole document or the project is checked by what it finds: an edit elsewhere in the document leaves it standing, and one to a value it changes makes the whole plan Out of date, with what it finds now and where. Nothing is written. Asking for a new plan replaces one still waiting, and the older card reads Replaced. A plan that stops partway reads Partly applied, with how many of its changes were written, or for a parse how many sentences. Ask the assistant to finish it. A plan that only keeps values as they are reads Nothing to write, and can only be discarded. One plan makes at most 20,000 changes in at most 500 documents.
Conversations are private to you and kept per project, though your server’s administrator can read them. A reply arrives even if you close the tab, and is there when you open the conversation again. Stop ends a turn early, even while the model is still working on its answer, and Retry asks the question again.
A conversation can be downloaded or copied as Markdown, or saved with Download as web page as one HTML file that opens in any browser. The pencil beside a conversation in the list renames it, and deleting one asks first and cannot be undone. A conversation is worked on in one tab at a time. Opened in a second tab it is read-only there, and Continue here moves it.
The assistant picker names who runs each assistant. An assistant run by a maintainer of the project or by an administrator answers every member, and one run by anyone else answers only that person. Where the assistant’s model runs, and where what you send it goes, depends on your Plaid site. When the site has said, the panel shows it under About this assistant.
9.1. Other Projects
Where the assistant allows it, Add project, beside the paperclip, lets a question read other UD projects you can open. Each project added is shown above the message box. The assistant only reads other projects, and plans changes in the conversation’s own project only. A citation into another project carries that project’s name.
9.2. Examples in a Reply
When a reply rests on a particular sentence, that sentence is shown, drawn one of two ways. Tree gives the dependency arcs over the words, and CoNLL-U the sentence’s rows as a table. The assistant picks whichever fits, and you can switch. The heading of each example opens that sentence in the annotation editor.
9.3. Files and PDFs
The paperclip in the message box attaches up to five files to a question, and dropping or pasting a file does the same.
The assistant reads tables (.csv, .tsv), text (.txt, .md), JSON, XML, .conllu files, and PDFs, up to 4 MB of text each.
A .conllu file is only read. To bring it into the project, import it.
A reply can also give you a file the assistant made, such as a cleaned table, to download.
A PDF must have text in it, so a scan is refused. Ask about "§9" or "page 41" and the assistant reads that part. Text taken from a PDF with old fonts can come out wrong, so check a form against the PDF before relying on it. Where the assistant can look things up on the web, it also reads a PDF from a link or a DOI pasted into a question.
9.4. The Panel
The Assistant button in the header, or the handle halfway down the right edge of the window, opens the assistant in a panel beside whatever you are looking at. The panel stays open as you move around, and can be dragged wider or narrower. A narrow window has no room for it.
The panel picks up your last conversation in the project you are in. Its history lists your other conversations, and ⤢ moves the open one to the Assistant tab. On a screen outside any project, it asks which project to use. An answer that ends while you are in another tab or window is announced with a short sound, a ● at the start of the tab’s title, and a notification if you allowed them when you first sent a message.
A question asked on a document is about that document, unless you name another.
Ask, under a sentence, puts that sentence into the question.
Typing @ in the message box names a sentence of the open document, or a document of the project, by what it says.
Sentences are numbered as in the Annotate tab, and you can also name one by its sent_id.
9.5. Conversation Size
The bar in the panel’s header shows how full the conversation is. It fills as the conversation grows. Past 85 percent it turns amber, and a full conversation sends nothing more. New conversation then starts a new one with what you had typed.
9.6. Assistant Availability
The assistant is a service that your server’s operator runs and picks the model for. In a project with none connected, the tab, the button, the handle and Ask are not shown. On the project list the button is shown whether or not an assistant is connected, and the panel says so when the project you pick has none. Saved conversations still open from their own links.
10. Additional Features
10.1. Search
Each project has a Search screen that finds structures across all its documents using Grew-style patterns. For example, this finds verbs with a nominal subject:
pattern { V [upos=VERB]; S [upos=NOUN|PROPN]; V -[nsubj]-> S }
Type a pattern and run it with the Search button or Ctrl/Cmd+Enter.
Results are grouped by document. Click a match to jump straight to that sentence in the Annotate view.
A built-in syntax reference with more examples is available on the search screen.
is_tree and is_not_tree are not supported.
is_forest, is_not_forest, is_cyclic and is_not_cyclic take every word to have one head, as import and the editor keep it.
10.1.1. Looking for a Word
The box above the pattern finds a word without writing Grew. It looks in one field (form, lemma, UPOS, XPOS, dependency relation or features) for text that contains, is, or matches (as a regular expression) what you type. It writes the Grew pattern for you and runs it, so you can then edit the pattern.
In features, each Key=Value pair is matched on its own: contains Number finds every word with a Number feature.
An accent matches whether typed as one character or two.
In a regular expression, \w, \d and \s cover every script, and \p{IsArabic} matches any character of the Arabic script, and likewise for other scripts.
10.1.2. Searching the Enhanced Graph
A relation that the enhanced graph adds to the tree carries Grew’s E: prefix: E:nsubj is an enhanced nsubj.
-
V -[nsubj]→ Sreads the tree alone. -
V -[E:nsubj]→ Sreads only the relations added in the enhanced graph. -
V → S,V -[1=nsubj]→ SandV -[^det]→ Sread both. Addenhanced=yesor!enhancedto a feature label to read one. -
V -[nsubj|E:nsubj]→ Sreads each label where it belongs.
is_forest, is_projective and an unlabelled X →> Y are about the tree and ignore the added relations.
10.1.3. Counting Matches
After a search, Count by turns the results into a tally.
Pick a pattern node and a field (V.upos, S.lemma, V.FEATS, or a named edge’s e.label) to get a sortable table of values with how many matches each accounts for.
Clicking a node’s value adds it to the pattern and runs the search again.
10.1.4. Changing Many Sentences at Once
A pattern followed by a commands block is a rule, and the Search screen previews what it would change instead of listing matches.
This one retags that wherever it was tagged as a subordinator:
pattern { X [lemma="that", upos=SCONJ] }
commands { X.upos = PRON }
With a rule in the box the button reads Preview changes. A preview writes nothing.
It lists every sentence the rule would change, grouped by document, with one line per change such as that: upos SCONJ → PRON.
Every sentence starts ticked. Untick the sentences or documents you do not want, or use Select none and tick the ones you do, then press Apply and confirm. Anyone who can search can preview, and applying is for maintainers.
As in Grew, a rule is applied to a sentence again and again until it no longer matches.
So a rule has to stop matching what it has just changed. One that adds something usually needs a without block naming what it adds:
pattern { V1 -[conj]-> V2; V1 -[nsubj]-> S }
without { V2 -[E:nsubj]-> S }
commands { add_edge V2 -[E:nsubj]-> S }
A rule that changes nothing, or is still matching after 1000 applications in one sentence, stops with an error on that sentence.
The commands cover features (X.upos = VERB, X.Number = D.Number, del_feat X.Number), relations (add_edge, del_edge, e.label = "obj", shift), and deleting a word (del_node).
Words cannot be added or reordered, since they come from the text.
Several rules can be written as rule name { … } blocks. The syntax reference on the Search screen lists every command with examples.
Some things to know before applying:
-
The changes count as yours. A rule that changes a value a parser wrote confirms it, the same as editing it by hand.
-
A sentence the rule would leave with a word of two heads, or with a cycle, is an error in the preview and is not changed.
-
Where a vocabulary is closed (see Suggestions or Rules), a sentence in which the rule writes a value off the list is refused.
-
Each document’s changes are one entry in its history. To undo a rewrite, restore the document to the entry before it (see Restoring an Earlier State).
-
A document someone else changes between your preview and your apply is left untouched, and the preview runs again so you can apply the rest.
10.2. History
Every change is recorded. History, at the end of a document’s tab row or on the rail at the left edge of the editor, lists the actions taken on it, such as "Update UPOS", "Create relation", or a parser run with the name of the person who ran it. Selecting an entry shows the document as it was right after that action, and an entry made of several changes can be expanded to select each one. Past states are read-only, and Return to current goes back to the live document.
Entries of these kinds are marked with it, here and in the Activity tab:
-
Import: a file brought in.
-
Automatic: a parser’s run, or another service’s.
-
Assistant: an approved assistant plan.
-
Bulk edit: a Grew rewrite (see Changing Many Sentences at Once).
-
Review: a prediction accepted as it stands.
-
Repair: a correction made when the document was opened.
10.2.1. Restoring an Earlier State
Maintainers can put a document back the way it was. While viewing a past state, choose Restore, and the dialog lists what would be removed, brought back, and changed back. Anything that comes back keeps its comments. A few things cannot come back, such as links to vocabulary entries that no longer exist, and the dialog names them.
A restore is one entry in the history, so it can itself be undone the same way. The toast that confirms it also offers Undo for fifteen seconds.
10.3. Sharing and Permissions
Projects are shared by adding members with a role:
-
Reader: can view everything but make no edits.
-
Writer: can edit and delete documents and annotations.
-
Maintainer: can also change project settings, manage members, and delete the project.
Add members from the project’s Access settings by searching for a name or an email address. When you only have Reader access, the editor shows everything but its controls are inert. If a maintainer makes you a Reader or removes you while you have a document open, the page turns read-only within a minute.
Each member also has a Review work checkbox, independent of their role. Tick it and everything that person enters is marked as needing review until someone else confirms it (see Machine vs. Human Annotations).
10.4. Settings
A project’s maintainers tailor the annotation scheme from the project’s Settings tab.
UD settings holds the controlled vocabularies (UPOS, XPOS and dependency relations), the colors of UPOS tags and relations, the feature inventory offered in the Features column, and the document and sentence fields for Metadata. XPOS starts empty and the others from the universal sets. Each can be changed. Save writes only what you changed. If another maintainer changed the same setting since you opened the tab, nothing is saved and the message names the setting.
General holds the project’s language, the tokenizer locale used by Tokenize, and project deletion.
The language is a BCP-47 tag such as en, de or zh-Hans. A connected parser starts on it, and Tokenize uses it unless a more specific tokenizer locale is set.
10.4.1. Suggestions or Rules
A vocabulary suggests its values and accepts anything else. To make a list binding, turn on the switch beneath it: Refuse tags outside this list for UPOS and XPOS, Refuse relations outside this list for dependency relations, or Refuse features outside this list for features.
-
UPOS and XPOS: the tag must be on the list.
-
Dependency relations: the base relation must be on the list, so
nsubj:passis allowed wherevernsubjis. -
Features: the key must be in the inventory and the value in that key’s list. A key listed with no values accepts any value.
A value off a closed list is refused where you type it, and from an assistant plan or a script. Values from an import, a copy or a restore, and a parser’s unconfirmed predictions, are kept anyway, and the Validation tab lists them. Accept predictions skips a prediction that is off the list and names it.
A list cannot be closed, or a value removed from a closed one, while the project’s annotations hold a value off it that a person entered or confirmed. The message names one such value. Change those values first, or add them to the list.
10.4.2. Value Definitions
Each value can carry a one-line definition, shown beside it while you pick one.
The 17 UPOS tags and the 37 universal relations come with theirs, and you can edit them or describe your own values under the list.
Features are described as whole pairs, such as Number=Plur.
10.5. Comments
A document’s Comments tab shows every thread on it, the document’s own first, then one per sentence. Comments are for discussing the work. They are not part of the data, so they do not appear in an export or in the history. A comment can be attached to a sentence or to the document, and a note about one word belongs in its sentence’s thread.
In the Annotate view, a badge under each sentence’s grid shows how many comments it has and opens the thread in place. Bodies are Markdown. You may edit and delete your own comments, a maintainer may delete anyone’s, and Readers can read but not write. Comments from other people appear as they are written.
When the sentence a comment is on is merged away or re-segmented, the comment is kept and marked Outdated. A project shared with another app can hold threads on things this editor does not show. They are listed under On other layers.
10.6. Validation
Maintainers get a Validation tab listing every value the project uses that its vocabularies do not list, with how many times each occurs. Each value opens to the sentences it occurs in, each a link to that sentence in the editor. Where the list is incomplete rather than the values wrong, Add N values puts them on it.
A relation subtype whose base is listed is not reported, nor is a feature value under a key listed with no values.
10.7. Activity
A project’s maintainers get an Activity tab showing who has been working on it, and on what. It tallies each person’s changes and documents over the last 7, 30 or 90 days, or all time, with when they were last seen. Below that is the project’s history, newest first, each entry linking to the document it changed.
10.8. Administration
Server administration lives in Plaid IGT, and administrators reach it from Admin in the header. It covers accounts, invitations, every project on the server and its members, the connected services, and server-wide activity. A project set up for Universal Dependencies is labelled UD there, and its name opens it in Plaid UD.
10.9. Services
Plaid UD can call external services for two tasks, tokenizing and parsing. A service is a small program, often a Python script wrapping a model, that connects to your project and appears wherever its task is offered. A maintainer can make one the default for its task under Settings → Services.
-
Tokenization splits the saved text into sentences, tokens and words. The built-in method uses the browser’s own segmentation and is always available.
-
Parsing fills in lemmas, parts of speech, features and dependency relations. There is no built-in method, so it needs a connected service, such as the bundled Stanza parser. Sentences a person has annotated are left alone unless you say otherwise.
Tokenize and Parse each open a dialog with the Method to use and its options, which closes when the run starts. While a run writes to a document, the document is read-only, and a banner at the top shows the run’s progress and how long it has been going. A run keeps going if you leave the document or close the tab, and reopening the document picks it up. Closing or reloading the tab asks first. Stop on the banner ends a run, and whatever it wrote stays. A parse that stops partway can be run again to finish.
Everything a service produces is marked in violet until a person confirms it (see Machine vs. Human Annotations). A run that changed nothing says so. The Services chapter of the Plaid manual explains how to write a service of your own.
10.10. Scripting
On your user profile you can create named API tokens for scripts and services. The project’s Access settings link there with Manage API tokens. A token carries your permissions and does not expire, and actions taken with one are attributed to it in the history. A token cannot change a password, create an account, create or revoke tokens, or make an invitation or password-reset link. Do these signed in.
For writing your own scripts against a project, see the Python client and JavaScript client references.
11. Keyboard Shortcuts
Every shortcut that takes Ctrl also takes Cmd on a Mac. The shortcuts below are fixed and cannot be rebound. The editor carries a short version of this under Gestures and marks, above the first sentence.
11.1. Annotation Grid
| Shortcut | Action |
|---|---|
Tab / Shift+Tab |
Move to the next or previous cell in the row. |
↑ / ↓ |
Move between rows. ← / → move along the row from the ends of a value, following the words: in a right-to-left sentence ← moves forward. See Text Direction. |
Enter |
Commit the cell, taking the highlighted suggestion if the list is open. |
Esc |
Cancel the edit, or close an open list or panel. |
Backspace (Features) |
On the empty input, select the last feature, so a second press removes it. |
Alt+↓ |
In a lemma, XPOS or Features cell, list what the project has given words like this one, with counts, and pick one. See Earlier Values in This Project. |
Ctrl/Cmd+Enter |
Accept everything proposed on the current word, then move to the next word. |
Ctrl/Cmd+Backspace |
Discard the machine’s unconfirmed work on the current word, then move to the next word. Anything a person entered or confirmed stays. |
Ctrl/Cmd+Shift+↑ / ↓ |
Jump to the previous or next word that still needs a look. |
Alt+click a word |
Open that sentence in the Text Editor. |
↑ out of the top row |
Jump to this word’s relation label in the tree. |
11.2. Dependency Tree
| Shortcut | Action |
|---|---|
Drag from one word to another |
Draw a relation, or re-point an existing one. |
Ctrl/Cmd+drag from one word to another |
Add a relation to the enhanced graph. Over a relation of the tree, give it a different label there. |
Click a label |
Edit it. |
Ctrl/Cmd+click a relation of the tree, or Ctrl/Cmd+E on its label |
Leave it out of the enhanced graph, or put it back. |
Ctrl/Cmd+D |
Jump from the grid into the relation labels. From the tree’s labels, move to the enhanced relations under the words, and from there back. |
← / →, or Tab / Shift+Tab |
Move between labels. Enter edits the one you are on, ↓ drops back into the grid. |
Tab / Shift+Tab in an open label |
Commit it and open the next or previous label. |
Esc |
Leave the labels, or cancel an edit. |
Shift+Backspace in a label you have just opened |
Delete the relation. Once you have typed, it is an ordinary Backspace. |
11.3. Text Editor
| Gesture | Action |
|---|---|
Click a token |
Start or end a sentence there. |
Hover a token |
Open its panel, to edit its words or delete it. |
Alt+click a token |
Open that sentence in the Annotate tab. |
Select a stretch of text |
Make a token of it. |
11.4. Search
| Shortcut | Action |
|---|---|
Enter (word box) |
Write the pattern for what you typed and run it. |
Ctrl/Cmd+Enter (pattern box) |
Run the pattern. |
Tab (pattern box) |
Indent. Shift+Tab, or Escape then Tab, leaves the box. |