Named Entity Recognition

Tagging names, concepts or key phrases is a crucial task for natural language understanding pipelines. Prodigy lets you label NER training data or improve an existing model’s accuracy.


If you’re starting from scratch, you can use the ner.manual recipe with raw text and one or more labels and start highlighting entity spans. See the docs on fully manual annotation for an example. To make the process faster and more efficient, you can also use patterns to pre-highlight entities, so you only need to correct them. If you know that “Barack Obama” is pretty much always a PERSON, you can write a pattern for that. If you have a spaCy pipeline that’s already doing an okay job at predicting some of your entities, you can use ner.correct to pre-highlight them for you, so you only have to correct the suggestions and add new categories.

Once you’ve collected a dataset of maybe a few hundred annotations, you can run training experiments to see if you’re on the right track. The train recipe takes one or more Prodigy datasets, trains a model and outputs statistics and results. You can also use data-to-spacy to export data in spaCy’s format to use with spacy train, or db-out to export your annotations to use in any other process or application.

If you want to improve and correct an existing model on your data, you can use the ner.correct recipe to pre-highlight the model’s predictions, correct them manually and then update the model with the new data. See here for an example of the annotation workflow.

Instead of annotating every example, you can also use the model to suggest the analyses it’s most uncertain about. This workflow is implemented in the ner.teach recipe and you can see an example project in this section. Annotation can be very efficient, because you only have to press accept or reject. Once you’re done annotating, you can use train to update your model from the binary annotations so that both accepted and rejected examples have an impact.

If your goal is to extract longer phrases or nested and overlapping spans, a named entity recognizer won’t be a good fit for your problem: it will predict token-based tags and expects that a token can only be part of one entity. It also takes advantage of exact span boundaries, which is very effective for named entities like proper nouns, but less helpful for longer phrases.

Instead of using the named entity recognition workflows, check out the documentaton on span categorization and the spans.manual recipe, which was introduced in Prodigy v1.11. It lets you label any number of potentially overlapping or nested spans and you can then use the data to train spaCy’s SpanCategorizer component or a similar model.

Prodigy represents entity annotations in a simple JSON format with a "text", a "spans" property describing the start and end offsets and entity label of each entity in the text, and a list of "tokens". So you could extract the suggestions from your model in this format, and then use the mark recipe with --view-id ner_manual to label the data exactly as it comes in.

You can also write a custom recipe with a custom stream (a regular Python generator!) to plug in your model. If you can load and run your model in Python, you can use it with Prodigy. See the section on custom models for an example. If you want to use active learning with a custom model, you can make your recipe return an update callback that’s called whenever a new batch of answers is sent back from the web app. However, there are a few more considerations here, like how sensitive your model is to small updates. See this section for details and an example.

If your goal is to fine-tune a transformer model, you can use Prodigy to make annotation more efficient by streaming in data tokenized with the algorithm you need for your model. This also ensures that the data you create always maps to valid tokens produced by your model – no matter how intuitive or unintuitive the word piece tokenization is. You can read more about this in the docs on efficient annotation for transformers.

If you have existing annotations, you can convert them to Prodigy’s format and use the db-in command to import them to a new dataset. Each record should have a "text" and a list of "spans". You can then run train to train your model, use ner.manual to add more annotations, or run the review recipe to correct mistakes and resolve conflicts.

If all you want to do is train and you don’t need to collect or correct any annotations, you might find it more efficient to just train with spaCy (or any other library) directly.

Named entity recognition models work best at detecting relatively short phrases that have fairly distinct start and end points. A good way to think about how easy the model will find the task is to imagine you had to look at only the first word of the entity, with no context. How accurately would you be able to tell how that word should be labelled? Now imagine you had one word of context on either side, and ask yourself the same question.

With spaCy’s current defaults (as of v2.2), the model gets to see four words on either side of each token (it uses a convolutional neural network with four layers). You don’t have to use spaCy, and even if you do, you can reconfigure the model so that it has a wider contextual window. However, if you find you’re working on a task that requires distant information to make the decisions, you should consider learning at least part of the information you need with a text classification approach.

Entity recognition models will especially struggle on problems where the annotators disagree about the exact end points of the phrases. This is a common problem if your category is somewhat vague, like “cause of problem” or “complaint”. These meanings can be expressed in a variety of ways, and you can’t always pin down the part of a sentence that expresses a particular opinion. If you find that annotators can’t agree on exactly which words should be tagged as an entity, that’s probably a sign that you’re trying to mark something that’s a bit too semantic, in which case text classification would be a better approach.

Choosing the right recipe and workflow

So you have an NER problem you want to solve, and data to annotate. And you want to get it done as efficiently as possible. But how do you pick the right workflow for your use case?

  1. Fully manual: This is the most classic way of annotating data. You’re shown a raw text and highlight all entity spans in the text. At the end of the process, you export “gold-standard” data that you can train your model with. Fully manual annotation is often the best choice if you want to create a new dataset, but are starting completely from scratch without any existing resources you can leverage. In Prodigy, you can use this workflow with the ner.manual recipe.

  2. Manual with suggestions from patterns: Fully manual annotation can easily get tedious, which leads to mistakes and inconsistent data. You’re often doing the same thing over and over – for instance, in your data, mentions of “New York” may pretty much always refer to the location. Instead of labelling it by hand every single time, you can use keyword lists and patterns describing the tokens you’re looking for and make Prodigy pre-highlight the candidates for you. Even if your patterns only help 50% of the time, that’s still 50% less work for you. In Prodigy, you can use this workflow with ner.manual and the --patterns option. Also see the docs on patterns.

  3. Manual with suggestions from model: Instead of patterns, you can also use an existing model to highlight suggestions and save you time. This workflow is also useful if you want to train a model with a mix of new and existing categories. You can use the existing model to help you label the entity types you’re interested in, correct any mistakes it makes and add a new category on top. In Prodigy, you can use this workflow with the ner.correct recipe for spaCy pipeline or a custom recipe for any other model that predicts named entities.

  4. Binary with active learning and a model in the loop: This workflow is useful if you already have a model and want to fine-tune it on more data. Instead of annotating every example, you can use the model to suggest you the most relevant examples to annotate and give it feedback on its predictions. There are many different ways you can select the “best” examples, and a whole line of research dedicated to exploring active learning techniques. Prodigy’s ner.teach recipe implements simple uncertainty sampling with beam search: for each example, the annotation model gets a number of analyses and asks you to accept or reject the entity analyses it’s most uncertain about. Based on your decisions, the model is updated in the loop and guided towards better predictions. Prodigy also includes utilities that let you implement custom workflows with a model in the loop.

Video tutorial: training a model to predict ingredients

The following video shows an end-to-end workflow for training a named entity recognition model to recognize food ingredients from scratch, taking advantage of semi-automatic annotation with ner.manual and ner.correct, as well as modern transfer learning techniques. It took around 2.5 hours to create 949 annotations, including 20% evaluation examples, which was enough to achieve 85% accuracy. You can download the raw and annotated datasets from GitHub.

Fully manual annotation

To get started with manual NER annotation, all you need is a file with raw input text you want to annotate and a spaCy pipeline for tokenization (so the web app knows what a word is and can allow more efficient highlighting). The following command will start the web server with the ner.manual recipe, stream in headlines from news_headlines.jsonl and provides the label options PERSON, ORG and PRODUCT.

Download news_headlines.jsonl

Recipe command

prodigy ner.manual ner_news_headlines blank:en ./news_headlines.jsonl --label PERSON,ORG,PRODUCT,LOCATION

Try it live and highlight entities!

This live demo requires JavaScript to be enabled.

When you hit the accept, reject or ignore buttons, your answer will be submitted and Prodigy will add an "answer" key to the annotation task dict – for example, "answer": "accept". When you’re annotating manually, you typically only want to use accepted answers. This should also include examples with no entities! Those are just as important for the model to learn from than examples with entities. Ignoring an answer typically means that you want to skip it completely and exclude it from everything – for example, because you don’t know the answer or because the question is confusing or not representative. The reject button is less relevant in manual annotation mode, because there’s nothing to say no to – however, you can use it to reject examples that have actual problems that need fixing, like messy tokenization (that won’t let you highlight the entities you need), unicode issues and so on. When you view or export your data later, e.g. with db-out, you can then explicitly filter out those examples and deal with them.

Pre-tokenizing the text for the manual interfaces allows more efficient annotation, because the selection can “snap” to the token boundaries and doesn’t require pixel-perfect highlighting. Even if you only select parts of a word, the word is still locked in as an entity. (Pro tip: For single-token entities, you can even double-click on the word!)

Surfacing the tokenization like this also lets you spot potential problems early: if your text isn’t tokenized correctly and you’re updating your model with token-based annotations, it may never actually learn anything meaningful because it’ll never actually produce tokens consistent with the annotations. If you’re using your own model and tokenization, you can pass in data with a "tokens" property in Prodigy’s format instead of using spaCy to tokenize.

As of v1.10, you can also set the --highlight-chars flag on ner.manual to highlight individual characters instead of full tokens.

The manual interface for labelling spans was primarily designed for sequence tagging tasks where the spans are represented as a sequence of token-based tags with one tag per token. This is also how most NER model implementations are designed and what makes them work so well: they expect each token to only have one label that the model can predict. Under the hood, those tags often look like B-PERSON (beginning of a person entity) or I-PRODUCT (token inside a product entity).

Allowing overlapping spans in the annotation interface would easily be confusing and misleading, because you wouldn’t be able to use the data collected this way for the most common use cases. Aside from this, it’d also make the UI much more complex and there’s not really a satisfying answer for visualizing multiple nested overlapping spans while still making the interface efficient and intuitive to use.

For most use cases, there’s usually a better way to structure the task to avoid this problem. You could make multiple passes over the data if you have hierarchical labels. Or you could write a stream that keeps sending out the same example so you can label it multiple times until you reject it or send it back empty (which means all spans are added).

Character/Token-based highlighting toggle New: 1.14.5

The ner.manual recipe also lets you set a --highlight-chars flag to allow highlighting individual characters instead of only tokens. As of v1.14.5 the flag adds a checkbox button to the UI for switching between character and character based highlighting while annotating.

Character-based highlighting will only store the character offsets of your annotation and won’t add a "tokens" property to the saved task.

This live demo requires JavaScript to be enabled.
This live demo requires JavaScript to be enabled.

Manual annotation with patterns

To annotate with patterns, you’ll need a patterns file, a spaCy pipeline for tokenization (or your own custom tokenizer) and a source of input text. In the following example, we used 100 match patterns for fashion brands. You can find the full code and data of this project in our projects repo.

fashion_brands_patterns.jsonl (excerpt){"label": "FASHION_BRAND", "pattern": [{"lower": "ann"}, {"lower": "taylor"}]}
{"label": "FASHION_BRAND", "pattern": [{"lower": "topshop"}]}
{"label": "FASHION_BRAND", "pattern": [{"lower": "naked"}, {"lower": "and"}, {"lower": "famous"}]}

Using the ner.manual, we can now stream in and annotate the data. All pattern matches will be pre-highlighted in the UI, so we only need to make corrections (unselect wrong or incomplete matches) and add the entities that the patterns missed.

Recipe command

prodigy ner.manual ner_fashion_brands en_core_web_sm ./reddit_fashion.jsonl --label FASHION_BRAND --patterns ./fashion_brand_patterns.jsonl

In about 2 hours, we collected 1735 annotations (1235 training examples, 500 evaluation examples). The trained spaCy pipeline with word vectors and pretraining achieved 82.1% accuracy on the task.

Datasets & full evaluation

Working with patterns

Match patterns are typically provided as a JSONL (newline-delimited JSON) file and can be used to pre-highlight spans for faster and more efficient annotation. If you have words and phrases that you know are pretty much always an entity, you can use patterns so you only have to annotate the exceptions. Prodigy supports two types of patterns:

patterns.jsonl{"pattern": [{"lower": "new"}, {"lower": "york"}], "label": "GPE"}
{"pattern": "Berlin", "label": "GPE"}
  1. Token patterns: These patterns are lists of dictionaries with one dictionary describing one token to match. The token attributes to match on can be the token’s "text" or lowercase form "lower", but also lexical attributes like "is_punct" or linguistic features like "lemma" or "pos". You can find more details in the spaCy’s documentation on rule-based matching.

  2. String matches: If the pattern value is a string, it will be used for exact string matching. While {"lower": "berlin"} matches “Berlin”, “berlin” and so on, "Berlin" will only match “Berlin”. The advantage of string patterns is that you don’t have to worry about the tokenization and whether the patterns describe the correct tokens. They also make it easy to re-use existing word lists and dictionaries.

More about Prodigy pattern files

Manual annotation with suggestions from a model

Using a model to suggest entities is a great way to bootstrap training data for named entity recognition. For example, most general-purpose models were trained on large corpora of news and web text, annotated with at least a few generic entity types. Even if what you need is slightly different, you can still take advantage of what the model already predicts and let it highlight suggestions for you.

Let’s say you want to train a model for financial news with labels for person names, organizations, monetary amounts and ticker symbols. This is a very achievable task for named entity recognition. spaCy’s English models already predict PERSON, ORG and MONEY, so you can correct its suggestions for these labels and add annotations for your new TICKER label.

Download raw_shares-newsapi.jsonl

Recipe command

prodigy ner.correct ner_finance_news en_core_web_sm ./raw_shares-newsapi.jsonl --label PERSON,ORG,MONEY,TICKER --unsegmented

Instead of labelling all texts and entity types from scratch, you only need to fill in the blanks and fix incorrect predictions. If the model is decent enough, this can save you a lot of time. The workflow is also very helpful if you’ve already trained a custom model and want to get a better feeling for its predictions and perform error analysis.

Binary annotation with active learning and a model in the loop

The binary ner.teach workflow implements uncertainty sampling with beam search: for each example, the annotation model gets a number of analyses and asks you to accept or reject the entity analyses it’s most uncertain about. Based on your decisions, the model is updated in the loop and guided towards better predictions. Even if you don’t update with the full gold-standard annotation, confirming correct analyses and eliminating “dead ends” can very quickly move the model towards the right answer, since all token-based tagging decisions influence each other. Collecting binary feedback is also very fast and allows the annotator to focus on one concept and decision at a time, reducing the potential for human error and inconsistent data.

The following command starts the server and will make suggestions for the entity type ORG. The progress indicator in the sidebar shows an estimate of how much you still need to annotate until there’s nothing left to learn – or, phrased differently, an estimate of when the loss is going to hit zero.

Recipe command

prodigy ner.teach ner_news_headlines_binary en_core_web_sm ./news_headlines.jsonl --label ORG

Try it live and accept or reject!

This live demo requires JavaScript to be enabled.

If the suggested entity is fully correct, you can hit accept. If it’s entirely or partially wrong, you can hit reject. It’s important to only accept suggestions that are fully correct. As of v1.11, the ner.teach recipe will also ask you about examples containing no entities at all, which can improve overall accuracy of your model. So if you see an example with no highlighted suggestions, you can accept it if the text contains no entities, or reject it if it does contain entities of the labels you’re annotating.

This live demo requires JavaScript to be enabled.

The ner.teach recipe also supports match patterns via the --patterns argument. Pattern matches will be mixed in with the suggestions from the model and can help the model get over the “cold start” and make sure it starts off with enough positive examples to make meaningful suggestions. This can be used for entity types with very poor performance or even to add new entity types to an existing model.

If you come across a partially correct suggestion – for instance the entity “Facebook Inc” with only “Facebook” highlighted as a suggested ORG – you should always reject them. The active learning-powered recipes will look at all possible analyses for the parse, so the correct boundaries are likely in there – it might just not be the suggestion you see first. By rejecting incorrect boundaries, you’re essentially telling the model to try again, moving it towards the correct boundaries. Each token can only be part of one entity, so if you accepted a partial match like “Facebook” in “Facebook Inc”, the feedback the model would get from this is “Yes, in contexts like this, ‘Facebook’ is a single-token ORG entity and wins over all other possible analyses containing this token.” That’s obviously not what you want.

The “score” field in the bottom right corner of the annotation card shows you the score of the current entity suggestion. Even though the recipe tries to present you with the most uncertain scores, it can sometimes happen that you see very different scores instead. So why does this happen?

Streams are generators and only operate on one batch at a time. They can also stream from huge files or potentially infinite sources of data, so Prodigy can’t just load it all into memory and keep sorting the whole stream. Instead, it uses an exponential moving average to decide whether to send out a score, based on the distribution of previous scores. This also prevents it from getting stuck if the model suddenly produces higher or lower scores. If the scores are confusing and the model isn’t producing meaningful suggestions, try collecting some gold-standard data first before switching to the binary workflow.

When you annotate with a model in the loop, the model is also updated in the background. So why do you still need to train your model on the annotations afterwards, and can’t just export the model that was updated in the loop? The main reason is that the model in the loop is only updated once each new annotation. This is never going to be as effective as batch training a model on the whole dataset, making multiple passes over the data, shuffling on each epoch and using other deep learning tricks like dropout rates, compounding batch sizes and so on. If you batch train your model with the collected annotations afterwards, you should receive the same model you had in the loop, just better.

When you stop the recipe, the model in the loop is discarded and you can use train to train a better version of it using your annotations. If you just restart the recipe with the base model, it’ll start again at the beginning – otherwise, Prodigy would have to first batch train it behind the scenes and you might have to wait for quite a while until you can get started annotating. If you want to start with the updated model, you can train it with your annotations, output it to a directory and then initialize ner.teach with the updated model:

prodigy train ./batch-trained-model --ner ner_dataset --base-model en_core_web_sm
prodigy ner.teach ner_dataset ./batch-trained-model ./data.jsonl --label PERSON

To prevent unintended side-effects, you typically want to train the base model from scratch using all annotations every time you train – for example, you want to update en_core_web_sm with all annotations from one or more datasets and not update batch-trained-model, save the result, update that again and so on.

Efficient annotation for transformers like BERT New: 1.10

Transformer models like BERT typically use subword tokenization algorithms like WordPiece or Byte Pair Encoding (BPE) that are optimized for efficient embedding of large vocabularies and not necessarily linguistic definitions of what’s considered a “word”. If you’re creating training data for fine-tuning a transformer, you can use its tokenizer to preprocess your texts to make sure that the data you annotate is compatible with the transformer tokenization. It also makes annotation faster, because your selection can snap to token boundaries. The following recipe implementation uses Hugging Face’s easy-to-use tokenizers library under the hood and you can adjust it for your custom tokenization needs.

View bert.ner.manual recipe

Recipe command

prodigy bert.ner.manual ner_reddit ./reddit_comments.jsonl --label PERSON,ORG --tokenizer-vocab ./bert-base-uncased-vocab.txt --lowercase --hide-wp-prefix -F
This live demo requires JavaScript to be enabled.

The recipe takes advantage of the "ws" property of each token to reflect whether tokens are followed by whitespace or not. This way, the text stays readable, even if words are split into several pieces. Setting --hide--wp--prefix hides the prefix added to word pieces by the tokenizer so the tokens are displayed as ["bi", "eber"] instead of ["bi", "##eber"]. You can also set --hide-special to hide tokens like [CLS] and [SEP] instead of just disabling them.

The data format produced by Prodigy retains the original texts and offsets produced by the tokenizer, as well as the encoded IDs as "tokenizer_id".

JSON data format{
  "text": "Justin Bieber - agree 100000%",
  "tokens": [
    {"text": "[CLS]", "id": 0, "start": 0, "end": 0, "tokenizer_id": 101, "disabled": true, "ws": true},
    {"text": "justin", "id": 1, "start": 0, "end": 6, "tokenizer_id": 6796, "disabled": false, "ws": true},
    {"text": "bi", "id": 2, "start": 7, "end": 9, "tokenizer_id": 12170, "disabled": false, "ws": false},
    {"text": "eber", "id": 3, "start": 9, "end": 13, "tokenizer_id": 22669, "disabled": false, "ws": true},
    {"text": "-", "id": 4, "start": 14, "end": 15, "tokenizer_id": 1011, "disabled": false, "ws": true},
    {"text": "agree", "id": 5, "start": 16, "end": 21, "tokenizer_id": 5993, "disabled": false, "ws": true},
    {"text": "1000", "id": 6, "start": 22, "end": 26, "tokenizer_id": 6694, "disabled": false, "ws": false},
    {"text": "00", "id": 7, "start": 26, "end": 28, "tokenizer_id": 8889, "disabled": false, "ws": false},
    {"text": "%", "id": 8, "start": 28, "end": 29, "tokenizer_id": 1003, "disabled": false, "ws": true},
    {"text": "[SEP]", "id": 9, "start": 0, "end": 0, "tokenizer_id": 102, "disabled": true, "ws": true}
  "spans": [
     {"start": 0, "end": 13, "token_start": 1, "token_end": 3, "label": "PERSON"}

Working with longer texts

You probably noticed that most of the examples on this page show short texts like sentences or paragraphs. For NER annotation, there’s often no benefit in annotating long documents at once, especially if you’re planning on training a model on the data. Annotating with a model in the loop is also much faster if the texts aren’t too long, which is why recipes like ner.teach and ner.correct split sentences by default. NER model implementations also typically use a narrow contextual window of a few tokens on either side. If a human annotator can’t make a decision based on the local context, the model will struggle to learn from the data.

That said, there are always exceptions, and if you’re using the ner.manual workflow with whole documents, you can customize the UI theme to fit more text on the screen. For example:

prodigy.json (excerpt){
    "custom_theme": {"cardMaxWidth": "95%", "smallText": 16}

Screenshot of manual NER UI with long text

Prodigy will show you a little symbol to mark newlines, so you know where they are in your raw text. Tokens that only contain newlines are also unselectable by default, so you can’t accidentally include them in your annotated spans and confuse your model this way. See the ner_manual docs for more details about newline tokens.

Using a custom model

You don’t need to use spaCy to let a model highlight suggestions for you. Under the hood, the concept is pretty straightforward: if you stream in examples with pre-defined "spans", Prodigy will accept and pre-highlight them. This means you can either stream in pre-labelled data, or write a custom recipe that uses your model to add "spans" to the examples.

Expected format{
  "text": "Apple updates its analytics service with new metrics",
  "spans": [{"start": 0, "end": 5, "label": "ORG"}]

For inspiration, see the example recipe script that shows how to add spaCy entities to the stream of incoming examples. For a custom model, the code could look like this:

Step 1: Write the stream generatorpseudocode 
def add_entities_to_stream(stream): custom_model = load_your_custom_model() for eg in stream: ents = custom_model(eg["text"]) eg["spans"] = [{"start": start, "end": end, "label"} for start, end, label in ents] yield eg

If you want to extract and add the entities at runtime, you can write a custom recipe that loads the raw data, uses your custom model to add "spans" to the stream, pre-tokenizes the text and then renders it all using the ner_manual interface.

Step 2: Putting it all together in a recipepseudocode 
import prodigy from import get_stream from prodigy.components.preprocess import add_tokens import spacy @prodigy.recipe("custom-ner") def custom_ner_recipe(dataset, source): stream = get_stream(source) # load the data stream = add_entities_to_stream(stream) # add custom entities stream = add_tokens(spacy.blank("en"), stream) # add "tokens" to stream return { "dataset": dataset, # dataset to save annotations to "stream": stream, # the incoming stream of examples "view_id": "ner_manual", # annotation interface to use "labels": ["PERSON", "ORG"] # labels to annotate }

If your model only outputs token-based BILUO tags and doesn’t give you the character offsets, you can use spaCy’s biluo_tags_to_offsets helper to quickly convert the tokens and tags to character offsets. The function returns a list of (start, end, label) tuples that you can use to pre-populate the "spans" in Prodigy.

from spacy.tokens import Doc
from spacy.vocab import Vocab
from import biluo_tags_to_offsets

doc = Doc(Vocab(), words=["I", "like", "New", "York"])
tags = ["O", "O", "B-LOC", "L-LOC"]
offsets = biluo_tags_to_offsets(doc, tags)  # [(7, 15, 'LOC')]

Pre-tokenizing the text for the manual interfaces allows more efficient annotation, because the selection can “snap” to the token boundaries and doesn’t require pixel-perfect highlighting. You can try it out in the live demo – even if you only select parts of a word, the word is still locked in as an entity. (Pro tip: For single-token entities, you can even double-click on the word!)

Surfacing the tokenization like this also lets you spot potential problems early: if your text isn’t tokenized correctly and you’re updating your model with token-based annotations, it may never actually learn anything meaningful because it’ll never actually produce tokens consistent with the annotations.

If you’re using your own model and tokenization, you can pass in data with a "tokens" property in Prodigy’s format instead of using spaCy to tokenize. Prodigy will respect those tokens and split up the text accordingly. If you do want to use spaCy to train your final model, you can modify the tokenization rules to match your annotations or set skip=True in the add_tokens preprocessor to just ignore the mismatches.

Active learning with a custom model

If you want to update your model in the loop as you annotate, make sure you pick an implementation that supports updates in small batches and that’s sensitive enough to small updates (since you want your annotations to have an effect).

Step 1: Use the model to predict and score entitiespseudocode 
def predict(stream): for eg in stream: predictions = your_model(eg["text"]) for score, start, end, label in predictions: example = copy.deepcopy(eg) example["spans"] = [{"start": start, "end": end, "label": label}] yield (score, example)

On their own, the scores and examples aren’t that interesting yet – you typically want to use the scores to only select the most relevant examples for annotation. In the built-in ner.teach recipe, we’re using beam search to get a number of different possible analyses and then use the prefer_uncertain sorter that takes the stream of (score, example) tuples and yields example dicts with scores closest to 0.5.

Step 2: Sort the stream by scorepseudocode 
from prodigy.components.sorters import prefer_uncertain stream = predict(stream) stream = prefer_uncertain(stream)
Step 3: Update the model with answerspseudocode 
def update(answers): accepted = [eg for eg in answers if eg["answer"] == "accept"] rejected = [eg for eg in answers if eg["answer"] == "reject"] update_your_model(accepted, rejected)

By default, Prodigy streams are generators and Prodigy will only ever ask for the next batch from the stream. So as you annotate and update the model, future batches will receive scores from your updated model in the loop. For a simplified example of that loop, check out the recipe script. It shows a text classification use case, but the principle is very similar and is illustrated using a dummy model that “predicts” random numbers.

Step 4: Putting it all together in a recipepseudocode 
import prodigy from import get_stream from prodigy.components.sorters import prefer_uncertain @prodigy.recipe("custom-ner") def custom_ner_recipe(dataset, source): stream = get_stream(source) # load the data stream = predict(stream) # call custom predict function stream = prefer_uncertain(stream) # sort to prefer uncertain scores return { "dataset": dataset, # dataset to save annotations to "stream": stream, # the incoming stream of examples "update": update, # the update callback "view_id": "ner" # annotation interface to use }

Command-line usage

prodigy custom-ner ner_dataset ./your_data.jsonl -F

Training NER models

Once you’ve labelled some data with Prodigy, you can start your training experiments. If you’ve collected annotations from different sources or multiple annotators, it’s often a good idea to use the review recipe to resolve any conflicts and double-check the data. It’s also recommended to create a separate, dedicated evaluation set that you can compare different approaches against.

  1. Train a spaCy pipeline using Prodigy’s CLI. The train recipe is a wrapper around spaCy’s training API and optimized for training straight from Prodigy datasets and quick experiments. It reads from a dataset, holds back data for evaluation and outputs nicely-formatted results. This workflow is the best choice if you just want to get going or quickly check if you’re “on the right track” and your model is learning things.

  2. Train a model with spaCy directly. Once you’re getting more serious, it often makes sense to train your model directly with the library you’re using – e.g. spaCy. This gives you more control over the training process and hyperparameters, and lets you train all model components at once. The data-to-spacy command lets you convert Prodigy datasets to spaCy’s format and auto-generates a config to use with the spacy train command. It’s recommended to use the review recipe on the different annotation types first to resolve conflicts properly. To check if your data is valid and contains no issues, you can run spaCy’s debug-data command.

  3. Train a model with any other implementation or framework. The db-out exports annotations in a straightforward JSONL format. Entity spans are provided as a list of "spans" with their respective start and end character offsets, and the entity label. This should make it easy to convert and use it to train any model. If you annotated using a manual interface, the data will also include a "tokens" property so you can restore the tokenization.

spaCy’s NER architecture was designed to support continuous updates with more examples and even adding new labels to existing trained models. Updating an existing model makes sense if you want to keep using the same label scheme and only want to fine-tune it on more data or more specific data. However, it can easily lead to inconsistent behavior if you’re adding new entity types and/or annotations that conflict with the data the model was trained on. For instance, if you suddenly want to predict all cities as CITY instead of GPE. Instead of trying to “fight” the existing weights trained on millions of words, it often makes more sense to train a new model from scratch.

Even if you’re training from scratch, you can still use a trained model to help you create training data more efficiently. Prodigy’s ner.correct will stream in the model’s predictions for the given labels and lets you manually correct the entity spans. This way, you can let a model label the entity types you want to keep, add your new types on top, and make corrections along the way. This is a very effective method for bootstrapping large gold-standard training corpora without having to do all the labelling from scratch.

If your training process needs token-based BILUO or IOB tags instead of character offsets, you can use spaCy’s biluo_tags_from_offsets helper or the Token.ent_iob_ and Token.ent_type attributes to to quickly convert the annotations.

import spacy
from import biluo_tags_from_offsets

text = "I like New York City"
spans = [{"start": 7, "end": 20, "label": "LOC"}]
nlp = spacy.blank("en")
doc = nlp(text)

offsets = [(span["start"], span["end"], span["label"]) for span in spans]
biluo_tags = biluo_tags_from_offsets(doc, offsets)
# ['O', 'O', 'B-LOC', 'I-LOC', 'L-LOC']
doc.ents = [doc.char_span(start, end, label) for start, end, label in offsets]
iob_tags = [f"{t.ent_iob_}-{t.ent_type_}" if t.ent_iob_ else "O" for t in doc]
# ['O', 'O', 'B-LOC', 'I-LOC', 'I-LOC']

Improving accuracy and using transfer learning

After you’ve done your initial annotations and run your first training process, you’ll usually be interested in improving the accuracy of your model. After all, you’re unlikely to get to 100% accuracy right away – and if you ever do, that’s probably a sign that something’s wrong! Here are the steps we generally recommend to improve your initial accuracies:

  1. Check if you just need more data. The train-curve recipe is a helpful utility for checking if your model would improve with more data. It takes the same arguments as the train and will train with different samples of the data – for instance, 25%, 50%, 75% and 100%. As a rule of thumb, if accuracy improves within the last segment, it could indicate that more annotations of the same type will improve the model.

  2. Initialize with word vectors. Using pretrained word embeddings to initialize your model is easy and can make a big difference. If you’re using spaCy, try using the en_core_web_lg model as the base model. If you’re working with domain-specific texts, you can train your own vectors and create a base model with them.

  3. Initialize with pretrained tok2vec weights or transformer embeddings. Transfer learning has proven to be a very effective way to “share” knowledge between tasks, by initializing models with pretrained representations. Transfer learning is quite easy to use, but there are a few steps in the workflow. For an end-to-end example, check out this project with code, data and downloadable weights. Compared to just training with word vectors, an NER model trained on the same data achieved +8.7% accuracy. You can use spaCy’s config generator or init config command to generate a config for a transformer-based pipeline, and data-to-spacy to export your annotations for training.