Compared to using regular expressions on raw text, spaCy’s rule-based matcher engines and components not only let you find the words and phrases you’re looking for – they also give you access to the tokens within the document and their relationships. This means you can easily access and analyze the surrounding tokens, merge spans into single tokens or add entries to the named entities in doc.ents.

Token-based matching

spaCy features a rule-matching engine, the Matcher, that operates over tokens, similar to regular expressions. The rules can refer to token annotations (e.g. the token text or tag_, and flags like IS_PUNCT). The rule matcher also lets you pass in a custom callback to act on matches – for example, to merge entities and apply custom labels. You can also associate patterns with entity IDs, to allow some basic entity linking or disambiguation. To match large terminology lists, you can use the PhraseMatcher, which accepts Doc objects as match patterns.

Adding patterns

Let’s say we want to enable spaCy to find a combination of three tokens:

1. A token whose lowercase form matches “hello”, e.g. “Hello” or “HELLO”.
2. A token whose is_punct flag is set to True, i.e. any punctuation.
3. A token whose lowercase form matches “world”, e.g. “World” or “WORLD”.

[{"LOWER": "hello"}, {"IS_PUNCT": True}, {"LOWER": "world"}]

First, we initialize the Matcher with a vocab. The matcher must always share the same vocab with the documents it will operate on. We can now call matcher.add() with an ID and a list of patterns.

import spacy
from spacy.matcher import Matcher

nlp = spacy.load("en_core_web_sm")
matcher = Matcher(nlp.vocab)
# Add match ID "HelloWorld" with no callback and one pattern
pattern = [{"LOWER": "hello"}, {"IS_PUNCT": True}, {"LOWER": "world"}]
matcher.add("HelloWorld", [pattern])

doc = nlp("Hello, world! Hello world!")
matches = matcher(doc)
for match_id, start, end in matches:
    string_id = nlp.vocab.strings[match_id]  # Get string representation
    span = doc[start:end]  # The matched span
    print(match_id, string_id, start, end, span.text)

The matcher returns a list of (match_id, start, end) tuples – in this case, [('15578876784678163569', 0, 3)], which maps to the span doc[0:3] of our original document. The match_id is the hash value of the string ID “HelloWorld”. To get the string value, you can look up the ID in the StringStore.

for match_id, start, end in matches:
    string_id = nlp.vocab.strings[match_id]  # 'HelloWorld'
    span = doc[start:end]                    # The matched span

Optionally, we could also choose to add more than one pattern, for example to also match sequences without punctuation between “hello” and “world”:

patterns = [
    [{"LOWER": "hello"}, {"IS_PUNCT": True}, {"LOWER": "world"}],
    [{"LOWER": "hello"}, {"LOWER": "world"}]
]
matcher.add("HelloWorld", patterns)

By default, the matcher will only return the matches and not do anything else, like merge entities or assign labels. This is all up to you and can be defined individually for each pattern, by passing in a callback function as the on_match argument on add(). This is useful, because it lets you write entirely custom and pattern-specific logic. For example, you might want to merge some patterns into one token, while adding entity labels for other pattern types. You shouldn’t have to create different matchers for each of those processes.

Available token attributes

The available token pattern keys correspond to a number of Token attributes. The supported attributes for rule-based matching are:

Extended pattern syntax and attributes

Instead of mapping to a single value, token patterns can also map to a dictionary of properties. For example, to specify that the value of a lemma should be part of a list of values, or to set a minimum character length. The following rich comparison attributes are available:

Regular expressions

In some cases, only matching tokens and token attributes isn’t enough – for example, you might want to match different spellings of a word, without having to add a new pattern for each spelling.

pattern = [{"TEXT": {"REGEX": "^[Uu](\.?|nited)$"}},
           {"TEXT": {"REGEX": "^[Ss](\.?|tates)$"}},
           {"LOWER": "president"}]

The REGEX operator allows defining rules for any attribute string value, including custom attributes. It always needs to be applied to an attribute like TEXT, LOWER or TAG:

# Match different spellings of token texts
pattern = [{"TEXT": {"REGEX": "deff?in[ia]tely"}}]

# Match tokens with fine-grained POS tags starting with 'V'
pattern = [{"TAG": {"REGEX": "^V"}}]

# Match custom attribute values with regular expressions
pattern = [{"_": {"country": {"REGEX": "^[Uu](nited|\.?) ?[Ss](tates|\.?)$"}}}]

Matching regular expressions on the full text

If your expressions apply to multiple tokens, a simple solution is to match on the doc.text with re.finditer and use the Doc.char_span method to create a Span from the character indices of the match. If the matched characters don’t map to one or more valid tokens, Doc.char_span returns None.

import spacy
import re

nlp = spacy.load("en_core_web_sm")
doc = nlp("The United States of America (USA) are commonly known as the United States (U.S. or US) or America.")

expression = r"[Uu](nited|\.?) ?[Ss](tates|\.?)"
for match in re.finditer(expression, doc.text):
    start, end = match.span()
    span = doc.char_span(start, end)
    # This is a Span object or None if match doesn't map to valid token sequence
    if span is not None:
        print("Found match:", span.text)

chars_to_tokens = {}
for token in doc:
    for i in range(token.idx, token.idx + len(token.text)):
        chars_to_tokens[i] = token.i

span = doc.char_span(start, end)
if span is not None:
    print("Found match:", span.text)
else:
    start_token = chars_to_tokens.get(start)    end_token = chars_to_tokens.get(end)    if start_token is not None and end_token is not None:        span = doc[start_token:end_token + 1]        print("Found closest match:", span.text)

Operators and quantifiers

The matcher also lets you use quantifiers, specified as the 'OP' key. Quantifiers let you define sequences of tokens to be matched, e.g. one or more punctuation marks, or specify optional tokens. Note that there are no nested or scoped quantifiers – instead, you can build those behaviors with on_match callbacks.

Using wildcard token patterns

While the token attributes offer many options to write highly specific patterns, you can also use an empty dictionary, {} as a wildcard representing any token. This is useful if you know the context of what you’re trying to match, but very little about the specific token and its characters. For example, let’s say you’re trying to extract people’s user names from your data. All you know is that they are listed as “User name: {username}“. The name itself may contain any character, but no whitespace – so you’ll know it will be handled as one token.

[{"ORTH": "User"}, {"ORTH": "name"}, {"ORTH": ":"}, {}]

Validating and debugging patterns

The Matcher can validate patterns against a JSON schema with the option validate=True. This is useful for debugging patterns during development, in particular for catching unsupported attributes.

import spacy
from spacy.matcher import Matcher

nlp = spacy.load("en_core_web_sm")
matcher = Matcher(nlp.vocab, validate=True)
# Add match ID "HelloWorld" with unsupported attribute CASEINSENSITIVE
pattern = [{"LOWER": "hello"}, {"IS_PUNCT": True}, {"CASEINSENSITIVE": "world"}]
matcher.add("HelloWorld", [pattern])
# 🚨 Raises an error:
# MatchPatternError: Invalid token patterns for matcher rule 'HelloWorld'
# Pattern 0:
# - [pattern -> 2 -> CASEINSENSITIVE] extra fields not permitted

Adding on_match rules

To move on to a more realistic example, let’s say you’re working with a large corpus of blog articles, and you want to match all mentions of “Google I/O” (which spaCy tokenizes as ['Google', 'I', '/', 'O']). To be safe, you only match on the uppercase versions, avoiding matches with phrases such as “Google i/o”.

from spacy.lang.en import English
from spacy.matcher import Matcher
from spacy.tokens import Span

nlp = English()
matcher = Matcher(nlp.vocab)

def add_event_ent(matcher, doc, i, matches):
    # Get the current match and create tuple of entity label, start and end.
    # Append entity to the doc's entity. (Don't overwrite doc.ents!)
    match_id, start, end = matches[i]
    entity = Span(doc, start, end, label="EVENT")
    doc.ents += (entity,)
    print(entity.text)

pattern = [{"ORTH": "Google"}, {"ORTH": "I"}, {"ORTH": "/"}, {"ORTH": "O"}]
matcher.add("GoogleIO", [pattern], on_match=add_event_ent)
doc = nlp("This is a text about Google I/O")
matches = matcher(doc)

A very similar logic has been implemented in the built-in EntityRuler by the way. It also takes care of handling overlapping matches, which you would otherwise have to take care of yourself.

We can now call the matcher on our documents. The patterns will be matched in the order they occur in the text. The matcher will then iterate over the matches, look up the callback for the match ID that was matched, and invoke it.

doc = nlp(YOUR_TEXT_HERE)
matcher(doc)

When the callback is invoked, it is passed four arguments: the matcher itself, the document, the position of the current match, and the total list of matches. This allows you to write callbacks that consider the entire set of matched phrases, so that you can resolve overlaps and other conflicts in whatever way you prefer.

Creating spans from matches

Creating Span objects from the returned matches is a very common use case. spaCy makes this easy by giving you access to the start and end token of each match, which you can use to construct a new span with an optional label. As of spaCy v3.0, you can also set as_spans=True when calling the matcher on a Doc, which will return a list of Span objects using the match_id as the span label.

import spacy
from spacy.matcher import Matcher
from spacy.tokens import Span

nlp = spacy.blank("en")
matcher = Matcher(nlp.vocab)
matcher.add("PERSON", [[{"lower": "barack"}, {"lower": "obama"}]])
doc = nlp("Barack Obama was the 44th president of the United States")

# 1. Return (match_id, start, end) tuples
matches = matcher(doc)
for match_id, start, end in matches:
    # Create the matched span and assign the match_id as a label
    span = Span(doc, start, end, label=match_id)
    print(span.text, span.label_)

# 2. Return Span objects directly
matches = matcher(doc, as_spans=True)
for span in matches:
    print(span.text, span.label_)

Using custom pipeline components

Let’s say your data also contains some annoying pre-processing artifacts, like leftover HTML line breaks (e.g. <br> or <BR/>). To make your text easier to analyze, you want to merge those into one token and flag them, to make sure you can ignore them later. Ideally, this should all be done automatically as you process the text. You can achieve this by adding a custom pipeline component that’s called on each Doc object, merges the leftover HTML spans and sets an attribute bad_html on the token.

import spacy
from spacy.language import Language
from spacy.matcher import Matcher
from spacy.tokens import Token

# We're using a component factory because the component needs to be
# initialized with the shared vocab via the nlp object
@Language.factory("html_merger")
def create_bad_html_merger(nlp, name):
    return BadHTMLMerger(nlp.vocab)

class BadHTMLMerger:
    def __init__(self, vocab):
        patterns = [
            [{"ORTH": "<"}, {"LOWER": "br"}, {"ORTH": ">"}],
            [{"ORTH": "<"}, {"LOWER": "br/"}, {"ORTH": ">"}],
        ]
        # Register a new token extension to flag bad HTML
        Token.set_extension("bad_html", default=False)
        self.matcher = Matcher(vocab)
        self.matcher.add("BAD_HTML", patterns)

    def __call__(self, doc):
        # This method is invoked when the component is called on a Doc
        matches = self.matcher(doc)
        spans = []  # Collect the matched spans here
        for match_id, start, end in matches:
            spans.append(doc[start:end])
        with doc.retokenize() as retokenizer:
            for span in spans:
                retokenizer.merge(span)
                for token in span:
                    token._.bad_html = True  # Mark token as bad HTML
        return doc

nlp = spacy.load("en_core_web_sm")
nlp.add_pipe("html_merger", last=True)  # Add component to the pipeline
doc = nlp("Hello<br>world! <br/> This is a test.")
for token in doc:
    print(token.text, token._.bad_html)

Instead of hard-coding the patterns into the component, you could also make it take a path to a JSON file containing the patterns. This lets you reuse the component with different patterns, depending on your application. When adding the component to the pipeline with nlp.add_pipe, you can pass in the argument via the config:

@Language.factory("html_merger", default_config={"path": None})
def create_bad_html_merger(nlp, name, path):
    return BadHTMLMerger(nlp, path=path)

nlp.add_pipe("html_merger", config={"path": "/path/to/patterns.json"})

Example: Using linguistic annotations

Let’s say you’re analyzing user comments and you want to find out what people are saying about Facebook. You want to start off by finding adjectives following “Facebook is” or “Facebook was”. This is obviously a very rudimentary solution, but it’ll be fast, and a great way to get an idea for what’s in your data. Your pattern could look like this:

[{"LOWER": "facebook"}, {"LEMMA": "be"}, {"POS": "ADV", "OP": "*"}, {"POS": "ADJ"}]

This translates to a token whose lowercase form matches “facebook” (like Facebook, facebook or FACEBOOK), followed by a token with the lemma “be” (for example, is, was, or ‘s), followed by an optional adverb, followed by an adjective. Using the linguistic annotations here is especially useful, because you can tell spaCy to match “Facebook’s annoying”, but not “Facebook’s annoying ads”. The optional adverb makes sure you won’t miss adjectives with intensifiers, like “pretty awful” or “very nice”.

To get a quick overview of the results, you could collect all sentences containing a match and render them with the displaCy visualizer. In the callback function, you’ll have access to the start and end of each match, as well as the parent Doc. This lets you determine the sentence containing the match, doc[start:end].sent, and calculate the start and end of the matched span within the sentence. Using displaCy in “manual” mode lets you pass in a list of dictionaries containing the text and entities to render.

import spacy
from spacy import displacy
from spacy.matcher import Matcher

nlp = spacy.load("en_core_web_sm")
matcher = Matcher(nlp.vocab)
matched_sents = []  # Collect data of matched sentences to be visualized

def collect_sents(matcher, doc, i, matches):
    match_id, start, end = matches[i]
    span = doc[start:end]  # Matched span
    sent = span.sent  # Sentence containing matched span
    # Append mock entity for match in displaCy style to matched_sents
    # get the match span by ofsetting the start and end of the span with the
    # start and end of the sentence in the doc
    match_ents = [{
        "start": span.start_char - sent.start_char,
        "end": span.end_char - sent.start_char,
        "label": "MATCH",
    }]
    matched_sents.append({"text": sent.text, "ents": match_ents})

pattern = [{"LOWER": "facebook"}, {"LEMMA": "be"}, {"POS": "ADV", "OP": "*"},
           {"POS": "ADJ"}]
matcher.add("FacebookIs", [pattern], on_match=collect_sents)  # add pattern
doc = nlp("I'd say that Facebook is evil. – Facebook is pretty cool, right?")
matches = matcher(doc)

# Serve visualization of sentences containing match with displaCy
# set manual=True to make displaCy render straight from a dictionary
# (if you're not running the code within a Jupyer environment, you can
# use displacy.serve instead)
displacy.render(matched_sents, style="ent", manual=True)

Example: Phone numbers

Phone numbers can have many different formats and matching them is often tricky. During tokenization, spaCy will leave sequences of numbers intact and only split on whitespace and punctuation. This means that your match pattern will have to look out for number sequences of a certain length, surrounded by specific punctuation – depending on the national conventions.

The IS_DIGIT flag is not very helpful here, because it doesn’t tell us anything about the length. However, you can use the SHAPE flag, with each d representing a digit (up to 4 digits / characters):

[{"ORTH": "("}, {"SHAPE": "ddd"}, {"ORTH": ")"}, {"SHAPE": "dddd"},
 {"ORTH": "-", "OP": "?"}, {"SHAPE": "dddd"}]

This will match phone numbers of the format (123) 4567 8901 or (123) 4567-8901. To also match formats like (123) 456 789, you can add a second pattern using 'ddd' in place of 'dddd'. By hard-coding some values, you can match only certain, country-specific numbers. For example, here’s a pattern to match the most common formats of international German numbers:

[{"ORTH": "+"}, {"ORTH": "49"}, {"ORTH": "(", "OP": "?"}, {"SHAPE": "dddd"},
 {"ORTH": ")", "OP": "?"}, {"SHAPE": "dddd", "LENGTH": 6}]

Depending on the formats your application needs to match, creating an extensive set of rules like this is often better than training a model. It’ll produce more predictable results, is much easier to modify and extend, and doesn’t require any training data – only a set of test cases.

import spacy
from spacy.matcher import Matcher

nlp = spacy.load("en_core_web_sm")
matcher = Matcher(nlp.vocab)
pattern = [{"ORTH": "("}, {"SHAPE": "ddd"}, {"ORTH": ")"}, {"SHAPE": "ddd"},
           {"ORTH": "-", "OP": "?"}, {"SHAPE": "ddd"}]
matcher.add("PHONE_NUMBER", [pattern])

doc = nlp("Call me at (123) 456 789 or (123) 456 789!")
print([t.text for t in doc])
matches = matcher(doc)
for match_id, start, end in matches:
    span = doc[start:end]
    print(span.text)

Example: Hashtags and emoji on social media

Social media posts, especially tweets, can be difficult to work with. They’re very short and often contain various emoji and hashtags. By only looking at the plain text, you’ll lose a lot of valuable semantic information.

Let’s say you’ve extracted a large sample of social media posts on a specific topic, for example posts mentioning a brand name or product. As the first step of your data exploration, you want to filter out posts containing certain emoji and use them to assign a general sentiment score, based on whether the expressed emotion is positive or negative, e.g. 😀 or 😞. You also want to find, merge and label hashtags like #MondayMotivation, to be able to ignore or analyze them later.

By default, spaCy’s tokenizer will split emoji into separate tokens. This means that you can create a pattern for one or more emoji tokens. Valid hashtags usually consist of a #, plus a sequence of ASCII characters with no whitespace, making them easy to match as well.

from spacy.lang.en import English
from spacy.matcher import Matcher

nlp = English()  # We only want the tokenizer, so no need to load a pipeline
matcher = Matcher(nlp.vocab)

pos_emoji = ["😀", "😃", "😂", "🤣", "😊", "😍"]  # Positive emoji
neg_emoji = ["😞", "😠", "😩", "😢", "😭", "😒"]  # Negative emoji

# Add patterns to match one or more emoji tokens
pos_patterns = [[{"ORTH": emoji}] for emoji in pos_emoji]
neg_patterns = [[{"ORTH": emoji}] for emoji in neg_emoji]

# Function to label the sentiment
def label_sentiment(matcher, doc, i, matches):
    match_id, start, end = matches[i]
    if doc.vocab.strings[match_id] == "HAPPY":  # Don't forget to get string!
        doc.sentiment += 0.1  # Add 0.1 for positive sentiment
    elif doc.vocab.strings[match_id] == "SAD":
        doc.sentiment -= 0.1  # Subtract 0.1 for negative sentiment

matcher.add("HAPPY", pos_patterns, on_match=label_sentiment)  # Add positive pattern
matcher.add("SAD", neg_patterns, on_match=label_sentiment)  # Add negative pattern

# Add pattern for valid hashtag, i.e. '#' plus any ASCII token
matcher.add("HASHTAG", [[{"ORTH": "#"}, {"IS_ASCII": True}]])

doc = nlp("Hello world 😀 #MondayMotivation")
matches = matcher(doc)
for match_id, start, end in matches:
    string_id = doc.vocab.strings[match_id]  # Look up string ID
    span = doc[start:end]
    print(string_id, span.text)

Because the on_match callback receives the ID of each match, you can use the same function to handle the sentiment assignment for both the positive and negative pattern. To keep it simple, we’ll either add or subtract 0.1 points – this way, the score will also reflect combinations of emoji, even positive and negative ones.

With a library like Emojipedia, we can also retrieve a short description for each emoji – for example, 😍‘s official title is “Smiling Face With Heart-Eyes”. Assigning it to a custom attribute on the emoji span will make it available as span._.emoji_desc.

from emojipedia import Emojipedia  # Installation: pip install emojipedia
from spacy.tokens import Span  # Get the global Span object

Span.set_extension("emoji_desc", default=None)  # Register the custom attribute

def label_sentiment(matcher, doc, i, matches):
    match_id, start, end = matches[i]
    if doc.vocab.strings[match_id] == "HAPPY":  # Don't forget to get string!
        doc.sentiment += 0.1  # Add 0.1 for positive sentiment
    elif doc.vocab.strings[match_id] == "SAD":
        doc.sentiment -= 0.1  # Subtract 0.1 for negative sentiment
    span = doc[start:end]
    emoji = Emojipedia.search(span[0].text)  # Get data for emoji
    span._.emoji_desc = emoji.title  # Assign emoji description

To label the hashtags, we can use a custom attribute set on the respective token:

import spacy
from spacy.matcher import Matcher
from spacy.tokens import Token

nlp = spacy.load("en_core_web_sm")
matcher = Matcher(nlp.vocab)

# Add pattern for valid hashtag, i.e. '#' plus any ASCII token
matcher.add("HASHTAG", [[{"ORTH": "#"}, {"IS_ASCII": True}]])

# Register token extension
Token.set_extension("is_hashtag", default=False)

doc = nlp("Hello world 😀 #MondayMotivation")
matches = matcher(doc)
hashtags = []
for match_id, start, end in matches:
    if doc.vocab.strings[match_id] == "HASHTAG":
        hashtags.append(doc[start:end])
with doc.retokenize() as retokenizer:
    for span in hashtags:
        retokenizer.merge(span)
        for token in span:
            token._.is_hashtag = True

for token in doc:
    print(token.text, token._.is_hashtag)

Efficient phrase matching

If you need to match large terminology lists, you can also use the PhraseMatcher and create Doc objects instead of token patterns, which is much more efficient overall. The Doc patterns can contain single or multiple tokens.

Adding phrase patterns

import spacy
from spacy.matcher import PhraseMatcher

nlp = spacy.load("en_core_web_sm")
matcher = PhraseMatcher(nlp.vocab)
terms = ["Barack Obama", "Angela Merkel", "Washington, D.C."]
# Only run nlp.make_doc to speed things up
patterns = [nlp.make_doc(text) for text in terms]
matcher.add("TerminologyList", patterns)

doc = nlp("German Chancellor Angela Merkel and US President Barack Obama "
          "converse in the Oval Office inside the White House in Washington, D.C.")
matches = matcher(doc)
for match_id, start, end in matches:
    span = doc[start:end]
    print(span.text)

Since spaCy is used for processing both the patterns and the text to be matched, you won’t have to worry about specific tokenization – for example, you can simply pass in nlp("Washington, D.C.") and won’t have to write a complex token pattern covering the exact tokenization of the term.

Matching on other token attributes

By default, the PhraseMatcher will match on the verbatim token text, e.g. Token.text. By setting the attr argument on initialization, you can change which token attribute the matcher should use when comparing the phrase pattern to the matched Doc. For example, using the attribute LOWER lets you match on Token.lower and create case-insensitive match patterns:

from spacy.lang.en import English
from spacy.matcher import PhraseMatcher

nlp = English()
matcher = PhraseMatcher(nlp.vocab, attr="LOWER")
patterns = [nlp.make_doc(name) for name in ["Angela Merkel", "Barack Obama"]]
matcher.add("Names", patterns)

doc = nlp("angela merkel and us president barack Obama")
for match_id, start, end in matcher(doc):
    print("Matched based on lowercase token text:", doc[start:end])

Another possible use case is matching number tokens like IP addresses based on their shape. This means that you won’t have to worry about how those string will be tokenized and you’ll be able to find tokens and combinations of tokens based on a few examples. Here, we’re matching on the shapes ddd.d.d.d and ddd.ddd.d.d:

from spacy.lang.en import English
from spacy.matcher import PhraseMatcher

nlp = English()
matcher = PhraseMatcher(nlp.vocab, attr="SHAPE")
matcher.add("IP", [nlp("127.0.0.1"), nlp("127.127.0.0")])

doc = nlp("Often the router will have an IP address such as 192.168.1.1 or 192.168.2.1.")
for match_id, start, end in matcher(doc):
    print("Matched based on token shape:", doc[start:end])

In theory, the same also works for attributes like POS. For example, a pattern nlp("I like cats") matched based on its part-of-speech tag would return a match for “I love dogs”. You could also match on boolean flags like IS_PUNCT to match phrases with the same sequence of punctuation and non-punctuation tokens as the pattern. But this can easily get confusing and doesn’t have much of an advantage over writing one or two token patterns.

Dependency Matcher v3.0Needs model

The DependencyMatcher lets you match patterns within the dependency parse using Semgrex operators. It requires a model containing a parser such as the DependencyParser. Instead of defining a list of adjacent tokens as in Matcher patterns, the DependencyMatcher patterns match tokens in the dependency parse and specify the relations between them.

A pattern added to the dependency matcher consists of a list of dictionaries, with each dictionary describing a token to match and its relation to an existing token in the pattern. Except for the first dictionary, which defines an anchor token using only RIGHT_ID and RIGHT_ATTRS, each pattern should have the following keys:

Each additional token added to the pattern is linked to an existing token LEFT_ID by the relation REL_OP. The new token is given the name RIGHT_ID and described by the attributes RIGHT_ATTRS.

Dependency matcher operators

The following operators are supported by the DependencyMatcher, most of which come directly from Semgrex:

Designing dependency matcher patterns

Let’s say we want to find sentences describing who founded what kind of company:

• Smith founded a healthcare company in 2005.
• Williams initially founded an insurance company in 1987.
• Lee, an experienced CEO, has founded two AI startups.

The dependency parse for “Smith founded a healthcare company” shows types of relations and tokens we want to match:

The relations we’re interested in are:

• the founder is the subject (nsubj) of the token with the text founded
• the company is the object (dobj) of founded
• the kind of company may be an adjective (amod, not shown above) or a compound (compound)

The first step is to pick an anchor token for the pattern. Since it’s the root of the dependency parse, founded is a good choice here. It is often easier to construct patterns when all dependency relation operators point from the head to the children. In this example, we’ll only use >, which connects a head to an immediate dependent as head > child.

The simplest dependency matcher pattern will identify and name a single token in the tree:

import spacy
from spacy.matcher import DependencyMatcher

nlp = spacy.load("en_core_web_sm")
matcher = DependencyMatcher(nlp.vocab)
pattern = [
  {
    "RIGHT_ID": "anchor_founded",       # unique name
    "RIGHT_ATTRS": {"ORTH": "founded"}  # token pattern for "founded"
  }
]
matcher.add("FOUNDED", [pattern])
doc = nlp("Smith founded two companies.")
matches = matcher(doc)
print(matches) # [(4851363122962674176, [1])]

Now that we have a named anchor token (anchor_founded), we can add the founder as the immediate dependent (>) of founded with the dependency label nsubj:

Step 1
pattern = [
    {
        "RIGHT_ID": "anchor_founded",
        "RIGHT_ATTRS": {"ORTH": "founded"}
    },
    {
        "LEFT_ID": "anchor_founded",
        "REL_OP": ">",        "RIGHT_ID": "founded_subject",
        "RIGHT_ATTRS": {"DEP": "nsubj"},    }
    # ...
]

The direct object (dobj) is added in the same way:

Step 2
pattern = [
    #...
    {
        "LEFT_ID": "anchor_founded",
        "REL_OP": ">",
        "RIGHT_ID": "founded_object",
        "RIGHT_ATTRS": {"DEP": "dobj"},
    }
    # ...
]

When the subject and object tokens are added, they are required to have names under the key RIGHT_ID, which are allowed to be any unique string, e.g. founded_subject. These names can then be used as LEFT_ID to link new tokens into the pattern. For the final part of our pattern, we’ll specify that the token founded_object should have a modifier with the dependency relation amod or compound:

Step 3
pattern = [
    # ...
    {
        "LEFT_ID": "founded_object",
        "REL_OP": ">",
        "RIGHT_ID": "founded_object_modifier",
        "RIGHT_ATTRS": {"DEP": {"IN": ["amod", "compound"]}},    }
]

You can picture the process of creating a dependency matcher pattern as defining an anchor token on the left and building up the pattern by linking tokens one-by-one on the right using relation operators. To create a valid pattern, each new token needs to be linked to an existing token on its left. As for founded in this example, a token may be linked to more than one token on its right:

The full pattern comes together as shown in the example below:

import spacy
from spacy.matcher import DependencyMatcher

nlp = spacy.load("en_core_web_sm")
matcher = DependencyMatcher(nlp.vocab)

pattern = [
    {
        "RIGHT_ID": "anchor_founded",
        "RIGHT_ATTRS": {"ORTH": "founded"}
    },
    {
        "LEFT_ID": "anchor_founded",
        "REL_OP": ">",
        "RIGHT_ID": "founded_subject",
        "RIGHT_ATTRS": {"DEP": "nsubj"},
    },
    {
        "LEFT_ID": "anchor_founded",
        "REL_OP": ">",
        "RIGHT_ID": "founded_object",
        "RIGHT_ATTRS": {"DEP": "dobj"},
    },
    {
        "LEFT_ID": "founded_object",
        "REL_OP": ">",
        "RIGHT_ID": "founded_object_modifier",
        "RIGHT_ATTRS": {"DEP": {"IN": ["amod", "compound"]}},
    }
]

matcher.add("FOUNDED", [pattern])
doc = nlp("Lee, an experienced CEO, has founded two AI startups.")
matches = matcher(doc)

print(matches) # [(4851363122962674176, [6, 0, 10, 9])]
# Each token_id corresponds to one pattern dict
match_id, token_ids = matches[0]
for i in range(len(token_ids)):
    print(pattern[i]["RIGHT_ID"] + ":", doc[token_ids[i]].text)

Rule-based entity recognition

The EntityRuler is a component that lets you add named entities based on pattern dictionaries, which makes it easy to combine rule-based and statistical named entity recognition for even more powerful pipelines.

Entity Patterns

Entity patterns are dictionaries with two keys: "label", specifying the label to assign to the entity if the pattern is matched, and "pattern", the match pattern. The entity ruler accepts two types of patterns:

1. Phrase patterns for exact string matches (string).

   {"label": "ORG", "pattern": "Apple"}
   

2. Token patterns with one dictionary describing one token (list).

   {"label": "GPE", "pattern": [{"LOWER": "san"}, {"LOWER": "francisco"}]}
   

Using the entity ruler

The EntityRuler is a pipeline component that’s typically added via nlp.add_pipe. When the nlp object is called on a text, it will find matches in the doc and add them as entities to the doc.ents, using the specified pattern label as the entity label. If any matches were to overlap, the pattern matching most tokens takes priority. If they also happen to be equally long, then the match occurring first in the Doc is chosen.

from spacy.lang.en import English

nlp = English()
ruler = nlp.add_pipe("entity_ruler")
patterns = [{"label": "ORG", "pattern": "Apple"},
            {"label": "GPE", "pattern": [{"LOWER": "san"}, {"LOWER": "francisco"}]}]
ruler.add_patterns(patterns)

doc = nlp("Apple is opening its first big office in San Francisco.")
print([(ent.text, ent.label_) for ent in doc.ents])

The entity ruler is designed to integrate with spaCy’s existing pipeline components and enhance the named entity recognizer. If it’s added before the "ner" component, the entity recognizer will respect the existing entity spans and adjust its predictions around it. This can significantly improve accuracy in some cases. If it’s added after the "ner" component, the entity ruler will only add spans to the doc.ents if they don’t overlap with existing entities predicted by the model. To overwrite overlapping entities, you can set overwrite_ents=True on initialization.

import spacy

nlp = spacy.load("en_core_web_sm")
ruler = nlp.add_pipe("entity_ruler")
patterns = [{"label": "ORG", "pattern": "MyCorp Inc."}]
ruler.add_patterns(patterns)

doc = nlp("MyCorp Inc. is a company in the U.S.")
print([(ent.text, ent.label_) for ent in doc.ents])

Validating and debugging EntityRuler patterns

The entity ruler can validate patterns against a JSON schema with the config setting "validate". See details under Validating and debugging patterns.

ruler = nlp.add_pipe("entity_ruler", config={"validate": True})

Adding IDs to patterns

The EntityRuler can also accept an id attribute for each pattern. Using the id attribute allows multiple patterns to be associated with the same entity.

from spacy.lang.en import English

nlp = English()
ruler = nlp.add_pipe("entity_ruler")
patterns = [{"label": "ORG", "pattern": "Apple", "id": "apple"},
            {"label": "GPE", "pattern": [{"LOWER": "san"}, {"LOWER": "francisco"}], "id": "san-francisco"},
            {"label": "GPE", "pattern": [{"LOWER": "san"}, {"LOWER": "fran"}], "id": "san-francisco"}]
ruler.add_patterns(patterns)

doc1 = nlp("Apple is opening its first big office in San Francisco.")
print([(ent.text, ent.label_, ent.ent_id_) for ent in doc1.ents])

doc2 = nlp("Apple is opening its first big office in San Fran.")
print([(ent.text, ent.label_, ent.ent_id_) for ent in doc2.ents])

If the id attribute is included in the EntityRuler patterns, the ent_id_ property of the matched entity is set to the id given in the patterns. So in the example above it’s easy to identify that “San Francisco” and “San Fran” are both the same entity.

Using pattern files

The to_disk and from_disk let you save and load patterns to and from JSONL (newline-delimited JSON) files, containing one pattern object per line.

patterns.jsonl
{"label": "ORG", "pattern": "Apple"}
{"label": "GPE", "pattern": [{"LOWER": "san"}, {"LOWER": "francisco"}]}

ruler.to_disk("./patterns.jsonl")
new_ruler = nlp.add_pipe("entity_ruler").from_disk("./patterns.jsonl")

When you save out an nlp object that has an EntityRuler added to its pipeline, its patterns are automatically exported to the pipeline directory:

nlp = spacy.load("en_core_web_sm")
ruler = nlp.add_pipe("entity_ruler")
ruler.add_patterns([{"label": "ORG", "pattern": "Apple"}])
nlp.to_disk("/path/to/pipeline")

The saved pipeline now includes the "entity_ruler" in its config.cfg and the pipeline directory contains a file entityruler.jsonl with the patterns. When you load the pipeline back in, all pipeline components will be restored and deserialized – including the entity ruler. This lets you ship powerful pipeline packages with binary weights and rules included!

Using a large number of phrase patterns

When using a large amount of phrase patterns (roughly > 10000) it’s useful to understand how the add_patterns function of the entity ruler works. For each phrase pattern, the EntityRuler calls the nlp object to construct a doc object. This happens in case you try to add the EntityRuler at the end of an existing pipeline with, for example, a POS tagger and want to extract matches based on the pattern’s POS signature. In this case you would pass a config value of "phrase_matcher_attr": "POS" for the entity ruler.

Running the full language pipeline across every pattern in a large list scales linearly and can therefore take a long time on large amounts of phrase patterns. As of spaCy v2.2.4 the add_patterns function has been refactored to use nlp.pipe on all phrase patterns resulting in about a 10x-20x speed up with 5,000-100,000 phrase patterns respectively. Even with this speedup (but especially if you’re using an older version) the add_patterns function can still take a long time. An easy workaround to make this function run faster is disabling the other language pipes while adding the phrase patterns.

ruler = nlp.add_pipe("entity_ruler")
patterns = [{"label": "TEST", "pattern": str(i)} for i in range(100000)]
with nlp.select_pipes(enable="tagger"):
    ruler.add_patterns(patterns)

Combining models and rules

You can combine statistical and rule-based components in a variety of ways. Rule-based components can be used to improve the accuracy of statistical models, by presetting tags, entities or sentence boundaries for specific tokens. The statistical models will usually respect these preset annotations, which sometimes improves the accuracy of other decisions. You can also use rule-based components after a statistical model to correct common errors. Finally, rule-based components can reference the attributes set by statistical models, in order to implement more abstract logic.

Example: Expanding named entities

When using a trained named entity recognition model to extract information from your texts, you may find that the predicted span only includes parts of the entity you’re looking for. Sometimes, this happens if statistical model predicts entities incorrectly. Other times, it happens if the way the entity type was defined in the original training corpus doesn’t match what you need for your application.

For example, the corpus spaCy’s English pipelines were trained on defines a PERSON entity as just the person name, without titles like “Mr.” or “Dr.”. This makes sense, because it makes it easier to resolve the entity type back to a knowledge base. But what if your application needs the full names, including the titles?

import spacy

nlp = spacy.load("en_core_web_sm")
doc = nlp("Dr. Alex Smith chaired first board meeting of Acme Corp Inc.")
print([(ent.text, ent.label_) for ent in doc.ents])

While you could try and teach the model a new definition of the PERSON entity by updating it with more examples of spans that include the title, this might not be the most efficient approach. The existing model was trained on over 2 million words, so in order to completely change the definition of an entity type, you might need a lot of training examples. However, if you already have the predicted PERSON entities, you can use a rule-based approach that checks whether they come with a title and if so, expands the entity span by one token. After all, what all titles in this example have in common is that if they occur, they occur in the previous token right before the person entity.

from spacy.language import Language
from spacy.tokens import Span

@Language.component("expand_person_entities")
def expand_person_entities(doc):
    new_ents = []
    for ent in doc.ents:
        # Only check for title if it's a person and not the first token
        if ent.label_ == "PERSON" and ent.start != 0:            prev_token = doc[ent.start - 1]            if prev_token.text in ("Dr", "Dr.", "Mr", "Mr.", "Ms", "Ms."):                new_ent = Span(doc, ent.start - 1, ent.end, label=ent.label)                new_ents.append(new_ent)            else:
                new_ents.append(ent)
        else:
            new_ents.append(ent)
    doc.ents = new_ents
    return doc

The above function takes a Doc object, modifies its doc.ents and returns it. Using the @Language.component decorator, we can register it as a pipeline component so it can run automatically when processing a text. We can use nlp.add_pipe to add it to the current pipeline.

import spacy
from spacy.language import Language
from spacy.tokens import Span

nlp = spacy.load("en_core_web_sm")

@Language.component("expand_person_entities")
def expand_person_entities(doc):
    new_ents = []
    for ent in doc.ents:
        if ent.label_ == "PERSON" and ent.start != 0:
            prev_token = doc[ent.start - 1]
            if prev_token.text in ("Dr", "Dr.", "Mr", "Mr.", "Ms", "Ms."):
                new_ent = Span(doc, ent.start - 1, ent.end, label=ent.label)
                new_ents.append(new_ent)
        else:
            new_ents.append(ent)
    doc.ents = new_ents
    return doc

# Add the component after the named entity recognizer
nlp.add_pipe("expand_person_entities", after="ner")

doc = nlp("Dr. Alex Smith chaired first board meeting of Acme Corp Inc.")
print([(ent.text, ent.label_) for ent in doc.ents])

An alternative approach would be to use an extension attribute like ._.person_title and add it to Span objects (which includes entity spans in doc.ents). The advantage here is that the entity text stays intact and can still be used to look up the name in a knowledge base. The following function takes a Span object, checks the previous token if it’s a PERSON entity and returns the title if one is found. The Span.doc attribute gives us easy access to the span’s parent document.

def get_person_title(span):
    if span.label_ == "PERSON" and span.start != 0:
        prev_token = span.doc[span.start - 1]
        if prev_token.text in ("Dr", "Dr.", "Mr", "Mr.", "Ms", "Ms."):
            return prev_token.text

We can now use the Span.set_extension method to add the custom extension attribute "person_title", using get_person_title as the getter function.

import spacy
from spacy.tokens import Span

nlp = spacy.load("en_core_web_sm")

def get_person_title(span):
    if span.label_ == "PERSON" and span.start != 0:
        prev_token = span.doc[span.start - 1]
        if prev_token.text in ("Dr", "Dr.", "Mr", "Mr.", "Ms", "Ms."):
            return prev_token.text

# Register the Span extension as 'person_title'
Span.set_extension("person_title", getter=get_person_title)

doc = nlp("Dr Alex Smith chaired first board meeting of Acme Corp Inc.")
print([(ent.text, ent.label_, ent._.person_title) for ent in doc.ents])

Example: Using entities, part-of-speech tags and the dependency parse

Let’s say you want to parse professional biographies and extract the person names and company names, and whether it’s a company they’re currently working at, or a previous company. One approach could be to try and train a named entity recognizer to predict CURRENT_ORG and PREVIOUS_ORG – but this distinction is very subtle and something the entity recognizer may struggle to learn. Nothing about “Acme Corp Inc.” is inherently “current” or “previous”.

However, the syntax of the sentence holds some very important clues: we can check for trigger words like “work”, whether they’re past tense or present tense, whether company names are attached to it and whether the person is the subject. All of this information is available in the part-of-speech tags and the dependency parse.

import spacy

nlp = spacy.load("en_core_web_sm")
doc = nlp("Alex Smith worked at Acme Corp Inc.")
print([(ent.text, ent.label_) for ent in doc.ents])

In this example, “worked” is the root of the sentence and is a past tense verb. Its subject is “Alex Smith”, the person who worked. “at Acme Corp Inc.” is a prepositional phrase attached to the verb “worked”. To extract this relationship, we can start by looking at the predicted PERSON entities, find their heads and check whether they’re attached to a trigger word like “work”. Next, we can check for prepositional phrases attached to the head and whether they contain an ORG entity. Finally, to determine whether the company affiliation is current, we can check the head’s part-of-speech tag.

person_entities = [ent for ent in doc.ents if ent.label_ == "PERSON"]
for ent in person_entities:
    # Because the entity is a span, we need to use its root token. The head
    # is the syntactic governor of the person, e.g. the verb
    head = ent.root.head
    if head.lemma_ == "work":
        # Check if the children contain a preposition
        preps = [token for token in head.children if token.dep_ == "prep"]
        for prep in preps:
            # Check if tokens part of ORG entities are in the preposition's
            # children, e.g. at -> Acme Corp Inc.
            orgs = [token for token in prep.children if token.ent_type_ == "ORG"]
            # If the verb is in past tense, the company was a previous company
            print({"person": ent, "orgs": orgs, "past": head.tag_ == "VBD"})

To apply this logic automatically when we process a text, we can add it to the nlp object as a custom pipeline component. The above logic also expects that entities are merged into single tokens. spaCy ships with a handy built-in merge_entities that takes care of that. Instead of just printing the result, you could also write it to custom attributes on the entity Span – for example ._.orgs or ._.prev_orgs and ._.current_orgs.

import spacy
from spacy.language import Language
from spacy import displacy

nlp = spacy.load("en_core_web_sm")

@Language.component("extract_person_orgs")
def extract_person_orgs(doc):
    person_entities = [ent for ent in doc.ents if ent.label_ == "PERSON"]
    for ent in person_entities:
        head = ent.root.head
        if head.lemma_ == "work":
            preps = [token for token in head.children if token.dep_ == "prep"]
            for prep in preps:
                orgs = [token for token in prep.children if token.ent_type_ == "ORG"]
                print({'person': ent, 'orgs': orgs, 'past': head.tag_ == "VBD"})
    return doc

# To make the entities easier to work with, we'll merge them into single tokens
nlp.add_pipe("merge_entities")
nlp.add_pipe("extract_person_orgs")

doc = nlp("Alex Smith worked at Acme Corp Inc.")
# If you're not in a Jupyter / IPython environment, use displacy.serve
displacy.render(doc, options={"fine_grained": True})

If you change the sentence structure above, for example to “was working”, you’ll notice that our current logic fails and doesn’t correctly detect the company as a past organization. That’s because the root is a participle and the tense information is in the attached auxiliary “was”:

To solve this, we can adjust the rules to also check for the above construction:

@Language.component("extract_person_orgs")
def extract_person_orgs(doc):
    person_entities = [ent for ent in doc.ents if ent.label_ == "PERSON"]
    for ent in person_entities:
        head = ent.root.head
        if head.lemma_ == "work":
            preps = [token for token in head.children if token.dep_ == "prep"]
            for prep in preps:
                orgs = [t for t in prep.children if t.ent_type_ == "ORG"]
                aux = [token for token in head.children if token.dep_ == "aux"]                past_aux = any(t.tag_ == "VBD" for t in aux)                past = head.tag_ == "VBD" or head.tag_ == "VBG" and past_aux                print({'person': ent, 'orgs': orgs, 'past': past})
    return doc

In your final rule-based system, you may end up with several different code paths to cover the types of constructions that occur in your data.