[nlp] How do I do word Stemming or Lemmatization?

I've tried PorterStemmer and Snowball but both don't work on all words, missing some very common ones.

My test words are: "cats running ran cactus cactuses cacti community communities", and both get less than half right.

See also:

This question is related to nlp stemming lemmatization

The answer is


The stemmer vs lemmatizer debates goes on. It's a matter of preferring precision over efficiency. You should lemmatize to achieve linguistically meaningful units and stem to use minimal computing juice and still index a word and its variations under the same key.

See Stemmers vs Lemmatizers

Here's an example with python NLTK:

>>> sent = "cats running ran cactus cactuses cacti community communities"
>>> from nltk.stem import PorterStemmer, WordNetLemmatizer
>>>
>>> port = PorterStemmer()
>>> " ".join([port.stem(i) for i in sent.split()])
'cat run ran cactu cactus cacti commun commun'
>>>
>>> wnl = WordNetLemmatizer()
>>> " ".join([wnl.lemmatize(i) for i in sent.split()])
'cat running ran cactus cactus cactus community community'

Do a search foR Lucene, im not sure if theres a PHP port but i do know Lucene is available for many platforms. Lucene is an OSS (from Apache) indexing and search library. Naturally it and community extras might have something interesting to look at. At the very least you can learn how its done in one language so you can translate the "idea" into PHP


df_plots = pd.read_excel("Plot Summary.xlsx", index_col = 0)
df_plots
# Printing first sentence of first row and last sentence of last row
nltk.sent_tokenize(df_plots.loc[1].Plot)[0] + nltk.sent_tokenize(df_plots.loc[len(df)].Plot)[-1]

# Calculating length of all plots by words
df_plots["Length"] = df_plots.Plot.apply(lambda x : 
len(nltk.word_tokenize(x)))

print("Longest plot is for season"),
print(df_plots.Length.idxmax())

print("Shortest plot is for season"),
print(df_plots.Length.idxmin())



#What is this show about? (What are the top 3 words used , excluding the #stop words, in all the #seasons combined)

word_sample = list(["struggled", "died"])
word_list = nltk.pos_tag(word_sample)
[wnl.lemmatize(str(word_list[index][0]), pos = word_list[index][1][0].lower()) for index in range(len(word_list))]

# Figure out the stop words
stop = (stopwords.words('english'))

# Tokenize all the plots
df_plots["Tokenized"] = df_plots.Plot.apply(lambda x : nltk.word_tokenize(x.lower()))

# Remove the stop words
df_plots["Filtered"] = df_plots.Tokenized.apply(lambda x : (word for word in x if word not in stop))

# Lemmatize each word
wnl = WordNetLemmatizer()
df_plots["POS"] = df_plots.Filtered.apply(lambda x : nltk.pos_tag(list(x)))
# df_plots["POS"] = df_plots.POS.apply(lambda x : ((word[1] = word[1][0] for word in word_list) for word_list in x))
df_plots["Lemmatized"] = df_plots.POS.apply(lambda x : (wnl.lemmatize(x[index][0], pos = str(x[index][1][0]).lower()) for index in range(len(list(x)))))



#Which Season had the highest screenplay of "Jesse" compared to "Walt" 
#Screenplay of Jesse =(Occurences of "Jesse")/(Occurences of "Jesse"+ #Occurences of "Walt")

df_plots.groupby("Season").Tokenized.sum()

df_plots["Share"] = df_plots.groupby("Season").Tokenized.sum().apply(lambda x : float(x.count("jesse") * 100)/float(x.count("jesse") + x.count("walter") + x.count("walt")))

print("The highest times Jesse was mentioned compared to Walter/Walt was in season"),
print(df_plots["Share"].idxmax())
#float(df_plots.Tokenized.sum().count('jesse')) * 100 / #float((df_plots.Tokenized.sum().count('jesse') + #df_plots.Tokenized.sum().count('walt') + #df_plots.Tokenized.sum().count('walter')))

Martin Porter's official page contains a Porter Stemmer in PHP as well as other languages.

If you're really serious about good stemming though you're going to need to start with something like the Porter Algorithm, refine it by adding rules to fix incorrect cases common to your dataset, and then finally add a lot of exceptions to the rules. This can be easily implemented with key/value pairs (dbm/hash/dictionaries) where the key is the word to look up and the value is the stemmed word to replace the original. A commercial search engine I worked on once ended up with 800 some exceptions to a modified Porter algorithm.


Martin Porter wrote Snowball (a language for stemming algorithms) and rewrote the "English Stemmer" in Snowball. There are is an English Stemmer for C and Java.

He explicitly states that the Porter Stemmer has been reimplemented only for historical reasons, so testing stemming correctness against the Porter Stemmer will get you results that you (should) already know.

From http://tartarus.org/~martin/PorterStemmer/index.html (emphasis mine)

The Porter stemmer should be regarded as ‘frozen’, that is, strictly defined, and not amenable to further modification. As a stemmer, it is slightly inferior to the Snowball English or Porter2 stemmer, which derives from it, and which is subjected to occasional improvements. For practical work, therefore, the new Snowball stemmer is recommended. The Porter stemmer is appropriate to IR research work involving stemming where the experiments need to be exactly repeatable.

Dr. Porter suggests to use the English or Porter2 stemmers instead of the Porter stemmer. The English stemmer is what's actually used in the demo site as @StompChicken has answered earlier.


I highly recommend using Spacy (base text parsing & tagging) and Textacy (higher level text processing built on top of Spacy).

Lemmatized words are available by default in Spacy as a token's .lemma_ attribute and text can be lemmatized while doing a lot of other text preprocessing with textacy. For example while creating a bag of terms or words or generally just before performing some processing that requires it.

I'd encourage you to check out both before writing any code, as this may save you a lot of time!


The most current version of the stemmer in NLTK is Snowball.

You can find examples on how to use it here:

http://nltk.googlecode.com/svn/trunk/doc/api/nltk.stem.snowball2-pysrc.html#demo


.Net lucene has an inbuilt porter stemmer. You can try that. But note that porter stemming does not consider word context when deriving the lemma. (Go through the algorithm and its implementation and you will see how it works)


Take a look at LemmaGen - open source library written in C# 3.0.

Results for your test words (http://lemmatise.ijs.si/Services)

  • cats -> cat
  • running
  • ran -> run
  • cactus
  • cactuses -> cactus
  • cacti -> cactus
  • community
  • communities -> community

If you know Python, The Natural Language Toolkit (NLTK) has a very powerful lemmatizer that makes use of WordNet.

Note that if you are using this lemmatizer for the first time, you must download the corpus prior to using it. This can be done by:

>>> import nltk
>>> nltk.download('wordnet')

You only have to do this once. Assuming that you have now downloaded the corpus, it works like this:

>>> from nltk.stem.wordnet import WordNetLemmatizer
>>> lmtzr = WordNetLemmatizer()
>>> lmtzr.lemmatize('cars')
'car'
>>> lmtzr.lemmatize('feet')
'foot'
>>> lmtzr.lemmatize('people')
'people'
>>> lmtzr.lemmatize('fantasized','v')
'fantasize'

There are other lemmatizers in the nltk.stem module, but I haven't tried them myself.


If I may quote my answer to the question StompChicken mentioned:

The core issue here is that stemming algorithms operate on a phonetic basis with no actual understanding of the language they're working with.

As they have no understanding of the language and do not run from a dictionary of terms, they have no way of recognizing and responding appropriately to irregular cases, such as "run"/"ran".

If you need to handle irregular cases, you'll need to either choose a different approach or augment your stemming with your own custom dictionary of corrections to run after the stemmer has done its thing.


You could use the Morpha stemmer. UW has uploaded morpha stemmer to Maven central if you plan to use it from a Java application. There's a wrapper that makes it much easier to use. You just need to add it as a dependency and use the edu.washington.cs.knowitall.morpha.MorphaStemmer class. Instances are threadsafe (the original JFlex had class fields for local variables unnecessarily). Instantiate a class and run morpha and the word you want to stem.

new MorphaStemmer().morpha("climbed") // goes to "climb"

I tried your list of terms on this snowball demo site and the results look okay....

  • cats -> cat
  • running -> run
  • ran -> ran
  • cactus -> cactus
  • cactuses -> cactus
  • community -> communiti
  • communities -> communiti

A stemmer is supposed to turn inflected forms of words down to some common root. It's not really a stemmer's job to make that root a 'proper' dictionary word. For that you need to look at morphological/orthographic analysers.

I think this question is about more or less the same thing, and Kaarel's answer to that question is where I took the second link from.


The top python packages (in no specific order) for lemmatization are: spacy, nltk, gensim, pattern, CoreNLP and TextBlob. I prefer spaCy and gensim's implementation (based on pattern) because they identify the POS tag of the word and assigns the appropriate lemma automatically. The gives more relevant lemmas, keeping the meaning intact.

If you plan to use nltk or TextBlob, you need to take care of finding the right POS tag manually and the find the right lemma.

Lemmatization Example with spaCy:

# Run below statements in terminal once. 
pip install spacy
spacy download en

import spacy

# Initialize spacy 'en' model
nlp = spacy.load('en', disable=['parser', 'ner'])

sentence = "The striped bats are hanging on their feet for best"

# Parse
doc = nlp(sentence)

# Extract the lemma
" ".join([token.lemma_ for token in doc])
#> 'the strip bat be hang on -PRON- foot for good'

Lemmatization Example With Gensim:

from gensim.utils import lemmatize
sentence = "The striped bats were hanging on their feet and ate best fishes"
lemmatized_out = [wd.decode('utf-8').split('/')[0] for wd in lemmatize(sentence)]
#> ['striped', 'bat', 'be', 'hang', 'foot', 'eat', 'best', 'fish']

The above examples were borrowed from in this lemmatization page.


I use stanford nlp to perform lemmatization. I have been stuck up with a similar problem in the last few days. All thanks to stackoverflow to help me solve the issue .

import java.util.*; 
import edu.stanford.nlp.pipeline.*;
import edu.stanford.nlp.ling.*; 
import edu.stanford.nlp.ling.CoreAnnotations.*;  

public class example
{
    public static void main(String[] args)
    {
        Properties props = new Properties(); 
        props.put("annotators", "tokenize, ssplit, pos, lemma"); 
        pipeline = new StanfordCoreNLP(props, false);
        String text = /* the string you want */; 
        Annotation document = pipeline.process(text);  

        for(CoreMap sentence: document.get(SentencesAnnotation.class))
        {    
            for(CoreLabel token: sentence.get(TokensAnnotation.class))
            {       
                String word = token.get(TextAnnotation.class);      
                String lemma = token.get(LemmaAnnotation.class); 
                System.out.println("lemmatized version :" + lemma);
            }
        }
    }
}

It also might be a good idea to use stopwords to minimize output lemmas if it's used later in classificator. Please take a look at coreNlp extension written by John Conwell.


Look into WordNet, a large lexical database for the English language:

http://wordnet.princeton.edu/

There are APIs for accessing it in several languages.


This looks interesting: MIT Java WordnetStemmer: http://projects.csail.mit.edu/jwi/api/edu/mit/jwi/morph/WordnetStemmer.html


Based on various answers on Stack Overflow and blogs I've come across, this is the method I'm using, and it seems to return real words quite well. The idea is to split the incoming text into an array of words (use whichever method you'd like), and then find the parts of speech (POS) for those words and use that to help stem and lemmatize the words.

You're sample above doesn't work too well, because the POS can't be determined. However, if we use a real sentence, things work much better.

import nltk
from nltk.corpus import wordnet

lmtzr = nltk.WordNetLemmatizer().lemmatize


def get_wordnet_pos(treebank_tag):
    if treebank_tag.startswith('J'):
        return wordnet.ADJ
    elif treebank_tag.startswith('V'):
        return wordnet.VERB
    elif treebank_tag.startswith('N'):
        return wordnet.NOUN
    elif treebank_tag.startswith('R'):
        return wordnet.ADV
    else:
        return wordnet.NOUN


def normalize_text(text):
    word_pos = nltk.pos_tag(nltk.word_tokenize(text))
    lemm_words = [lmtzr(sw[0], get_wordnet_pos(sw[1])) for sw in word_pos]

    return [x.lower() for x in lemm_words]

print(normalize_text('cats running ran cactus cactuses cacti community communities'))
# ['cat', 'run', 'ran', 'cactus', 'cactuses', 'cacti', 'community', 'community']

print(normalize_text('The cactus ran to the community to see the cats running around cacti between communities.'))
# ['the', 'cactus', 'run', 'to', 'the', 'community', 'to', 'see', 'the', 'cat', 'run', 'around', 'cactus', 'between', 'community', '.']

http://wordnet.princeton.edu/man/morph.3WN

For a lot of my projects, I prefer the lexicon-based WordNet lemmatizer over the more aggressive porter stemming.

http://wordnet.princeton.edu/links#PHP has a link to a PHP interface to the WN APIs.


Try this one here: http://www.twinword.com/lemmatizer.php

I entered your query in the demo "cats running ran cactus cactuses cacti community communities" and got ["cat", "running", "run", "cactus", "cactus", "cactus", "community", "community"] with the optional flag ALL_TOKENS.

Sample Code

This is an API so you can connect to it from any environment. Here is what the PHP REST call may look like.

// These code snippets use an open-source library. http://unirest.io/php
$response = Unirest\Request::post([ENDPOINT],
  array(
    "X-Mashape-Key" => [API KEY],
    "Content-Type" => "application/x-www-form-urlencoded",
    "Accept" => "application/json"
  ),
  array(
    "text" => "cats running ran cactus cactuses cacti community communities"
  )
);

In Java, i use tartargus-snowball to stemming words

Maven:

<dependency>
        <groupId>org.apache.lucene</groupId>
        <artifactId>lucene-snowball</artifactId>
        <version>3.0.3</version>
        <scope>test</scope>
</dependency>

Sample code:

SnowballProgram stemmer = new EnglishStemmer();
String[] words = new String[]{
    "testing",
    "skincare",
    "eyecare",
    "eye",
    "worked",
    "read"
};
for (String word : words) {
    stemmer.setCurrent(word);
    stemmer.stem();
    //debug
    logger.info("Origin: " + word + " > " + stemmer.getCurrent());// result: test, skincar, eyecar, eye, work, read
}