A Rust library to support natural language processing with pure Rust implementation and Python bindings
Rust Docs | Crates Home Page | Tests | NER-Kit
The rsnltk
library integrates various existing Python-written NLP toolkits for powerful text analysis in Rust-based applications.
This toolkit is based on the Python-written Stanza and other important NLP crates.
A list of functions from Stanza and others we bind here include:
- Tokenize
- Sentence Segmentation
- Multi-Word Token Expansion
- Part-of-Speech & Morphological Features
- Named Entity Recognition
- Sentiment Analysis
- Language Identification
- Dependency Tree Analysis
Some amazing crates are also included in rsnltk
but with simplified APIs for actual use:
Additionally, we can calculate the similarity between words based on WordNet though the semantic-kit
PyPI project via pip install semantic-kit
.
-
Make sure you install Python 3.6.6+ and PIP environment in your computer. Type
python -V
in the Terminal should print no error message; -
Install our Python-based ner-kit (version>=0.0.5a2) for binding the
Stanza
package viapip install ner-kit==0.0.5a2
; -
Then, Rust should be also installed in your computer. I use IntelliJ to develop Rust-based applications, where you can write Rust codes;
-
Create a simple Rust application project with a
main()
function. -
Add the
rsnltk
dependency to theCargo.toml
file, keep up the Latest version. -
After you add the
rsnltk
dependency in thetoml file
, install necessary language models from Stanza using the following Rust code for the first time you use this package.
fn init_rsnltk_and_test(){
// 1. first install the necessary language models
// using language codes
let list_lang=vec!["en","zh"];
//e.g. you install two language models,
// namely, for English and Chinese text analysis.
download_langs(list_lang);
// 2. then do test NLP tasks
let text="I like Beijing!";
let lang="en";
// 2. Uncomment the below codes for Chinese NER
// let text="我喜欢北京、上海和纽约!";
// let lang="zh";
let list_ner=ner(text,lang);
for ner in list_ner{
println!("{:?}",ner);
}
}
Or you can manually install those language models via the Python-written ner-kit
package which provides more features in using Stanza. Go to: ner-kit
If no error occurs in the above example, then it works. Finally, you can try the following advanced example usage.
Currently, we tested the use of English and Chinese language models; however, other language models should work as well.
Example 1: Part-of-speech Analysis
fn test_pos(){
//let text="我喜欢北京、上海和纽约!";
//let lang="zh";
let text="I like apple";
let lang="en";
let list_result=pos(text,lang);
for word in list_result{
println!("{:?}",word);
}
}
Example 2: Sentiment Analysis
fn test_sentiment(){
//let text="I like Beijing!";
//let lang="en";
let text="我喜欢北京";
let lang="zh";
let sentiments=sentiment(text,lang);
for sen in sentiments{
println!("{:?}",sen);
}
}
Example 3: Named Entity Recognition
fn test_ner(){
// 1. for English NER
let text="I like Beijing!";
let lang="en";
// 2. Uncomment the below codes for Chinese NER
// let text="我喜欢北京、上海和纽约!";
// let lang="zh";
let list_ner=ner(text,lang);
for ner in list_ner{
println!("{:?}",ner);
}
}
Example 4: Tokenize for Multiple Languages
fn test_tokenize(){
let text="我喜欢北京、上海和纽约!";
let lang="zh";
let list_result=tokenize(text,lang);
for ner in list_result{
println!("{:?}",ner);
}
}
Example 5: Tokenize Sentence
fn test_tokenize_sentence(){
let text="I like apple. Do you like it? No, I am not sure!";
let lang="en";
let list_sentences=tokenize_sentence(text,lang);
for sentence in list_sentences{
println!("Sentence: {}",sentence);
}
}
Example 6: Language Identification
fn test_lang(){
let list_text = vec!["I like Beijing!",
"我喜欢北京!",
"Bonjour le monde!"];
let list_result=lang(list_text);
for lang in list_result{
println!("{:?}",lang);
}
}
Example 7: MWT expand
fn test_mwt_expand(){
let text="Nous avons atteint la fin du sentier.";
let lang="fr";
let list_result=mwt_expand(text,lang);
}
Example 8: Estimate the similarity between words in WordNet
You need to firstly install semantic-kit
PyPI package!
fn test_wordnet_similarity(){
let s1="dog.n.1";
let s2="cat.n.2";
let sims=wordnet_similarity(s1,s2);
for sim in sims{
println!("{:?}",sim);
}
}
Example 9: Obtain a dependency tree from a text
fn test_dependency_tree(){
let text="I like you. Do you like me?";
let lang="en";
let list_results=dependency_tree(text,lang);
for list_token in list_results{
for token in list_token{
println!("{:?}",token)
}
}
}
Example 1: Word2Vec similarity
fn test_open_wv_bin(){
let wv_model=wv_get_model("GoogleNews-vectors-negative300.bin");
let positive = vec!["woman", "king"];
let negative = vec!["man"];
println!("analogy: {:?}", wv_analogy(&wv_model,positive, negative, 10));
println!("cosine: {:?}", wv_cosine(&wv_model,"man", 10));
}
Example 2: Text summarization
use rsnltk::native::summarizer::*;
fn test_summarize(){
let text="Some large txt...";
let stopwords=&[];
let summarized_text=summarize(text,stopwords,5);
println!("{}",summarized_text);
}
Example 3: Get token list from English strings
use rsnltk::native::token::get_token_list;
fn test_get_token_list(){
let s="Hello, Rust. How are you?";
let result=get_token_list(s);
for r in result{
println!("{}\t{:?}",r.text,r);
}
}
Example 4: Word segmentation for some language where no space exists between terms, e.g. Chinese text.
We implement three word segmentation methods in this version:
- Forward Maximum Matching (fmm), which is baseline method
- Backward Maximum Matching (bmm), which is considered better
- Bidirectional Maximum Matching (bimm), high accuracy but low speed
use rsnltk::native::segmentation::*;
fn test_real_word_segmentation(){
let dict_path="30wdict.txt"; // empty if only for tokenizing
let stop_path="baidu_stopwords.txt";// empty when no stop words
let _sentence="美国太空总署希望,在深海的探险发现将有助于解开一些外太空的秘密,\
同时也可以测试前往太阳系其他星球探险所需的一些设备和实验。";
let meaningful_words=get_segmentation(_sentence,dict_path,stop_path, "bimm");
// bimm can be changed to fmm or bmm.
println!("Result: {:?}",meaningful_words);
}
Thank Stanford NLP Group for their hard work in Stanza.
The rsnltk
library with MIT License is provided by Donghua Chen.