Rakuten MA

Introduction

Rakuten MA (morphological analyzer) is a morphological analyzer (word segmentor + PoS Tagger)
for Chinese and Japanese written purely in JavaScript.

Rakuten MA has the following unique features:
- Pure JavaScript implementation. Works both on modern browsers and node.js.
- Implements a language independent character tagging model. Outputs word segmentation and PoS tags for Chinese/Japanese.
- Supports incremental update of models by online learning (Soft Confidence Weighted, Wang et al. ICML 2012).
- Customizable feature set.
- Supports feature hashing, quantization, and pruning for compact model representation.
- Bundled with Chinese and Japanese models trained from general corpora (CTB [Xue et al. 2005] and BCCWJ [Maekawa 2008]) and E-commerce corpora.

Demo

You can try Rakuten MA on the demo page. (It may take a while to load this page.)

Usage

$3

Since Rakuten MA is a JavaScript library, there's no need for installation. Clone the git repository as

git clone https://github.com/rakuten-nlp/rakutenma.git

or download the zip archive from here: https://github.com/rakuten-nlp/rakutenma/archive/master.zip

If you have Node.js installed, you can run the demo by

node demo.js

which is identical to the usage example below.

$3

// RakutenMA demo

// Load necessary libraries
var RakutenMA = require('./rakutenma');
var fs = require('fs');

// Initialize a RakutenMA instance
// with an empty model and the default ja feature set
var rma = new RakutenMA();
rma.featset = RakutenMA.default_featset_ja;

// Let's analyze a sample sentence (from http://tatoeba.org/jpn/sentences/show/103809)
// With a disastrous result, since the model is empty!
console.log(rma.tokenize("彼は新しい仕事できっと成功するだろう。"));

// Feed the model with ten sample sentences from tatoeba.com
var tatoeba = JSON.parse(fs.readFileSync("tatoeba.json"));
for (var i = 0; i < 10; i ++) {
rma.train_one(tatoeba[i]);
}

// Now what does the result look like?
console.log(rma.tokenize("彼は新しい仕事できっと成功するだろう。"));

// Initialize a RakutenMA instance with a pre-trained model
var model = JSON.parse(fs.readFileSync("model_ja.json"));
rma = new RakutenMA(model, 1024, 0.007812); // Specify hyperparameter for SCW (for demonstration purpose)
rma.featset = RakutenMA.default_featset_ja;

// Set the feature hash function (15bit)
rma.hash_func = RakutenMA.create_hash_func(15);

// Tokenize one sample sentence
console.log(rma.tokenize("うらにわにはにわにわとりがいる"));

// Re-train the model feeding the right answer (pairs of [token, PoS tag])
var res = rma.train_one(
[["うらにわ","N-nc"],
["に","P-k"],
["は","P-rj"],
["にわ","N-n"],
["にわとり","N-nc"],
["が","P-k"],
["いる","V-c"]]);
// The result of train_one contains:
// sys: the system output (using the current model)
// ans: answer fed by the user
// update: whether the model was updated
console.log(res);

// Now what does the result look like?
console.log(rma.tokenize("うらにわにはにわにわとりがいる"));

$3

Include the following code snippet in the of your HTML.

The analysis and result looks like this:

$3

1. Load an existing model, e.g., model = JSON.parse(fs.readFileSync("model_file")); then rma = new RakutenMA(model); or rma.set_model(model);
2. Specify featset depending on your langage (e.g., rma.featset = RakutenMA.default_featset_zh; for Chinese and rma.featset = RakutenMA.default_featset_ja; for Japanese).
3. Remember to use 15-bit feature hashing function (rma.hash_func = RakutenMA.create_hash_func(15);) when using the bundled models (model_zh.json and model_ja.json).
4. Use rma.tokenize(input) to analyze your input.

$3

1. Prepare your training corpus (a set of training sentences where a sentence is just an array of correct [token, PoS tag].)
2. Initialize a RakutenMA instance with new RakutenMA().
3. Specify featset. (and optionally, ctype_func, hash_func, etc.)
4. Feed your training sentences one by one (from the first one to the last) to the train_one(sent) method.
5. Usually SCW converges enough after one epoch (one pass through the entire training corpus) but you can repeat Step 4. to achieve even better performance.

See scripts/train_zh.js (for Chinese) and scripts/train_ja.js (for Japanese) to see an example showing how to train your own model.

$3

1. Load an existing model and initialize a RakutenMA instance. (see "Using bundled models to analyze Chinese/Japanese sentences" above)
2. Prepare your training data (this could be as few as a couple of sentences, depending on what and how much you want to "re-train".)
3. Feed your training sentences one by one to the train_one(sent) method.

$3

The model size could still be a problem for client-side distribution even after applying feature hashing.
We included a script scripts/minify.js which applies feature quantization
(see [Hagiwara and Sekine COLING 2014] for the details) to reduce the trained model size.

You can run it node scripts/minify.js [input_model_file] [output_model_file] to make a minified version of the model file.
Remember: it also deletes the "sigma" part of the trained model, meaning that you are no longer able to re-train the minified model. If necessary, re-train the model first, then minify it.

API Documentation

| Constructor | Description |
| ----------------------------| ------------------------------------------- |
| RakutenMA(model, phi, c) | Creates a new RakutenMA instance. model (optional) specifies the model object to initialize the RakutenMA instance with. phi and c (both optional) are hyper parameters of SCW (default: phi = 2048, c = 0.003906). |

| Methods | Description |
| ----------------------------| ------------------------------------------- |
| tokenize(input) | Tokenizes input (string) and returns tokenized result ([token, PoS tag] pairs). |
| train_one(sent) | Updates the current model (if necessary) using the given answer sent ([token, PoS tag] pairs). The return value is an object with three properties ans, sys, and updated, where ans is the given answer (same as sent), sys is the system output using the (old) model, and updated is a binary (True/False) flag meaning whether the model was updated (because sys was different from ans) or not.|
| set_model(model) | Sets the Rakuten MA instance's model to model. |
| set_tag_scheme(scheme) | Sets the sequential labeling tag scheme. Currently, "IOB2" and "SBIEO" are supported. Specifying other tag schemes causes an exception. |

| Properties | Description |
| ----------------------------| ------------------------------------------- |
| featset | Specifies an array of feature templates (string) used for analysis. You can use RakutenMA.default_featset_ja and RakutenMA.default_featset_zh as the default feature sets for Japanese and Chinese, respectively. See below ("Supported feature templates") for the details of feature templates. |
| ctype_func | Specifies the function used to convert a character to its character type. RakutenMA.ctype_ja_default_func is the default character type function used for Japanese. Alternatively, you can call RakutenMA.create_ctype_chardic_func(chardic) to create a character type function which takes a character to look it up in chardic and return its value. (For example, RakutenMA.create_ctype_chardic_func({"A": "type1"}) returns a function f where f("A") returns "type1" and [] otherwise.) |
| hash_func | Specifies the hash function to use for feature hashing. Default = undefined (no feature hashing). A feature hashing function with bit-bit hash space can be created by calling RakutenMA.create_hash_func(bit). |

Terms and Conditions

Distribution, modification, and academic/commercial use of Rakuten MA is permitted, provided that
you conform with Apache License version 2.0 http://www.apache.org/licenses/LICENSE-2.0.html.

If you are using Rakuten MA for research purposes, please cite our paper on Rakuten MA [Hagiwara and Sekine 2014]

FAQ (Frequently Asked Questions)

Q. What are supported browsers and Node.js versions?

- A. We confirmed that Rakuten MA runs in the following environments:
- Internet Explorer 8 (ver. 8.0.7601.17414 or above)
- Google Chrome (ver. 35.0.1916.153 or above)
- Firefox (ver. 16.0.2 or above)
- Safari (ver. 6.1.5 or above)
- Node.js (ver. 0.10.13 or above)

Q. Is commercial use permitted?
- A. Yes, as long as you follow the terms and conditions. See "Terms and Conditions" above for the details.

Q. I found a bug / analysis error / etc. Where should I report?
- A. Please create an issue at Github issues https://github.com/rakuten-nlp/rakutenma/issues.
- Alternatively, you can create a pull request if you modify the code. Rakuten MA has a test suite using Jasmine http://jasmine.github.io/. Please make sure all the tests pass (no errors after running jasmine-node spec) and write your own (if necessary) before submitting a pull request.

Q. Tokenization results look strange (specifically, the sentence is split up to individual characters with no PoS tags)
- A. Check if you are using the same feature set (featset) and the feature hashing function (hash_func) used for training. Remember to use 15-bit feature hashing function (rma.hash_func = RakutenMA.create_hash_func(15);) when using the bundled models (model_zh.json and model_ja.json).

Q. What scripts (Simplified/Traditional) are supported for Chinese?
- A. Currently only simplified Chinese is supported.

Q. Can we use the same model file in the JSON format for browsers?
- A. Yes and no. Although internal data structure of models is the same, you need to add assignment (e.g., var model = [JSON representation];) in order to refer to it on browsers. See the difference between model_zh.json (for Node.js) and model_zh.js (for browsers). There is a mini script scripts/convert_for_browser.js which does this for you. We recommend you work on Node.js for model training etc. and then convert it for browser uses.

Appendix

$3

| Feature template | Description |
| ---------------- | ----------------------------- |
| w7 | Character unigram (c-3) |
| w8 | Character unigram (c-2) |
| w9 | Character unigram (c-1) |
| w0 | Character unigram (c0) |
| w1 | Character unigram (c+1) |
| w2 | Character unigram (c+2) |
| w3 | Character unigram (c+3) |
| c7 | Character type unigram (t-3) |
| c8 | Character type unigram (t-2) |
| c9 | Character type unigram (t-1) |
| c0 | Character type unigram (t0) |
| c1 | Character type unigram (t+1) |
| c2 | Character type unigram (t+2) |
| c3 | Character type unigram (t+3) |
| b7 | Character bigram (c-3 c-2) |
| b8 | Character bigram (c-2 c-1) |
| b9 | Character bigram (c-1 c0) |
| b1 | Character bigram (c0 c+1) |
| b2 | Character bigram (c+1 c+2) |
| b3 | Character bigram (c+2 c+3) |
| d7 | Character type bigram (t-3 t-2) |
| d8 | Character type bigram (t-2 t-1) |
| d9 | Character type bigram (t-1 t0) |
| d1 | Character type bigram (t0 t+1) |
| d2 | Character type bigram (t+1 t+2) |
| d3 | Character type bigram (t+2 t+3) |
| others | If you specify a customized feature function in the featset array, the function will be called with two arguments _t and i, where _t is a function which takes a position j and returns the character object at that position, and i is the current position. A character object is an object with two properties c and t which are character and character type, respectively. The return value of that function is used as the feature value. (For example, if you specify a function f(_t, i) which returns _t[i].t;, then it's returning the character type of the current position, which is basically the same as the template c0. )|

$3

| Tag | Description |
| --- | ---------------- |
| AD | Adverb |
| AS | Aspect Particle |
| BA | ba3 (in ba-construction) |
| CC | Coordinating conjunction |
| CD | Cardinal number |
| CS | Subordinating conjunction |
| DEC | de5 (Complementizer/Nominalizer) |
| DEG | de5 (Genitive/Associative) |
| DER | de5 (Resultative) |
| DEV | de5 (Manner) |
| DT | Determiner |
| ETC | Others |
| FW | Foreign word |
| IJ | Interjection |
| JJ | Other noun-modifier |
| LB | bei4 (in long bei-construction) |
| LC | Localizer |
| M | Measure word |
| MSP | Other particle |
| NN | Other noun |
| NN-SHORT | Other noun (abbrev.) |
| NR | Proper noun |
| NR-SHORT | Proper noun (abbrev.) |
| NT | Temporal noun |
| NT-SHORT | Temporal noun (abbrev.) |
| OD | Ordinal number |
| ON | Onomatopoeia |
| P | Preposition |
| PN | Pronoun |
| PU | Punctuation |
| SB | bei4 (in short bei-construction) |
| SP | Sentence-final Particle |
| URL | URL |
| VA | Predicative adjective |
| VC | Copula |
| VE | you3 (Main verb) |
| VV | Other verb |
| X | Others |

$3

| Tag | Original JA name | English |
| --- | ---------------- | ------------------ |
| A-c | 形容詞-一般 | Adjective-Common |
| A-dp | 形容詞-非自立可能 | Adjective-Dependent |
| C | 接続詞 | Conjunction |
| D | 代名詞 | Pronoun |
| E | 英単語 | English word |
| F | 副詞 | Adverb |
| I-c | 感動詞-一般 | Interjection-Common |
| J-c | 形状詞-一般 | Adjectival Noun-Common |
| J-tari | 形状詞-タリ | Adjectival Noun-Tari |
| J-xs | 形状詞-助動詞語幹 | Adjectival Noun-AuxVerb stem |
| M-aa | 補助記号-AA | Auxiliary sign-AA |
| M-c | 補助記号-一般 | Auxiliary sign-Common |
| M-cp | 補助記号-括弧閉 | Auxiliary sign-Open Parenthesis |
| M-op | 補助記号-括弧開 | Auxiliary sign-Close Parenthesis |
| M-p | 補助記号-句点 | Auxiliary sign-Period |
| N-n | 名詞-名詞的 | Noun-Noun |
| N-nc | 名詞-普通名詞 | Noun-Common Noun |
| N-pn | 名詞-固有名詞 | Noun-Proper Noun |
| N-xs | 名詞-助動詞語幹 | Noun-AuxVerb stem |
| O | その他 | Others |
| P | 助詞 | Particle |
| P-fj | 助詞-副助詞 | Particle-Adverbial |
| P-jj | 助詞-準体助詞 | Particle-Phrasal |
| P-k | 助詞-格助詞 | Particle-Case Marking |
| P-rj | 助詞-係助詞 | Particle-Binding |
| P-sj | 助詞-接続助詞 | Particle-Conjunctive |
| Q-a | 接尾辞-形容詞的 | Suffix-Adjective |
| Q-j | 接尾辞-形状詞的 | Suffix-Adjectival Noun |
| Q-n | 接尾辞-名詞的 | Suffix-Noun |
| Q-v | 接尾辞-動詞的 | Suffix-Verb |
| R | 連体詞 | Adnominal adjective |
| S-c | 記号-一般 | Sign-Common |
| S-l | 記号-文字 | Sign-Letter |
| U | URL | URL |
| V-c | 動詞-一般 | Verb-Common |
| V-dp | 動詞-非自立可能 | Verb-Dependent |
| W | 空白 | Whitespace |
| X | 助動詞 | AuxVerb |

Acknowledgements

The developers would like to thank Satoshi Sekine, Satoko Marumoto, Yoichi Yoshimoto, Keiji Shinzato, Keita Yaegashi, and Soh Masuko for
their contribution to this project.

References

Masato Hagiwara and Satoshi Sekine. Lightweight Client-Side Chinese/Japanese Morphological Analyzer Based on Online Learning. COLING 2014.

Kikuo Maekawa. 2008. Compilation of the Kotonoha-BCCWJ corpus (in Japanese). Nihongo no kenkyu (Studies in Japanese), 4(1):82–95.
(Some English information can be found here.) [Site]

Jialei Wang, Peilin Zhao, and Steven C. Hoi. 2012. Exact soft confidence-weighted learning. In Proc. of ICML 2012, pages 121–128. [PDF] [Site]

Naiwen Xue, Fei Xia, Fu-dong Chiou, and Marta Palmer. 2005. The Penn Chinese treebank: Phrase structure
annotation of a large corpus. Natural Language Engineering, 11(2):207–238. [PDF]

---

© 2014 Rakuten NLP Project. All Rights Reserved. / Sponsored by Rakuten, Inc. and Rakuten Institute of Technology.