pat-tree

pat-tree

In Information Retrieval, text segmentation on Chinese like documents has been a difficult task, since Chinese words are continuous and has no white space between them. But finding basic elements of a document is critical for all applications in information retrieval.

PAT tree is a Patricia tree, or called trie, that used particularly for text segmentation and word retrieval. This module can be used for PAT tree construction for Chinese documents. Provide functionality to add documents and construct PAT tree in memory, convert to JSON for storing to database, extract keywords, and text segmentation.

You can collect a corpus, adding all of them to construct a PAT tree, then extract significant lexical patterns, and do text segmentation on other documents.

example of result:

有時 喜歡   有時候 不喜歡
為什麼 會 這樣   … ？
20 點 求 解 哈哈

Installation

npm install pat-tree --save

Usage

Instantiate

var PATtree = require("pat-tree");
var tree = new PATtree();

Add document

tree.addDocument(doc);

doc is the document you want to add to the tree. data type: string

Extract Significant Lexical Patterns

var SLPs = tree.extractSLP(TFThreshold, SEThreshold, verbose); 
// SLPs: array of JSON objects, which are signifiant lexical patterns and their relative informations.

If the frequency of a pattern exceeds TFThreshold, it would appear in the result array. The higher the SEThreshold, the stricter to filter out longest substrings of a significant lexical pattern.

verbose: optional, if set to true, then will print out progress on console.

TFTreshold should be integer, SEThreshold should be float between 0 and 1.

Text segmentation

var result = tree.segmentDoc(doc, SLPs); 

doc is the document to be segmented, data type: string.

SLPs is array of SLP that extracted by tree.extractSLP(), data type: array of JSON object.

result is the result of document segmentation, data type: string.

Convert to JSON

var json = tree.toJSON(); 

The result json has following three content:

• json.header: JSON object,
• json.documents: array,
• json.tree: array

You could store them to database and use tree.reborn() to generate the tree again. In NoSQL database, you can store the three items to seperate collections, header collection would contain exactly one document, and documents and tree would contain lots of documents.

For Example, if using MongoDB native driver:

	var json = tree.toJSON();

	// One header object would be stored to database
	db.collection("header").insert(json.header, function(err, result) {
		if(err) throw err;
	});

	// All documents would be stored to database
	db.collection("documents").insert(json.documents, function(err, result) {
		if(err) throw err;
	});	

	// All nodes of the tree would be stored to database
	db.collection("tree").insert(json.tree, function(err, result) {
		if(err) throw err;				
	});	

Reborn

tree.reborn(json);

If you use tree.toJSON() to generate the JSON object and store the three objects to different collections, you can construct them to the original JSON object and use tree.reborn(json) to reborn the tree.

For example, if using MongoDB native driver:

	db.collection("header").find().toArray(function(err, headers) {
		db.collection("documents").find().toArray(function(err, documents) {
			db.collection("tree").find().toArray(function(err, tree) {
				var json = {};
				json.header = headers[0];  // there should be only one header.
				json.documents = documents;
				json.tree = tree;

				var patTree = new PATTree();
				patTree.reborn(json);
			})
		})
	})	

The patTree object would now be the same as the previously stored status, and you can do all operations like patTree.addDocuments(doc) to it.

CATUION If you reborn the tree by above method, and do some operations like patTree.addDocument(doc), and you want to store the tree back to database as illustrated in Convert to JSON, you MUST drop the collections(header, documents, tree) in the database first, avoiding any record that is previously stored.

Additional functions

Print tree content

tree.printTreeContent(printExternalNode, printDocuments);

Print the content of the tree on console. If printExternalNode is set to true, print out one external node for each internal node. If printDocuments is set to true, print out the whole collection of the tree.

Traversal

tree.traverse(preCallback, inCallback, postCallback);

For convenience, there are functions for each order of traversal

tree.preOrderTraverse(callback);
tree.inOrderTraverse(callback);
tree.postOrderTraverse(callback);

For example

tree.preOrderTraverse(function(node) {
	console.log("node id: " + node.id);
})

Data type

Node

Every nodes has some common informaitons, an node has the following structure:

	node = {
		id: 3,        // the id of this node, data type: integer, auto generated.
		parent: parentNode,       // the parent of this node, data type: Node
		left: leftChildNode,      // data type: Node 
		right: rightChildNode,    // data type: Node
	}

Other attributes in nodes are different for internal nodes and external nodes, Internal nodes has following structure:

Internal nodes

	internalNode = {
		// ... 

		type: "internal", 
        // indicates this is an internal node
		position: 13,
        // the branch position of external nodes, data type: integer
		prefix: "00101", 
        // the sharing prefix of external nodes, data type: string of 0s and 1s
		externalNodeNum: 87, 
        // number of external nodes contained in subtree of this node, 
        // data type: integer
		totalFrequency: 89, 
        // number of the total frequency of the external nodes in the collection,
        // data type: integer
		sistringRepres: node 
        // one of the external node in the subree of this internal node,
        // data type: Node
	}

External nodes

External nodes has following structure:

	externalNode = {
		// ...

		type: "external", 
        // indicates this is an external node,
		sistring: "00101100110101", 
        // binary representation of the character, data type: string
		indexes: ["0.1,3", "1.2.5"] 
        // the positions where the sistring appears in the collection,
        // data type: array
	}

Collection

The whole collection consists of documents, which consists of sentenses, which consists of words. An example could be this:

	[ [ '嗨你好',
    	'這是測試文件' ],
  	  [ '你好',
    	'這是另外一個測試文件' ] ]

An index is in following structure:

DocumentPosition.SentensePosition.wordPosition

For example, "0.1.2" is the index of the character "測".

Performance

All operations are fast, but require more memory and disk space to operate successfully. Running on Macbook Pro Retina, connected to local MongoDB, given 8GB memory size by specifying V8 option --max_old_space_size=8000, has following performance.

• Add 32,769 Facebook-like posts by tree.addDocument() takes about 5 minutes.
• After above operation, extract SLP by tree.extractSLP() takes about 5 minutes.
• After above operation, converting to JSON by tree.toJSON() and store three collections to database takes about 1 minutes and 5 GB disk space, and about 1,000,000 records of tree nodes.
• After above operation, find all collections in database and reborn the tree by tree.reborn() takes about 1 minutes.
• After above operation, do text segmentation on 32,769 posts by tree.segmentDoc(), given SLPs extracted above, takes about 5 minutes.

Release History

• 1.0.2 Gaurantee SLP sorting order when segmentDoc()
• 1.0.1 Modify README file
• 1.0.0 Stable release
• 0.2.8 Improve algorithm of segmentDoc()
• 0.2.7 Fix bug in reborn()
• 0.2.6 Greatly improve performance of extractSLP()
• 0.2.5 Greatly improve performance of addDocument()
• 0.2.4 Fix bug in reborn()
• 0.2.3 Add functions toJSON() and reborn()
• 0.2.2 Change function name of splitDoc() to segmentDoc()
• 0.2.1 Mofify README file
• 0.2.0 Add text segmentation functionality
• 0.1.8 Alter algorithm, improve simplicity
• 0.1.7 Improve performance
• 0.1.6 Improve performance
• 0.1.5 Add functionality of SLP extraction
• 0.1.4 Add external node number and term frequency to internal nodes
• 0.1.3 Able to restore Chinese characters
• 0.1.2 Construction complete
• 0.1.1 First release