LP Extractor Protocol

From edegan.com
Revision as of 21:43, 26 March 2019 by Ed (talk | contribs)
Jump to navigation Jump to search


Project
LP Extractor Protocol
Project logo 02.png
Project Information
Has title LP Extractor Protocol
Has owner Lasya Rajan
Has start date
Has deadline date
Has project status Active
Subsumed by: Listing Page Extractor
Copyright © 2019 edegan.com. All Rights Reserved.

Overview of Possible Methods

According to “Project Goal V2,” (E:\mcnair\Projects\Incubators) there are three proposed methods to organize and extract useful information from an HTML web page. The first method is text processing, analyzing and classifying the textual content of the HTML page. The second method is to use image based pattern recognition, likely through an off-the-shelf model that can extrapolate key HTML elements from web page screenshots. The third, and most novel method is to structurally analyze the HTML tree structure, and express that simplified HTML structure in a Domain Specific Language (DSL).

Text Processing

There are two possible classification methods for the processing the text of target HTML pages. The first is a "Bag of Words" approach, which uses Term Frequency – Inverse Document Frequency to do basic natural language processing and select words or phrases which have discriminant capabilities. The second is a Word2Vec approach which uses shallow 2 layer neural networks to reduce descriptions to a vector with high discriminant potential. (See "Memo for Evan" in E:\mcnair\Projects\Incubators for further detail.)

Image Processing

This method would likely rely on a conventional neural network (CNN) to classify HTML elements present in web page screenshots. Implementation could be achieved by combining the VGG16 model or ResNet architecture with batch normalization to increase accuracy in this context.

HTML Tree Structure Analysis

Structurally analyzing the HTML tree structure of a web page and expressing it in a DSL is the most innovative method of the three. It would require more than simply adapting off-the-shelf models. First, the DSL itself would need to be designed to optimize abstraction into the target domain, a web page. (See Domain Specific Language Research.) Then, the DSL would need to be integrated into the machine learning pipeline by encoding the DSL into an appropriately formatted input, such as a vector or matrix, for a neural network. Three proposed methods for this encoding are using an adjacency matrix, an edges to vertices matrix, or utilizing DFS (depth-first search) algorithms.

DFS Encoding

Currently, we are leaning towards utilizing DFS algorithms. DFS algorithms operate by starting at the root node (or an arbitrary node for a graph) and traverses the longest branch fully before backtracking back to the last split before the branch terminated. A DFS algorithm could traverse any given tree and record 1 when a new node is found, and 0 when that node is fully explored. This creates a numerical representation of that tree that can then be entered into a vector or matrix. A DFS algorithm has an efficiency of O(n).

Adjacency Matrix

By interpreting the tree as a graph, we can utilize an adjacency matrix to encode the tree. The elements of the matrix represent whether their corresponding vertices are adjacent in the graphical representation. In its simplest form, for a set of V number of vertices, the matrix would be a square matrix of dimensions |V| x |V|. The diagonal elements of such a matrix would all be zero. This approach has an algorithmic efficiency of O(n^2).

Edges to Vertices Matrix

For any given tree, we have n-1 (I'm assuming n = number of nodes) edges. For every edge, we can record the two ending vertices. This will result in a matrix of dimensions (n-1) x 2. This matrix approach has an algorithmic efficiency of O(n).

node2vec

Supervised Learning Approach (HTML to DSL)

Additionally, the HTML tree structure analysis method will require a subprocess by which to parse a complex HTML page into our DSL. An example of a similar process is Pix2Code, in which a DSL context and a GUI are feed into an architecture containing Long Short-Term Memory (LSTM) layers and a CNN-based vision model (see image below) which outputs a DSL token. After training with paired inputs is complete, this architecture can then take an empty context and a GUI input and output DSL code.

Image from "Project Goal V2" of Pix2Code architecture

Literature

HTML Tree Structure Analysis

This is the documentation for the Pix2Code architecture mentioned.
This article provides an overview of various web data extraction techniques. Section 2.2 describes a process of extracting a web page's DOM structure, and Section 2.3 includes a supervised extraction process that has some similar aspects to Pix2Code and other paired input architectures.
This approach to web content extraction focuses exclusively on less structured web pages, and classifying text blocks within those pages. In Section 3: Data Collection, an algorithm written in JavaScript is used to inspect DOM elements and organize them by parent element.
This survey of various web data extraction methods includes a section on tree-based analysis in Chapter 2: Techniques.
This paper presents a method of content extraction that analyzes the relationships between DOM elements based on a "chars-node ratio" that displays the relationship between text content and tags content in each node of the DOM tree. The authors of this paper implemented this technique in an open-source Firefox plugin.


DFS Encoding

This paper discusses different methods of encoding graph structures into low-dimensional embeddings that can be exploited by machine learning models. Section 2.2.2 (Random walk approaches) specifically compares the accuracy of using the random walk approach to traverse a graph, as opposed to BFS and DFS approaches.
node2vec was mentioned briefly in the above Hamilton et al. paper. node2vec is a scalable encoding algorithm that focuses on preserving network neighborhoods of nodes. The definition of a neighborhood can manipulated depending on the application context. In section 3.1 (Classic Search Strategies), node2vec is compared to BFS and DFS approaches.
V2V (Vector to Vertex) is a learning approach similar to node2vec, except that it takes the random-walk sequence results from graph data and encodes them using a Continous Bag of Word (CBOW) approach to create V2V vectors.


General

This article describes methods to simplify the content of a noisy HTML page, specifically through using machine learning to predict whether a block is content or non-content. This allows the classifier to remove boilerplate information.


This article classifies various web data extraction techniques into 5 different types of tools, and one category of web extraction specific-languages. Section 3.2 (HTML-aware Tools) describes several existing tools for parsing HTML tree structures in building wrappers. Section 3.4 (NLP-based Tools) includes several methods of text analysis that may be relevant to this project.