[paper reading] raw+:a two view graph propagation method with word coupling for readability assessment

Key Point:

• Graph based classification （ Use the relationship between documents ）
• Coupled word bag model （ Evaluate the relevance of words in reading difficulties ）
• Two view propagation method （ At the same time, we use word bag model and language features ）

Method

in general , It is divided into the following two stages ：

• Characteristic means （ Map the document into feature vectors ）
• Readability classification （ Graph construction 、 Graph merging and label propagation ）

One . Coupled word bag model （The Coupled Bag-of-Words Model）

1. Construct word coupling matrix （the word coupling matrix）

• hypothesis ： Simple words tend to appear in simple sentences , Difficult words tend to appear in difficult sentences .
• idea ： Assess the difficulty of reading a word , Through its co-occurrence probability distribution in sentences with different difficulty levels .
• Note： Use unlabeled sentences , Use heuristic functions （heuristic functions） Label the sentences with reading difficulty .

Step1：Per-sentence reading difficulty estimation

• Use heuristic function to make a rough estimation , Get the weak tag of each sentence （weak label）.
• utilize 8 Language features （ It can be applied to sentence level ）, To build the 8 A heuristic function is used to calculate the difficulty score of sentences .
  
• Discretize continuous fractions , Use the following formula to get the reading level of each sentence （3 In terms of ）
  

Step2：Per-word difficulty distribution estimation

• Every word $t$ The corresponding length is i Vector （$i$ Is the number of reading grades ）.
• Every dimension of a vector $i$ Is a probability value $p_t(i)$, Equals that the word appears at the corresponding level $i$ The proportion of the number of sentences in the total number of sentences in which the word appears , The calculation formula is as follows ：
  

Step3：Word coupling matrix construction

• The value of each element in the word coupling matrix corresponds to the relationship between two words , The value is the similarity of the vector between two words / The difference of probability distribution .
• A total of 3 Word coupling matrix $C^{sur},C^{lex},C^{syn}$.
• Filtering strategy （ Avoid that it will be very time-consuming when the vocabulary is too large ）： According to the entropy of words , Set the percentage , Filter low entropy words .

2. Generate a coupled word bag model

• Through the word coupling matrix and the basic word bag matrix （BoW matrix） Multiply , Get the coupled word bag matrix （coupled BoW matrix).
  

• A total of 3 Coupling TF-IDF matrix , Namely $M^{sur}$,$M^{lex}$,$M^{syn}$, Coupled word bag matrix （cBoW） Will be dense , And pay attention to the similarity of reading difficulties .

Two . Language features （The Linguistic Features）

• Build a matrix $M^l\in R^{n_l\times |D|}$, among $n_l$ Is the number of language features .
• The language features selected in this article are language-independent, To support the different language independent features of their proposed methods .
• Surface Features, Lexical Features, Syntactic Features

3、 ... and . Dual view propagation （Two-View Graph Propagation）

• The construction of graph
  

• Merging of graphs
  It is divided into the merging of isomorphic diagrams in views and the merging of heterogeneous diagrams between views .

• Intra-view homogeneous graph merging
  The basic idea is to keep the common edges and remove the edges with redundant information . First keep in 3 Neighbor nodes exist in all graphs , For nodes that exist in at least one graph , Select and point v Nodes with the least public neighbors （ The purpose is to ensure that the number of triangles in the graph is at least ）. The boundary right is 3 The average of the corresponding edges of a graph .