TextRank, 关键词和句子抽取

1. 简介

TextRank, 基于图模型的关键词和句子抽取. 与 google 的PageRank有一定的相通之处. 都是 unsupervised.

2. theory

术语约定:

• vote
  G=(V,E). 点 A links to B , 可以认为是 A 对 B 发起的一次vote, 即 A casts a vote for B.

TextRank model 的基本思想是:

1. 点A得到的投票数越多, A越重要.
2. 点A的重要性直接影响到它所投票的重要性.

2.1 formula

$$S(V_i)=(1-d)+d*\sum_{j \in In(V_i)} \frac 1 {|Out(V_j)|} S(V_j) \tag 1$$

• $S(V_i)$
  score of $V_i$
• d
  damping factor that can be set [0,1], usually set to 0.85.
• $In(V_i)$
  the set of vertices that point to $V_i$
• $Out(V_i)$
  the set of vertices that $V_i$ point to .

3.build graph

基于图的排序模型, 那么关键就是建图.

3.1 edge

two vertices are connected if they co-occur within a window of maximum N words, where N can be set anywhere between [2,10].

• directed edge
  若要有序, the direction was set following the natural flow of the text.
• undirected edge
  无向边可以认为是双向的弧.

3.2 vertex

句子由若干个word组成 , 需要对这些 word 作简单的处理, 才能当作 vertices , added to the graph.

• pre-processing
  根据stop word, 词性标注过滤一部分, 剩下的当作vertex.
  论文实验显示, 仅考虑 noun 和 adjective 是好的.
• processing
  根据文本构建 graph 的过程.
• post-processing
  相邻接的 key word sequence 会 collapsed into multi-word keyword.
  论文中是这么举例的. 在文章 Matlab code for plotting ambiguity functions 中, 如果 Matlab 与 code 总是邻接, 那么就可以考虑合成为 Matlab code, 丰富意思的表达.

4. discussion

论文中试验用的 data set is a collection of 500 abstracts from the Inspec database.
结果: precision:31.2, Recall:43.1, F-measure:36.2 .

• window size
  论文中用了(2,3,510), 取2时效果最好.
• directed or undirected edge
  论文中用 无向边 的时候, 效果最好.

5. sentence extraction

可以用于 自动生成摘要.

• 顶点
  不再是单词, 而是整个句子.
• 边
  不再以 co-occurrence 作为连接的条件, 因为 “co-occurrence” is not a meaningful relation for such large context.
  而是依据 similarity.
  记 句子 $S=w_1, w_2, ... , w_n$, 表示为单词的序列, 那么句子之间的相似度为:
  $$Similarity(S_i, S_j)=\frac {|\{ w_k | w_k \in S_i \cap w_k \in S_j \}|} {\log (|S_i|) + \log (|S_j|) } \tag {5-1}$$
  把相似度作为权重, 计算顶点的重要性:
  $$WS(V_i)=(1-d)+d* \sum_{ j \in In(V_i)} \frac { w_{j,i}} {\sum_{k \in Out(V_j)} w_{j,k} } WS(V_j) \tag {5-2}$$

参考

1. 原始论文, TextRank: Bringing Order into Texts