http://www.eurecom.fr/en/publication/6231/download/data-publi-6231.pdf
—————————————————————————————————————————
contributions：
1，describe a compact graph-based representation that allows the specification of a rich set of relationships inherent in the relational world.
2，propose how to derive sentences from such a graph that effectively “describe" the similarity
across elements.
3，effective optimization to improve the quality of the learned embeddings and the performance of integration tasks.
4，propose a diverse collection of criteria.
steps：
1，定义3种node：row(r)，attributions(A)，token；边：把共现的连起来
2，random walk（随机游走）生成一些路径
3，From Walks to Sentences：
4，Embedding Construction：We piggyback on the plethora of effective embeddings algorithms such as word2vec, GloVe, fastText, and so on.
Optimization：IMPROVING LOCAL EMBEDDINGS
1，Handling Imbalanced Relations
2，Handling Missing and Noisy Data
Criteria：准备好所有node的embedding以后，就可以用于后续的任务
1，Schema Matching (SM)
- 直接拿CID的embedding做KNN查找。
2，Entity Resolution (ER)
- 直接拿RID的embedding做KNN查找。
3，Token Matching (TM).
- - 直接拿TID的embedding做KNN查找。Top1就是 conceptual synonym。
Experiment
1，Evaluating Embeddings Quality
- MatchAttribute (MA)，MatchRow(MR)，MatchConcept (MC)
2，Data Integration Tasks
- Schema Matching（SM）
- Entity Resolution
S/F/O：对muti-word的不同处理方式
DeepERpl：用local embedding+deeper网络
ER，不同TOPn，有不同效果。
- Token Matching
