Dataset format and input format for new predictions about cnn-re-tf HOT 4 CLOSED

Hi,

Could you please refer issue#1 (how to check prediction) and issue#2 (how to create source.att), at first?

MLMI stands for multi-label multi-instance training, that is, the labels are not mutually exclusive.
My task is to predict the relationship between two entities. For example, if you have the entity pair:

Berlin <relation> Germany

then the correct prediction will be: country, located_in and capital_of. (In this case, three correct labels according to WikiData)

Let's say, I found the following sentence online:

The city of Berlin is located in eastern Germany on the River Spree. 

Then the training data looks like:

• source.txt (Berlin has entity ID: <Q64>, Germany: <Q183>)

  The city of <Q64> is located in eastern <Q183> on the River Spree . 
  

• source.left

  The city of <Q64>
  

• source.middle

  is located in eastern
  

• source.right

  <Q183> on the River Spree . 
  

• target.txt (assume that the first column represents whether the sentence has the label country, second column: located_in, third column: capital_of, etc.)

  1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
  

Here is the list of 23 labels (and the class-wise results):

If you still have any questions, describe them little bit more concretely, please? I don't have any clue what your task is, how your data look like etc...

from cnn-re-tf.

Hi, thanks for the explanation. My task is also relationship extraction, I need to find the relations that exist in a given set of documents. I plan to first create a dataset from the documents with the entities and relations annotated, and then train a model on it to predict relations for new documents. Will this project be suitable for my task?

from cnn-re-tf.

It sounds pretty similar to my task.
Ah, so, your question is how to do distant supervision, right?

I'm not sure if it helps, but I describe here the four steps I did in the distant_supervision.py script.

1. Download raw text data (if you already have the documents, you can replace this with your data)
   For example:

   The city of Berlin is located in eastern Germany on the River Spree. 
   

2. Find entity
   • execute StanfordNER (demo)

     The city of <Berlin_LOCATION> is located in eastern <Germany_LOCATION> on the <River Spree_LOCATION>. 
     

   • extract all combinations. In the case above:

     <Berlin_LOCATION> <Germany_LOCATION>
     <Berlin_LOCATION> <River Spree_LOCATION>
     <Germany_LOCATION> <River Spree_LOCATION>
     

3. Find entity ID (using SPARQL on WikiData API (sample))

   <Q64 Berlin> <Q183 Germany>
   <Q64 Berlin> <Q1684 River Spree>
   <Q183 Germany> <Q1684 River Spree>
   

4. Find relation (using SPARQL on WikiData API (sample))

   <Q64 Berlin> <Q183 Germany> --> <P17 country>, <P1376 capital_of>, <P131 located_in>
   <Q64 Berlin> <Q1684 River Spree> --> <P206 located_next_to>
   <Q183 Germany> <Q1684 River Spree> --> None
   

At the end of these four steps, we get

• source.txt

The city of <Q64> is located in eastern <Q183> on the River Spree . 
The city of <Q64> is located in eastern Germany on the <Q1684> . 

and

• target.txt

1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

If you want to learn the theory, I recommend to read:

• Minz et al. Distant supervision for relation extraction without labeled data ACL 2009.
• Dong et al. Knowledge Vault: A Web-Scale Approach to
  Probabilistic Knowledge Fusion ACM SIGKDD 2014.

from cnn-re-tf.

Thanks for the detailed explanation, I think I might be able to solve my task with this.

from cnn-re-tf.