print out cluster of relations for each relation prediction with the associated entities

[bhomass]

Hi, I am still reading thru the code, not sure if I am close. It would really help if you could give some code example, how the trained model would predict the relation per input sentence features, and at the end provide sample data (entities with their predicted relation, along with original text) for each relation cluster.

It would seem after loading the saved model, you still need to call OieInduction.compileFunction to create the theano graph, apply the parameters from the saved model, then call functions like OieInduction.getClustersSets() which depends on the output from OieInduction.compileFunction().

Is my assessment correct?

I know you can turn on the pickle output during learn(), but it seems you would be dumping out one version per iteration, when I only want the final version. In addition, I don't really want to call learn repeatedly after initial training, just want to do inference on new data using the trained model.

[bhomass]

I wrote a predict() method in OieInduction, went thru the compilation of the predict and labelTrain graphs. Ran thru the the dataset and indeed got good predict output. Will check into them in detail later.

[bhomass]

after loading the pickled model, and re-apply to the same training dataset, the prediction shows 970 entries in relation 4 and 30 entries in relation 7. Is this a matter of poor training?

when I increase the epoch count to 100, now all relations get stuck on relation 4.

apparently, it is necessary to boost the alpha parameter to spread out the relation assignment a bit more.

[chenhan97]

after loading the pickled model, and re-apply to the same training dataset, the prediction shows 970 entries in relation 4 and 30 entries in relation 7. Is this a matter of poor training?

when I increase the epoch count to 100, now all relations get stuck on relation 4.

apparently, it is necessary to boost the alpha parameter to spread out the relation assignment a bit more.

Hi bhomass, can you share your predict() method code? I'm also trying to visualize the prediction...but not familiar to Theano. Many thanks ahead!

[henuykc]

Hello, bhomass. Can you use the NYT data set to run this code? If so, could you please share how to make the NYT data set that can make this code run, or how to make a good data set. Thank you very much!

[bhomass]

gentleman, this is an old paper, hardly the state of the art. I suggest you move on to newer publications. I myself can no longer find the code.

[henuykc]

hahaha, I'm working on my graduation thesis, so I don't think too much about the thesis time. Thank you for your suggestion!