Inconsistent results with current deepdive version.

[ghost]

Hi, I am currently running the last version of deepdive on and EC2 instance (r3.xLarge) with ubuntu 16.04. I am using the spouse example with a 5000 lines subset of the signalmedia I created. signalmedia-5k.jsonl.gz I noticed I don't have the same results number of lines in the database each time I run the the process. Here are my results : Run Articles Sentences person_mention spouse_feature 1 5000 92074 39455 53618035 2 5000 92070 39452 53635745 3 5000 92069 39470 53636191 4 5000 92071 39457 53637198

I am about to write a python script that will compare the different databases to know what the differences are.

[ghost]

I looked at the differences in the tables. I noticed that even when I have the same number of rows between two sentences tables they don't contain the same results. I looked quickly at the differences and nothing obvious pops up. In one case it considers \u2013 as a group of four numbers instead of a single character. In another number nothing is clear in what the difference is.

[ghost]

Ok. I did multiple runs on my 5k spouse test to get different sentences results. I also created a small python program to compare the results. From what I see when I have differences it is always the last lines. It seems there is a reuse of a string variable somewhere and when there is a document that is smaller than the previous document then deepdive or coreNLP mixes both documents by adding the end of the big document to the small document.

[ghost]

I pushed further my tests. I think there is an issue with the CoreNLP. I have runs where in the same phrase of the same document words are cut in half such as "available" in one run and "availa", "ble" in the other run.

[manning]

Did you ever solve this? Can you provide a document where this can be directly reproduced just calling CoreNLP?

[ghost]

Hi Manning, we didn't solve this. We haven't tried to reproduce it by calling CoreNLP directly but we will probably do it in the future since we are using it.