# 38th place solution (NER + Heuristics)

I am afraid my solution is way too simple compared to some of the cool models I've seen. But here goes:
I used Spacy's entity recogniser to filter all possible candidate datasets. Then I wrote a simple CFG which determines if those entities indeed denote a dataset. This part generates candidate datasets that are strict matches, thereby increasing the rate of false negatives. To mitigate this, I did a second pass on all texts with key terms in these candidate datasets to generate the final discovered dataset. That's the model in a nutshell.
Few additional notes:
This model is nearly unsupervised.
I did not use the train dataset for the actual dataset prediction. I only used it to determine what kind of CFG I need to write.
One very direct observation is, all datasets that are referred in these publications are camel-cased. If nothing else, you can start with these as your candidates and even exclude the named-entity task. (That gave me a private score of 0.29. The final model I submitted with Spacy gave 036).
These are not free-form texts on which standard large-scale models have been trained. These are publications and they follow a certain standard. If you analyze the training dataset, you can automatically see that most of them are of the form
data collected from <some dataset>
,
Used <some report/study>
, and so on. The second pass in my model exists to capture those that don't follow these candidate phrasings.
I did not use Roberta for the same reason. See the first place solution post for a much more in-depth discussion on this.
As for the heuristics, you can just start with "data" and look for subtree (in the PCFG) matches that intersect with the matched entities.  I used a few more. Again, the input texts are almost free of grammatical errors and follow very similar patterns while quoting references and external datasets.
Spacy 3.0's Transformer gave me much better initial candidates. But it kept throwing some submission errors.
So I used Spacy 2.5 since I didn't have much time to compete in this competition.