Week Eight and Nine | Stage 4 | Hard Weeks, Slow performance gains
TL;DR
Pretty sloppy weeks! Here are the tasks I was supposed to accomplish this week:
- A new test was suggested by one of the Mentors which ensured that some composition template were exclusively present in the test:
- train will contain A,B
- train will contain AoC
- train will contain DoB
- test will contain AoB
- Here A and B represent ontology entities (Done)
- Improve the performance on this test (In progress)
- Report on experiments done. (Done)
- Add TLDR(To Long Didn’t Read) section to the blogs. (In progress)
- Improve and adapt whenever required. (In progress)
- Complete Blogs (Not Done)
Suggestions after meeting
- Introduce proper alignment information to the model. This might fix the
<unk>problem. As used in this paper: https://arxiv.org/pdf/1806.10478.pdf - Example:
- it means to literally append lines to the
train.{en,sparql}files with labels and resources, respectively. the result was faster convergence, so reaching a certain BLEU score in less time. an example could be:
- it means to literally append lines to the
train.en
--------
what is the capital of Germany?
...
germany
united states
japan
train.sparql
------------
select var_s where br_open dbr_Germany dbo_capital var_s br_close
...
dbr_Germany
dbr_United_States
dbr_Japan
- Try embeddings
- One is to use a word embedding model which supports OOV words, such as fastText
- Another is to train your own corpus also including Wikipedia
- Add more examples to allow the model to learn so that the prediction increases.
This Week Work
- Try more experiments, understand, evolve and proceed.