Gali Ün

ai engineer · istanbul, open to relocation or remote

← all work

Seq2Seq Language Model

project · 2018

Python · TensorFlow · Teknofest 2018

004 · find the answer in the text

q Which element did Marie Curie name after her homeland?

In 1898 Marie Curie named a newly discovered radioactive element polonium, after her native Poland.

→

extractive qa · seq2seq with attention · the 2018 model read turkish

A neural language model with attention that I built with TensorFlow in 2018: an encoder reads a sentence, and a decoder writes the reply word by word, attending to the parts of the input that matter. Attention was still new then. The Transformer paper had come out in 2017, and BERT followed in October 2018.

One config file sets it up: LSTM or GRU cells, a unidirectional or bidirectional encoder, Luong or Bahdanau attention, bucketing for sequences of different lengths, and word2vec or fastText embeddings, pre-trained or trained on your own data.

My Teknofest 2018 competition project grew out of it: I adapted it into a Turkish question-answering system for Teknofest's nationwide AI contest, where it placed 3rd.