We are a team of graduate students working voluntarily at Inzva Hacker Space on building a model for Generic Sentence Encoder for the Turkish Language. Generic Sentence Encoders are used to let the computer understand what the user meant by their sentences. They are typically used in conversational assistants to understand the meaning of the user utterances. There are a number of recent research that are proposed for Generic Sentence Encoder for the English language as exemplified below.

Universal Sentence Encoder (by Google AI) https://ai.google/research/pubs/pub46808

Learning General Purpose Distributed Sentence Representations via Large Scale Multi-task Learning (by Microsoft Research) https://openreview.net/forum?id=B18WgG-CZ

They leverage the Natural Language Inference (NLI) datasets to train their models as well as to evaluate their models as a downstream task. There are many NLI datasets that have been previously crowdsourced for the English language. Examples in an NLI dataset can be seen below.

An example of entailment (similarity).

• Sentence 1: A senior is waiting at the window of a restaurant that serves sandwiches.
• Sentence 2: A person waits to be served his food.

An example of contradiction (dissimilarity).

• Sentence 1: A senior is waiting at the window of a restaurant that serves sandwiches.
• Sentence 2: A man is waiting in line for the bus.

Examples are taken from The Stanford Natural Language Inference (SNLI) Corpus (https://nlp.stanford.edu/projects/snli).

We want to build a Generic Sentence Encoder model for the Turkish. However, there are not any NLI datasets in the Turkish language due to the high cost of crowdsourcing. We want to leverage the new Amazon Translate Service to translate the NLI datasets available in the English language into the Turkish language as a cost effective solution and publish the translated datasets publicly to boost Turkish NLP works.

The results of the following experiments were obtained.

Note: We are working on the presentation of the results in an more readable, understandable and interpretable way. Stay tuned :)

🎯 (2019-09-30) We have applied for the AWS Research Credits again to translate some well known NLI and Textual Entailment datasets into Turkish.

🐣 (2018-10-26) Our research has been awarded free GPU resource by TRUBA which is the Turkish Science e-Infrastructure that is provided by The Scientific and Technological Research Council of Turkey (TUBITAK).

🐣 (2019-01-26) Our research has been granted AWS Research Credits to translate the most commong NLI datasets into Turkish and evaluate our resulting NLI models.

🎯 (2019-03-31) We have applied for the AWS Research Credits again to translate all available NLI and Textual Entailment datasets into Turkish.

🐣 (2019-10-18) Our research has been granted AWS Research Credits to keep translating the NLI datasets into Turkish and evaluate our resulting NLI models.

This repository is based on the implementation of the following paper in this Github repository which belongs to the following paper.

Lajanugen Logeswaran, Honglak Lee, An efficient framework for learning sentence representations. In ICLR, 2018.