Different Language Support about txtai

[davidmezzetti]

Thank you for the support.

The best place to start is this notebook: https://colab.research.google.com/github/neuml/txtai/blob/master/examples/01_Introducing_txtai.ipynb

The index in the notebook above uses sentence-transformers. This link has a list of all the sentence transformer models available: https://huggingface.co/models?search=sentence-transformers

The following is an example modification using a multi-lingual model

# Create embeddings model, backed by sentence-transformers & transformers
embeddings = Embeddings({"method": "transformers", "path": "sentence-transformers/xlm-r-100langs-bert-base-nli-stsb-mean-tokens"})

If this doesn't work well, another model to try: sentence-transformers/LaBSE

Then change the sections text below to a couple examples in the target language you want to experiment with.

sections = ["US tops 5 million confirmed virus cases",
            "Canada's last fully intact ice shelf has suddenly collapsed, forming a Manhattan-sized iceberg",
            "Beijing mobilises invasion craft along coast as Taiwan tensions escalate",
            "The National Park Service warns against sacrificing slower friends in a bear attack",
            "Maine man wins $1M from $25 lottery ticket",
            "Make huge profits without work, earn up to $100,000 a day"]

Finally change the queries to the target language as well:

for query in ("feel good story", "climate change", "health", "war", "wildlife", "asia", "north america", "dishonest junk"):

I would just iterate over a couple different models until you find one that works well. Let me know how it works out.

[ByUnal]

I appreciate for your feedback and concern. I will try your recommendation as soon as I'm available, then I will let you know.

[ByUnal]

Hello again, now I tried to transformers that you suggested. I tried both of them for Turkish. First one worked in some cases but it is not efficient. Second one didn't work. I need Turkish language support of transformers

[davidmezzetti]

Another one to try: sentence-transformers/distilbert-multilingual-nli-stsb-quora-ranking

Otherwise you can also try any of the generic Turkish transformer models: https://huggingface.co/models?search=turkish

Those multilingual models are the ones that should support multiple languages. I suspect a model specifically trained for Turkish language and on a NLI/STSB like task for Turkish would work best.

txtai uses the sentence-transformers library to build transformer-based sentence embeddings. I would suggest trying as many of the models to see if any of them work at an acceptable level for your task.

[davidmezzetti]

Issue should now be resolved. Tokenization can be disabled by setting the config option:

Embeddings({"method": "transformers", path: "/path/to/model", "tokenize": False})