Yes, ChatGPT is multilingual, but...

...It does not yield the same high performances that you can get by querying it in English.

For non-native speakers this can represent an initial barrier, for two reasons:🚧

1. Engineering effective prompts

When English is not your first language, generating on-point questions that fully express what you mean can be hard, and it is not unusual that ChatGPT or other language model get confused about what you are asking for, at least with their first answers.🤔

2. Unreliable results in your mother-tongue

On the other hand, when trying to speak with the LLM in your native language (especially if it is not well represented in the World-Wide-Web cultural products), you can bump into awkward phrasing, errors or difficulties in interpreting idioms and other everyday expressions.🤨

What can we do?

It would be great if we could generate a multilingual LLM from scratch, and Bigscience, for instance, is doing a lot in this direction with Bloom🌸.

Nevertheless, we can also decide to build upon already-existent English-based models, without finetuning or retraining them, but with a clever workaround: we can use a filtering function that is able to translate the user's native-language query in English, feeding it to the LLM and retrieving the response, which will be eventually back-translated from English to the original language.㊗️

Curious of trying? Let's use some python to build it!🐍

1. Import all the necessary dependencies

To build a multi-lingual chatbot, you'll need several dependencies, which you can install via pip:

python3 -m pip install transformers==4.39.3 \
langdetect==1.0.9 \
deep-translator==1.11.4 \
torch==2.1.2 \
gradio==4.28.3

Let's see what these packages do:

transformers is a package by Hugging Face, that helps you interact with models on HF Hub (GitHub)
langdetect is a package for automated language detection (Github)
deep-translator is a package to translate sentences, based on several translation services (GitHub)
torch is a package to manage tensors and dynamic neural networks in python (GitHub)
gradio is a package developed to ease the development of app interfaces in python and other languages (GitHub)

2. Build the back-end architecture

We need to build a back-end architecture that looks like this (realized with Drawio):

Let's define a Translation class that helps us with detecting the original language and translating it:

from langdetect import detect
from deep_translator import GoogleTranslator

class Translation:
    def __init__(self, text, destination):
        self.text = text
        self.destination = destination
        try:
            self.original = detect(self.text) # detect original
        except Exception as e:
            self.original = "auto" # if it does not work, default to "auto"
    def translatef(self):
        translator = GoogleTranslator(source=self.original, target=self.destination) # use Google Translate, one of the fastest translators available
        translation = translator.translate(self.text)
        return translation

As you can see, the class takes, as arguments, the text we want to translate (text) and the language we want to translate it into (destination).

Let's now load the LLM that we want to use for our purposes: we'll start with Bigscience's Bloom-1.7B, which is a medium-sized LLM and a good match for a 16GB RAM, 2-core CPU hardware.

from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

model = AutoModelForCausalLM.from_pretrained("bigscience/bloom-1b7") # import the model
tokenizer = AutoTokenizer.from_pretrained("bigscience/bloom-1b7") # load the tokenizer

pipe = pipeline("text-generation", model=model, tokenizer=tokenizer, max_new_tokens=2048, repetition_penalty=1.2, temperature=0.4) # prepare the inference pipeline

We define a maximum number of generated tokens (2048), set the repetition penalty to 1.2 (fairly high) in order to avoid the model repeating the same thing over and over again, and we keep the temperature ("creativity" in generating the response) quite low.

Now, let's create a function that is able to take a message from the chat, translate it to English (unless it is already in English), feed it as a prompt to Bloom, retrieve the English response and back-translate it into the original language:

def reply(message, history):
    txt = Translation(message, "en")
    if txt.original == "en":
        response = pipe(message)
        return response[0]["generated_text"]
    else:
        translation = txt.translatef()
        response = pipe(translation)
        t = Translation(response[0]["generated_text"], txt.original)
        res = t.translatef()
        return res

We have all we need for our back-end architecture, it is time to build the front-end interface!

3. Build the front-end user interface

With Gradio, building the user's interface is as simple as one line of code:

demo = gr.ChatInterface(fn=reply, title="Multilingual-Bloom Bot")

Now we can launch the application with:

demo.launch()

And, imagining that we saved the whole script in a file titled chat.py, to make the chatbot run we go to our terminal and type:

python3 chat.py

Then we patiently wait and head over to the local server link that Gradio will give us once everything is loaded and ready to work!

If you want to find the source code, head over to this GitHub repository.

Demo

Do you want to try what we just created? Make sure to head over to this Hugging Face Space I built: as-cle-bert/bloom-multilingual-chat💻.

Support

If you like the idea, make sure to show your support by leaving a little ⭐ on GitHub!

If you please, support my open-source work by funding me on GitHub: in this way, it will be possible for me to improve my multilingual chatbot performances by hosting it on a more powerful hardware on HF.

I will be really grateful for your concrete support🥺

AI enthusiasm #9 - A multilingual chatbot📣🈸