Neural sentiment backend backed by Bumblebee.

Loads a pre-trained multilingual sentiment classifier from Hugging Face and serves predictions via Nx.Serving. The default model is cardiffnlp/twitter-xlm-roberta-base-sentiment, an XLM-RoBERTa model fine-tuned on multilingual Twitter data (English, French, German, Italian, Spanish, Portuguese, Arabic, and Hindi by training, with reasonable transfer to many more languages via the underlying multilingual base).

This backend is opt-in:

• The :bumblebee and :exla Hex packages must be added as dependencies of the host application.

• Either pass backend: Text.Sentiment.Backends.Bumblebee to Text.Sentiment.analyze/2, or set the application config:

  config :text, :sentiment_backend, Text.Sentiment.Backends.Bumblebee

Cold start

The first call to analyze/2 downloads the model weights (~280 MB on disk) from Hugging Face, traces the inference graph through Bumblebee, and compiles it under EXLA (or Nx.Defn.Evaluator if EXLA is not loaded). Expect a 10–30 second cold start. Subsequent calls in the same VM run in single-digit milliseconds — the compiled serving is cached in :persistent_term.

For production deployments where cold start is unacceptable, start a named serving process at boot via Bumblebee.Text.text_classification/3 + Nx.Serving.start_link/1, then pass serving: <name_or_pid> to analyze/2 to skip the cache entirely.

Tokenizer override

Some fine-tuned models on Hugging Face ship without the tokenizer.json Bumblebee expects — they have only the raw SentencePiece or WordPiece data. The default Cardiff sentiment model is one of those, so this backend loads its tokenizer from the base FacebookAI/xlm-roberta-base repo instead. Other models fall through to "use the same repo as the model" by default.

If you point :model at a fine-tune that itself lacks a tokenizer.json, pass :tokenizer_repo to point at a repo that has one (typically the base model the fine-tune was trained on).

Result shape

Returns the same map shape every backend produces:

• :label — :positive, :negative, or :neutral (mapped from the model's textual labels).

• :compound — P(positive) − P(negative), in [-1.0, +1.0].

• :scores — the full per-label probability map, e.g. %{positive: 0.86, neutral: 0.10, negative: 0.04}.

• :backend — Text.Sentiment.Backends.Bumblebee.

• :model — the model id used for the prediction.

Unlike the lexicon backend, no :matched/:tokens counts are included — the underlying model is opaque about which tokens it used.

Language

XLM-RoBERTa is intrinsically multilingual; no per-language routing is required. The :language option is accepted (for API consistency with the lexicon backend) but ignored by this backend; the language used at scoring time is reported in the result map for round-tripping.