docs: document speciesist bias in word vectors and NLP models

[stuckvgn]

Proposal

Document speciesist bias as a recognized form of bias in spaCy's documentation, particularly in sections about word vectors and model evaluation.

Context

Peer-reviewed research has demonstrated that word embeddings and NLP models — the foundation of spaCy's pipeline — encode speciesist associations:

• Takeshita et al. (2022). "Speciesist language and nonhuman animal bias in English masked language models." Information Processing & Management. Found BERT, DistilBERT, RoBERTa, and ALBERT all associate harmful words with nonhuman animals.
• Hagendorff et al. (2023). "Speciesist bias in AI." AI and Ethics. Documented speciesist bias across image recognition, word embeddings, and language models.
• Leach et al. (2023). Large-scale word embedding analysis establishing anthropocentric speciesism in everyday language. British Journal of Social Psychology.

Relevance to spaCy

• Word vectors (en_core_web_md, en_core_web_lg): Pre-trained vectors inherit speciesist associations from training corpora.
• NER: Named entity models trained on biased corpora may treat animal-related entities differently
• Text classification: Downstream models built on spaCy inherit these biases
• Sentiment: spaCy-based sentiment pipelines reflect speciesist patterns in training data

Suggested Change

A brief note in the word vectors or model evaluation documentation acknowledging speciesist bias as a documented phenomenon, with citations. This aligns with spaCy's existing documentation on model limitations and biases.