Five jumbled up sentences, related to a topic, are given below. Four of them can be put together to form a coherent paragraph. Identify the odd one out and key in the number of the sentence as your answer:

1. Machine learning models are prone to learning human-like biases from the training data that feeds these algorithms.

2. Hate speech detection is part of the on-going effort against oppressive and abusive language on social media.

3. The current automatic detection models miss out on something vital: context.

4. It uses complex algorithms to flag racist or violent speech faster and better than human beings alone.

5. For instance, algorithms struggle to determine if group identifiers like "gay" or "black" are used in offensive or prejudiced ways because they're trained on imbalanced datasets with unusually high rates of hate speech.

The correct answer is 3.

To identify the odd sentence out, let's analyze the theme and logical flow of the sentences to see which four can be combined into a coherent paragraph:

Sentence 2 introduces the broad topic: "Hate speech detection is part of the on-going effort against oppressive and abusive language on social media." This serves as a perfect introductory sentence for the paragraph.

Sentence 4 naturally follows Sentence 2 by explaining how hate speech detection works: "It uses complex algorithms to flag racist or violent speech faster and better than human beings alone." Here, the pronoun "It" refers back to "Hate speech detection" in Sentence 2.

Sentence 1 introduces a limitation or problem with these algorithmic models: "Machine learning models are prone to learning human-like biases from the training data that feeds these algorithms."

Sentence 5 provides a concrete example of the bias introduced in Sentence 1: "For instance, algorithms struggle to determine if group identifiers like 'gay' or 'black' are used in offensive or prejudiced ways because they're trained on imbalanced datasets with unusually high rates of hate speech." The transition phrase "For instance" directly connects the struggle with group identifiers to the biased training data mentioned in Sentence 1.

Together, sentences 2-4-1-5 form a logically coherent paragraph that moves from the definition of hate speech detection, to how it operates, to its limitations, and finally to a specific example of those limitations.

Sentence 3 ("The current automatic detection models miss out on something vital: context.") is the odd one out. While it also discusses automatic detection models, it introduces a completely new point (the lack of "context") that is not developed or connected to the discussion of training datasets, imbalanced data, or human-like biases discussed in sentences 1 and 5.