Alt text validation system

A method for validating alt text for images in web pages, the method comprising:

• extracting image data from the web pages, the image data including source data and alt text data for a plurality of image elements in the web pages;
• providing images defined by the source data as input to a computer vision component;
• receiving a natural language description of each of the images from the computer vision component;
• generating prompts for the image elements, each of the prompts being associated with an image and including at least the natural language description for the image and the alt text data for the image;
• providing each of the prompts as inputs to a semantic comparison model, the semantic comparison model being trained to output similarity scores for the prompts, each of the similarity scores being indicative of a similarity between the natural language description and the alt text data for an image element associated with a prompt;
• determining which image elements in the plurality of image elements have inaccurate alt text based on the similarity score;
• and
• flagging image elements with inaccurate alt text as requiring user action to correct the alt text.

A non-transitory computer readable medium on which are stored instructions that, when executed, cause a programmable device to perform functions of:

• extracting image data from the web pages, the image data including source data and alt text data for a plurality of image elements in the web pages;
• providing images defined by the source data as input to a computer vision component;
• receiving a natural language description of each of the images from the computer vision component;
• generating prompts for the image elements, each of the prompts being associated with an image and including at least the natural language description for the image and the alt text data for the image;
• providing each of the prompts as inputs to a semantic comparison model, the semantic comparison model being trained to output similarity scores for the prompts, each of the similarity scores being indicative of a similarity between the natural language description and the alt text data for an image element associated with a prompt;
• determining which image elements in the plurality of image elements require new alt text data based on the similarity score;
• and
• flagging image elements with inaccurate alt text as requiring user action to correct the alt text.