Explainability and Transparency

Both in trust & safety and law enforcement, ability to trace the system's reasoning when making a decision is paramount. Today, algorithmic transparency is also required by EU Digital Services Act.

Providing Human-Readable Explanation

If "explain":true is specified as a setting, then human-readable explaination is provided for every abuse and sentiment_expressions element.

Example

Request:

{
  "content":"u r a liar",
  "language":"en",
  "settings":{"snippets":true, =="explain":true==}
}

Response:

{
	"text": "u r a liar",
	"topics": [
		"dishonesty"
	],
	"abuse": [
		{
			"sentence_index": 0,
			"offset": 0,
			"length": 10,
			"text": "u r a liar",
			"type": "personal_attack",
			"severity": "high",
			"explanation": "Claim that discussion participant is unwelcome person"
		}
	],
	"sentiment_expressions": [
		{
			"sentence_index": 0,
			"offset": 0,
			"length": 10,
			"text": "u r a liar",
			"polarity": "negative",
			"explanation": "Unfavourable opinion",
			"targets": [
				"credibility"
			]
		}
	]
}

Tracing and Debugging Platform Decisions

In debugging mode, Tisane generates a log which can then be loaded in a special debugger environment.

Language Model Portal platform is available only for private cloud installations.

Contact us for more info.