Small, specialised, honest about it.

English family stories

Both models were trained only on English text written as short stories about families. Other languages, lists, family trees in tables, or questions typed as a sentence will not work reliably.

Names, in the third person

Every person you want to ask about needs a name in the story, and the story should talk about them, not as “I”.

• Works: “Ashley’s daughter Lillian asked her mom to read her a story.”
• Won’t work: “My mom took me to see my grandma.” There are no names to connect.

Exam questions

You can paste a textbook blood-relation question as it is. Kinthread finds the question (“How is C related to B?”), fills in the two people, ignores answer options like “(a) Uncle”, and solves it.

• Works: “A is the brother of B. C is the father of A. How is C related to B?” Letters work as names.
• Works: “What is C to B?”, “C is related to B as?” and “How C is related to B?”
• Not yet: spoken-clue puzzles such as “Pointing to a photograph, A said, ‘He is the son of my father.’” Rewrite the clue as plain facts first: “He is A’s brother.”
• Not yet: coded relations (“A + B means A is the father of B”).

About 280 words

The model reads at most 304 tokens of question and story. Punctuation and “’s” count as tokens, so the real limit is a little under 280 words. The meter under the story shows exactly how much of the model’s context the story uses. A story can mention at most 32 different people.

Eighteen relationships

Answers are always one of these:

• aunt
• brother
• daughter
• daughter-in-law
• father
• father-in-law
• granddaughter
• grandfather
• grandmother
• grandson
• mother
• mother-in-law
• nephew
• niece
• sister
• son
• son-in-law
• uncle

Stories can mention husbands and wives, and the model uses them as steps on a path, but they are not possible answers. Neither are cousins, great-grandparents, step-relatives, half-siblings, or in-laws beyond the four listed. Asking about those gives a wrong or tentative answer.

Short chains are more reliable

A chain is the number of family facts needed to get from one person to the other. Two to four steps work well; longer chains are much harder for a model this small.

Names can be missed

The name model is very accurate on ordinary names but not perfect. It can miss a name in an unusual position (“her coworker Dana”) or mistake a place for a person (“they visited Kyoto”). Names should start with a capital letter. If someone is missing, choose Missing someone? Add a name and type the name exactly as it appears.

Unknown versus tentative

• “The story doesn’t establish a relationship” means nothing in the text connects them as family, or the model could not verify any family path. It is a deliberate refusal to guess.
• “My best guess” means the model found a path but its reasoning paths disagree about the answer. Treat it as a guess: on held-out stories these were right about half the time.

Friends are the hardest case

When the only link is a friend, coworker or neighbour, the model correctly says the story doesn’t establish a family relationship only about half the time (47% of held-out friend and stranger questions). Two families that never meet are much easier: the text check catches 96.5% of them. Be sceptical of an answer about someone who is only a friend.

Writing tips

• Use possessives: “Rahul’s mother Sunita” is clearer than “Sunita, who is a mother, and Rahul”.
• One relationship per clause, with the names close to the relationship word.
• Keep pronouns near the person they refer to, and avoid an ambiguous “they”.
• Leave out sentences that don’t describe how people are related; they add length and distraction.

Ready to try? Go back to the workbench.