This model was trained on tail and entity prediction in a knowledge graph using the graph’s context represented by the node’s neighborhood. Textual representation was obtained from wikidata entities and relation titles. Entity description was used to disambiguate if two entities had the same title. If still, no disambiguation was possible, we assigned unique numerical ids for such entities. 1. sort the neighborhood by semantic similarity of relations from its triplets to the relation from the input triplet in order to prioritize more important information in the context; 2. limit the sorted neighborhood to 512 triplets, since this will always be at least as big as the size of the allowed context, and, after verbalization, specify the maximum length of 512 for the model tokenizer to fit the resulting verbalized neighborhood representation into the language model context. Neighborhood sorting by semantic proximity was performed using a pre-calculated matrix of cosine similarity of relations in KG, for similarity calculation the relations were embedded by the fasttext model. We trained the model on the Wikidata5M dataset for approximately 5M iterations on 8xA100 GPUs using a batch…
Модальности:
Генерация текста
Задача: Генерация текста
Автор: DeepPavlov
Теги: t5, text2text-generation, en, text-generation-inference, endpoints_compatible
Лайков: 3 | Загрузок: 17
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
