Generating summary content using supervised sentential extractive summarization

A computing system comprising:

• means for receiving training data comprising a training text input and a summary corresponding to the training text input;
• means for generating a binary tag for each word of the training text input using a greedy algorithm;
• means for training a summary generation model by training (i) a first set of trainable parameters of a sigmoid function identifying a probability that each word of the training text input is used in the summary and (ii) a second set of trainable parameters of a contextual embedding, wherein training the first set of trainable parameters and the second set of trainable parameters is accomplished by maximizing a conditional probability of an accuracy of the binary tag generated for each word;
• means for generating a set of word embeddings corresponding to each word of a text input;
• means for generating a set of selection probabilities corresponding to each word of the text input using the respective word embeddings;
• means for calculating a set of sentence saliency scores for a set of sentences of the text input using respective selection probabilities of the set of selection probabilities for each word of the text input;
• and
• means for generating a text input summary of the text input using a subset of sentences from the set of sentences with greatest sentence saliency scores from the set of sentence saliency scores.

A non-transitory computer-readable medium having instructions stored thereon, the instructions executable by one or more processing devices to perform operations comprising:

• generating a static channel word embedding and a contextual word embedding corresponding to each word of a text input;
• generating a set of selection probabilities corresponding to each word of the text input using the respective word embeddings, wherein generating the set of selection probabilities comprises:
• concatenating a static channel word embedding and a contextual word embedding of a first word of the text input into a first vector;
• generating a first word representation of the first word by inputting a first vector comprising a trainable neural network comprising a bidirectional long short-term memory neural network;
• and
• generating a first selection probability of the set of selection probabilities using a sigmoid function comprising the first word representation and at least one trainable neural parameter of the trainable neural network;
• calculating a set of sentence saliency scores for a set of sentences of the text input using respective selection probabilities of the set of selection probabilities for each word of the text input;
• and
• generating a summary of the text input using a subset of sentences from the set of sentences with greatest sentence saliency scores from the set of sentence saliency scores.