Abstract: The present application introduces improved methods for removing erroneous sections (e.g. hallucinated sentences) from computer-generated summaries. This improves the accuracy of the resultant summaries; but outputting corrected summaries for which the erroneous sentences have been removed. Importantly, the methods described herein do not require the training of any additional machine learning models, but instead work solely based on probabilities generated by the summary generation neural network that generates the summaries. Furthermore, the methodology described herein is able to work for any type of summary generation neural network.

Abstract: According to an aspect there is provided a computer-implemented method for determining summaries of text over multiple batches of text. The method comprises: for each of a plurality of batches of text, each batch comprising text for addition to a cumulative document: adding the batch of text to the cumulative document to produce an updated cumulative document; encoding the updated cumulative document using an encoder neural network to obtain one or more encoder hidden states; inputting the one or more encoder hidden states and a cumulative summary that summarises each preceding batch of text into a decoder neural network to generate a summary for the batch of text; and updating the cumulative summary by adding the summary to the cumulative summary. The method further comprises outputting each summary.

Abstract: This application provides systems and methods for training a language model to perform one or more specific natural language processing tasks. The embodiments described herein fine-tune language models for downstream tasks solely by pre-processing the training data set. Rather than fine-tuning via architecture changes (e.g., addition of classification layers on top of a language model), the embodiments described herein fine-tune language model(s) via dataset pre-processing alone. This is much simpler for the practitioner. Furthermore, it allows iterative additions of functionality to the language model without a complete restructure of the architecture. This is possible because of the general nature of the language-modelling task, which essentially consists of predicting what comes next in a sequence given some context. If training data can be framed in this manner, a language model can be used to solve that task directly without architecture modifications.

Abstract: Embodiments described herein provide a more flexible, effective, and computationally efficient means for determining multiple intents within a natural language input. Some methods rely on specifically trained machine learning classifiers to determine multiple intents within a natural language input. These classifiers typically require a large amount of labelled training data in order to work effectively, and are generally only applicable to determining specific types of intents (e.g., a specifically selected set of potential inputs). In contrast, the embodiments described herein avoid the use of specifically trained classifiers by determining inferred clauses from a syntactic graph of the input. This allows the methods described herein to function more efficiently and over a wider variety of potential inputs.

Abstract: The disclosed system addresses a technical problem tied to computer technology and arising in the realm of computer memory capacity, namely the technical problem of providing a flexible response dialogue system that can be utilised for a variety of different types of dialogue without requiring the system to be specifically trained for each situation. This therefore avoids the need for large amounts of labelled training data for each type of dialogue (each potential conversation flow or subject area for the conversation). The disclosed system solves this technical problem by using semantic similarity to match a user's input to one of a set of predefined inputs (predefined user responses). Various mechanisms are implemented to provide disambiguation in the event of multiple potential matches for the input. By using semantic similarity, the user's response in unconstrained. This therefore provides a user interface that is more user-friendly.

Abstract: Methods for determining whether two sets of words are similar are provided. In one aspect, a method includes receiving a first set of words and a second set of words, whichare subsets of a vocabulary, and each of the first and second sets of words include word embeddings corresponding to each word. The method also includes determining a word membership function for each word in the vocabulary. Determining the word membership includes determining a set of similarity values, each representing the similarity between the word and a respective word in the vocabulary. The method also includes determining a membership function for the first and second sets of words based on the determined word membership functions, and determining a set-based coefficient for the similarity between the first and second sets of words based on the membership function. Systems and devices are also provided.

Abstract: Embodiments described herein provide a more flexible, effective, and computationally efficient means for determining multiple intents within a natural language input. Some methods rely on specifically trained machine learning classifiers to determine multiple intents within a natural language input. These classifiers require a large amount of labelled training data in order to work effectively, and are generally only applicable to determining specific types of intents (e.g., a specifically selected set of potential inputs). In contrast, the embodiments described herein avoid the use of specifically trained classifiers by determining inferred clauses from a semantic graph of the input. This allows the methods described herein to function more efficiently and over a wider variety of potential inputs.

Abstract: The disclosed system addresses a technical problem tied to computer technology and arising in the realm of computer memory capacity, namely the technical problem of providing a flexible response dialogue system that can be utilised for a variety of different types of dialogue without requiring the system to be specifically trained for each situation. This therefore avoids the need for large amounts of labelled training data for each type of dialogue (each potential conversation flow or subject area for the conversation). The disclosed system solves this technical problem by using semantic similarity to match a user's input to one of a set of predefined inputs (predefined user responses). Various mechanisms are implemented to provide disambiguation in the event of multiple potential matches for the input. By using semantic similarity, the user's response in unconstrained. This therefore provides a user interface that is more user-friendly.

Abstract: Methods for determining whether two sets of words are similar are provided. In one aspect, a method includes receiving a first set of words and a second set of words, which are subsets of a vocabulary, and each of the first and second sets of words include word embeddings corresponding to each word. The method also includes determining a word membership function for each word in the vocabulary. Determining the word membership includes determining a set of similarity values, each representing the similarity between the word and a respective word in the vocabulary. The method also includes determining a membership function for the first and second sets of words based on the determined word membership functions, and determining a set-based coefficient for the similarity between the first and second sets of words based on the membership function.

Abstract: Embodiments described herein provide a more flexible, effective, and computationally efficient means for determining multiple intents within a natural language input. Some methods rely on specifically trained machine learning classifiers to determine multiple intents within a natural language input. These classifiers typically require a large amount of labelled training data in order to work effectively, and are generally only applicable to determining specific types of intents (e.g., a specifically selected set of potential inputs). In contrast, the embodiments described herein avoid the use of specifically trained classifiers by determining inferred clauses from a syntactic graph of the input. This allows the methods described herein to function more efficiently and over a wider variety of potential inputs.

Abstract: Embodiments described herein provide a more flexible, effective, and computationally efficient means for determining multiple intents within a natural language input. Some methods rely on specifically trained machine learning classifiers to determine multiple intents within a natural language input. These classifiers require a large amount of labelled training data in order to work effectively, and are generally only applicable to determining specific types of intents (e.g., a specifically selected set of potential inputs). In contrast, the embodiments described herein avoid the use of specifically trained classifiers by determining inferred clauses from a semantic graph of the input. This allows the methods described herein to function more efficiently and over a wider variety of potential inputs.