Microsoft Technology Licensing, LLC

Shahar Keren

Nir Nice

Noam Koenigstein

Ulrich Paquet

Daniel Sitton

Royi Ronen

Roy Eisenstadt

Dror Kremer

Alexander Tsvetkov

Yehuda Finkelstein

Yarin Kuper

Abed El Kader Asi

Shai Roitman

Shay Ben-Elazar

Amit Perelstein

Example solutions for reducing the likelihood of hallucinations by language models, such as large language models (LLMs) are disclosed. By injecting a sufficient range and quantity of curated factual data into a prompt, the likelihood of a hallucination by an LLM may be reduced. This enables language models to be used in a wider range of settings, in which fabrication of facts is problematic, while reducing the need for a human to carefully check the generated text for accuracy. Examples include: generating a summary of a transcript using a summarization model; extracting topic-specific data from stored data using a scoring model; dynamically generating a language model prompt using the topic-specific data and the summary; and generating an output text using a language model and the language model prompt.

Dynamic construction of large language model prompts

Example apparatus and methods access multiple sources of information concerning features for applications, clean the data from the multiple sources, extract features from the cleaned data, selectively weight the sources, data or extracted features and produce a feature vector. The feature vector may then be used in a single language feature space or in a multi-language feature space. Feature spaces may then be used to find similarities between applications to facilitate recommending applications. In one embodiment, different feature spaces may be connected using a graph where nodes represent items and edges represent similarity relationships between items based on related feature spaces. Traversing the graph may allow similarities to be found that might not otherwise be possible. For example, while there may be no direct English to Hebrew similarity relationship, there may be English to French and French to Hebrew relationships that can be followed in the graph.

Multilingual content based recommendation system

Example apparatus and methods perform matrix factorization (MF) on a usage matrix to create a latent space that describes similarities between users and items and between items and items in the usage matrix. The usage matrix relates users to items according to a collaborative filtering approach. A cell in the usage matrix may store a value that describes whether a user has acquired an item and the strength with which the user likes an item that has been acquired. The latent item space may reflect true relationships between items represented in the usage matrix and those relationships may be proportional to the strength in the usage matrix. The strength of the relationship may be encoded using continuous data that measures, for example, the amount of time a video game has been played, the amount of time content has been viewed, or other continuous or cumulative engagement measurements.

Recommendation system with dual collaborative filter usage matrix

Example apparatus and methods perform matrix factorization (MF) on a collaborative filter based usage matrix to create a multi-dimensional latent space that embeds users, items, and features. A full distance matrix is extracted from the latent space. The full distance matrix may be extracted from the latent space by defining a distance metric between item pairs based on the multi-dimensional representation in the latent space. The full distance matrix may be populated with values computed for item pairs using the distance metric. A plurality of vectors associated with a multi-dimensional Euclidean space are produced from the full distance matrix. The plurality of vectors produce a navigable data set. The plurality of vectors may be produced in a manner that minimizes strain on the distances vectors. A representation of the navigable data set may be presented as, for example, a virtually traversable landscape that supports an interactive user experience.

Recommendation system with multi-dimensional discovery experience

A technique for efficiently factoring a matrix in a recommendation system. Usage data for a large set of users relative to a set of items is provided in a usage matrix R. To reduce computational requirements, the usage matrix is sampled to provide a reduced matrix R'. R&#x2032; is factored into a user matrix U&#x2032; and an item matrix V. User vectors in U&#x2032; and V are initialized and then iteratively updated to arrive at an optimal solution. The reduced matrix can be factored using the computational resources of a single computing device, for instance. Subsequently, the full user matrix U is obtained by fixing V and analytically minimizing an error in UV=R+error. The computations of this analytic solution can be divided among a set of computing devices, such as by using a map and reduce technique. Each computing device solves the equation for different respective subset of users.

Efficient modeling system for user recommendation using matrix factorization

Growing community of inventors

Hemed, Israel

Shahar Keren