Microsoft Technology Licensing, LLC

Mingyu Wang

Weibin Zhu

Zheng Chen

Yi-Min Wang

Chi Gao

Ying Li

Weizhu Chen

Pengcheng He

Qiaolin Mao

Jiayuan Huang

Jianwen Zhang

Li Rong

Changhong Yuan

Jiantao Sun

Haiyun Xu

Yuangang Tu

Robert S Harrell

Shichao Hu

Liaoyuan Ning

Jiaye Xie

Systems and devices to perform knowledge-driven dynamic conversations and select content within automated agents such as chatbots and virtual assistants are disclosed. In an example, operations to facilitate a knowledge-based conversation session with a human user using an automated agent include: receiving a conversational input regarding a support issue; analyzing the conversational input to determine an intent and applicable entity properties associated with the intent; performing a multi-turn conversation to identify a solution using the intent and the applicable entity properties, by exchanging iterative questions and answers between the automated agent and the user to dynamically recalculate applicability of the solution to the support issue; and outputting information associated with the identified solution. In further examples, the operations include a dynamic application of a solution policy and a diagnosis policy in the multi-turn conversation, to determine whether to deliver a solution or ask diagnosis questions.

Knowledge-driven dialog support conversation system

Generally discussed herein are devices, systems, and methods for visualization of user intent in accessing a virtual agent. A method can include receiving sentences from the respective interaction sessions, projecting the sentences to a higher-dimensional space to create respective higher-dimensional vectors, projecting the higher-dimensional vectors to a lower-dimensional space to create respective lower-dimensional vectors, the lower-dimensional space including either two dimensions or three dimensions, plotting the lower-dimensional vectors as respective points on a graph, encoding the respective points consistent with the respective intents determined to be associated with the sentences by an intent classifier to create encoded points, and generating a visual representation of the encoded points.

Visualization of user intent in virtual agent interaction

A keyword may be expanded into related words, such as for use in information retrieval. The terms comprising words and/or phrases of a large number of documents (e.g., web pages) are processed into a graph data structure, in which the terms are represented as nodes and edges represent the relationships between the nodes, with weights for each edge representing the relevance of the relationship. The graph may be built by selecting each term of a document and considering the terms within a certain number of words to be associated with the selected term; for each such association the weight indicative of the relevance is increased. When the graph is accessed with a keyword, the edges from that keyword's node and their respective weights indicate which other nodes are most relevant to the keyword, thereby providing the corresponding expanded terms.

Graph-based keyword expansion

A group of keywords are received, wherein each keyword includes one or more words. A word list is formed from the group of keywords, where the word list includes a list of each word in the group of keywords. A misspelled keyword is corrected using analysis of the words in the word list. The corrected keyword is output.

Self learning contextual spell corrector

Described is a paid search advertising technology in which advertisements associated with bidding keywords are ranked by relevance when returning one or more advertisements in a response to a query. A relevance score is computed for an advertisement based on the bidding keyword and page data (text and/or other page content) of the advertisement. The relevance score may be based on a similarity vector score computed from a keyword vector and page data vector relationship, combined with a proximity score computed from the keyword's bigram set and the page data bigram set. When a query is received, advertisements are selected based on the proximity of the query to each advertisement's bidding keyword, providing candidate scores. Each candidate score is modified (e.g., multiplied) into a final score based on its respective advertisement's relevance score. The final scores are then used to re-rank the advertisements relative to one another.

Relevance score in a paid search advertisement system

Growing community of inventors

Beijing, China

Mingyu Wang