Understanding and processing speech links is a big challenge to guide natural language processing in the context of AI assistants like Amazon Alexa. This problem, as a rule, implies the correct association of pronouns in user queries with implied concepts, for example, matching the pronoun “them” in the statement “play their latest album” with some musical artist. AI specialists at Amazon are actively working on technology that could help AI with processing such requests through automatic reformulation and replacement. So, the request "Play their latest album" will automatically be replaced with "Play the latest Imagine Dragons album". At the same time, the word necessary for replacement is selected in accordance with the probabilistic approach calculated using machine learning.
Scientists a preliminary result of his work in a preprint with a rather complicated title - "Scaling the state tracking of a multi-domain dialogue using query reformulation." In the near future, it is planned to present this study at the North American branch of the Association for Computational Linguistics.
“Because our query reformulation mechanism uses the general principles of using speech links, it does not depend on any specific information about the application where it will be used, so it does not require retraining when we use it to extend the capabilities of Alexa,” explained Arit Gupta (Arit Gupta), Linguistics Expert at Amazon Alexa AI. He noted that their new technology, called CQR (contextual query rewriting), completely frees the voice assistant's internal code from any concern about speech links in queries.
First, AI determines the general context of the request: what information the user wants to receive or what action to perform. In the process of dialogue with the user, the AI classifies the keywords, saving them in special variables for further use. If the next request contains any link, then the AI will try to replace it with the most likely of the stored and semantically suitable words, and if this is not in memory, it will turn to the internal dictionary of the most frequently used values, and then rebuild the request with the replacement applied, to pass it on to the voice assistant for execution.
As Gupta and colleagues point out, CQR acts as a pre-processing layer for voice commands and focuses only on the syntactic and semantic meanings of words. In experiments with a specially prepared dataset, CQR improved query accuracy by 22% when the link in the current query refers to the word used in the most recent response, and by 25% when the link in the current utterance refers to a word from the previous utterance.
Source: 3dnews.ru