LOCATION EXTRACTION FOR VOICE CONTROLLED ROBOT GUIDES

Lin, Justin

Permalink

https://hdl.handle.net/2142/124836 Copy

Description

Title

LOCATION EXTRACTION FOR VOICE CONTROLLED ROBOT GUIDES

Author(s)

Lin, Justin

Issue Date

2023-05-01

Keyword(s)

Location Extraction, Visual Impairments, Natural Language Processing, Voice Controlled Robots

Abstract

Finding a restroom in unfamiliar buildings, is especially challenging for people with visual impairments (PwVI). A robot guide that can verbally communicate with PwVI and guide them to their desired destinations would significantly increase their independence and quality of life. Rooms and locations with similar names can confuse robots from determining the correct location from the speech commands given to them. Therefore, robots require more detailed commands to decipher the correct goal location compared to the human brain. In this thesis, we present a natural language understanding system that can extract the locations where PwVI want to go from natural language commands. The instructions in the voice commands can contain the target location and relational locations amidst a lot of noise. Na¨ ıve methods like a bag-of-words model and previous learning based methods either cannot differentiate the target and relational locations or require an additional component of image mapping alongside the natural language command. To this end, We introduce a novel location token classifier where we fine tune a token classification model based on Bidirectional Encoder Representation from Transformers (BERT) on a custom dataset. The model allows the user to give the robot any natural language input commands and outputs the labeled entities, target location, and relational location. We use a custom dataset curated for navigation for PwVI and show improvements over bagof-words basline in extracting goal locations. We also deploy our model in a robot navigation guide and show that we can locate the user’s goal location from speech commands to provide a successful/effective guiding experience.

Type of Resource

text

Language

eng

Owning Collections