Better Vocabularies for Assistive Communication Aids: Connecting Terms Using Semantic Networks and Untrained Annotators
Sonya Nikolova, Jordan Boyd-Graber, Christiane Fellbaum, Perry Cook · 2009 · Proceedings of the 11th International ACM SIGACCESS Conference on Computers and Accessibility (Assets '09) · doi:10.1145/1639642.1639673
Summary
This paper presents the design and evaluation of ViVA (Visual Vocabulary for Aphasia), an assistive communication tool that uses semantic networks to help people with aphasia find words more easily. The authors address a fundamental problem with existing AAC devices: their vocabularies are organized in rigid, hierarchical category structures that do not reflect how humans actually store and retrieve words in memory. For people with anomic aphasia — a condition where individuals know what they want to say but cannot access the right word — these arbitrary organizational schemes make word-finding even harder. ViVA takes a different approach by building its vocabulary network on WordNet, an electronic lexical database designed to model human semantic memory. Words are connected through meaningful semantic relationships rather than placed in fixed categories. To address the sparsity of connections in WordNet, the researchers augmented it with evocation ratings — measures of how strongly one word brings another to mind — collected from over 100,000 word pairs using Amazon Mechanical Turk. The system also incorporates usage data and logistic regression to predict additional useful links between words. A key design feature is ViVA's adaptability: it learns from user behavior, adjusting vocabulary connections based on which words an individual uses frequently and in what contexts, creating personalized navigation paths.
Key findings
The evocation data collected via Amazon Mechanical Turk from untrained annotators correlated well (0.54) with ratings from trained undergraduate annotators, validating crowdsourcing as a viable method for gathering semantic association data at scale. Adding evocation-based links and simulated usage data shortened browsing paths between related words by approximately 44% compared to the original Lingraphica vocabulary hierarchy. Logistic regression predictions of additional links further improved results by 8% on average, with 22% of paths becoming shorter by two or more steps. The specific path improvements were striking: for example, navigating from "rice" to "cheese" required traversing through "home," "dictionary," "things," "house," "kitchen," "refrigerator," "dairy products" in Lingraphica, but only needed one step through "rice-cheese" in ViVA. The system demonstrated that vocabulary organization modeled on human semantic memory significantly outperforms traditional hierarchical categorization for word retrieval tasks.
Relevance
This research has significant implications for the design of AAC devices and vocabulary-based assistive technologies. It demonstrates that basing vocabulary organization on cognitive science models of semantic memory — rather than arbitrary categories — can dramatically improve word-finding for people with language impairments. The crowdsourcing methodology for collecting semantic association data is particularly valuable, as it provides a scalable, cost-effective way to build rich vocabulary networks that reflect how diverse populations conceptualize word relationships. The adaptive learning component, which personalizes vocabulary pathways based on individual usage patterns, points toward more user-centred AAC design. While the evaluation used simulated rather than real user data, the magnitude of improvement in navigation paths suggests strong potential for real-world impact on communication efficiency for people with aphasia and related conditions.
Tags: aphasia · assistive communication · semantic networks · visual vocabularies · adaptive tools · crowdsourcing · vocabulary navigation · AAC