Skip to main navigation Skip to search Skip to main content

Using Compression Algorithms to Map the Mental Lexicon and Improve AI Alignment

Project: Research

Project Details

Description

In visual word recognition, people decompose words into constituents, and words that share constituents are linked in the mental lexicon. Some evidence suggests that people decompose words to the same constituents as a compression algorithm called byte-pair encoding (BPE), which would imply that visual word recognition approaches optimal efficiency and could improve models of the mental lexicon. Furthermore, most large language models (LLMs) use BPE to segment text into tokens, and if tokens map onto psychologically real constituents, it could facilitate concept alignment. That is to say,BPE could help LLMs interpret words like people do. However, supplying LLMs with human-like constituents is not enough; concept alignment also requires that LLMs use those constituents in a human-like manner. People use orthographic constituents selectively, with stronger formal associations between words with fewer semantic associates, which averts failures in word meaning access. So far, LLMs seem to lack this tendency, suggesting that adaptive strategies which people use to compensate for underdetermined meanings are not recovered through statistical learning, so LLMs miss this opportunity for concept alignment.To test these possibilities, we will conduct three sets of experiments. Study 1 will use masked priming in lexical decisions to provide evidence that BPE identifies psychologically real orthographic constituents in English. For example, according to one promising BPE model, people should recognize “allure” faster following “pasture”, which shares the final constituent “ure”, than following “stature”, which has the final constituent “ature”. Studies 2 and 3 will provide evidence that BPE constituents correspond to links in the mental lexicon and that people use those links to supplement representations of vague or unfamiliar words (e.g., “gypsum”) with orthographically related words (e.g., “gypsy”). To investigate these links in the mental lexicon, we will construct multiplex networks, which simultaneously consider orthographic and semantic connections between words, and we will evaluate their performance based on how well they model word associations in Dutch, English, and Spanish, generated by people and LLMs.We hypothesize that people internalize complex orthographic patterns, which influences how people retrieve and represent word meanings, and that although LLMs have direct access to those patterns, they lack adaptive strategies that are essential to human word processing. These findings would support theories of language evolution, improve models of the mental lexicon, and identify potential sources of AI misalignment.
Project number9044140
Grant typeGRF
StatusNot started
Effective start/end date1/01/27 → …

Fingerprint

Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.