Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Poster Slams

Probing the categorical structure of fifty abstract words

Poster D45 in Poster Session D with Social Hour, Friday, October 7, 5:30 - 7:15 pm EDT, Millennium Hall
Also presenting in Poster Slam D, Friday, October 7, 5:15 - 5:30 pm EDT, Regency Ballroom

Andrew Persichetti1, Jiayu Shao1, Stephen Gotts1, Juan Antonio Lossio-Ventura1, Francisco Pereira1, Alex Martin1; 1National Institute of Mental Health

Our lexicon can be divided into concrete concepts that describe perceivable entities, such as “dog” and “truck,” and abstract concepts that refer to intangible entities, such as “frustration” and “truth.” While concrete concepts are relatively easy to categorize based on shared physical and functional properties, attempts to characterize the categorical structure of abstract concepts have not been as successful. We sought to uncover the categorical structure of abstract concepts using a rigorous, multipronged approach in which we first found categorical boundaries within a set of abstract words using data from an implicit judgement task and then tested the behavioral validity of those category boundaries using automatic semantic priming. First, we chose fifty words that were identified as abstract in a prior study. Next, 414 participants (194 female; mean age=40.3, s.d.=12.2) made implicit similarity judgements about the fifty abstract words during an odd-one-out triplet task on Amazon’s Mechanical Turk. We used the results from this task to derive a similarity matrix for the fifty abstract words. To ensure that the similarity matrix was stable, we ran the same task online in a separate group of 414 participants. The matrices were highly similar (r=0.93). We then used principal components analysis and k-means clustering to analyze the matrix and divide the words into candidate categories. Next, we removed clusters that contained less than three words and then removed any unstable words from the remaining clusters. We determined cluster stability by comparing them with clusters from a word embedding obtained using fastText to represent the fifty abstract words in a 300-dimensional space, then applying UMAP to reduce the dimensions, followed by HDBSCAN and k-means to group similar words together. After comparing the clustering solutions, we were left with thirty abstract words that were separated into five categories for the automatic semantic priming experiment. We also included thirty concrete words from five categories as a necessary control condition. During the priming experiment each participant (N= 12, to-date) was presented with 1312 trials, so that every possible pair from within each category was shown twice along with an equal number of between-category and between-word-type pairs. In each trial, a prime word was presented for 100 ms, followed by a 50 ms mask and 100 ms blank screen, then a probe word for 250 ms, and finally a blank screen for 1000 ms. The participants were asked to respond using the keyboard whether the probe word was abstract or concrete. Overall, accuracy was very high (95.1%) and did not differ across within- and between-category trials for abstract or concrete words (both p’s>0.15). Paired t-tests revealed significantly faster response times to within-category word pairs relative to between-category pairs (i.e., priming) for both abstract and concrete words (both p’s<0.05). These results demonstrate that abstract words can be organized into behaviorally relevant categories assumed to reflect yet-to-be-determined shared properties. We plan to further probe the categorical structure of abstract words using other methods, including neuroimaging, behavioral tasks, and computational modeling.

Topic Areas: Meaning: Lexical Semantics, Methods

SNL Account Login

Forgot Password?
Create an Account