A Practical Primer on Processing Semantic Property Norm Data

Erin M. Buchanan, Simon De Deyne, & Maria Montefinese

11/13/2019

What is the property norm task?

Participants are given a list of cue words and asked to list properties of those cue words
Largely influential data for studying semantic representation and meaning, calculating various measures of similarity, feature informativeness, significance, and richness
Standardized stimuli are very useful!

Example of Data Formatted for Tidy Data
Cue	Participant Answer
airplane	you fly in it its big it is fast they are expensive they are at an airport you have to be trained to fly it there are lots of seats they get very high up
airplane	wings engine pilot cockpit tail
airplane	wings it flys modern technology has passengers requires a pilot can be dangerous runs on gas used for travel
airplane	wings flys pilot cockpit uses gas faster travel
airplane	wings engines passengers pilot(s) vary in size and color
airplane	wings body flies travel

Stemming versus lemmatization (tm vs korPus or textstem)
TreeTagger can be difficult to set up, but presents many language models
If you are willing to mix and match with python, spacyr is the R wrapper for spaCy

Lemma and Part of Speech (POS) Information from TreeTagger
Cue	Feature	POS	Lemma
airplane	is	verb	be
airplane	fast	adverb	fast
airplane	they	pronoun	they
airplane	are	verb	be
airplane	expensive	adjective	expensive
airplane	they	pronoun	they

You guys know what stopwords are
Should we remove stopwords?
However, one should be explicit about their removal of stopwords because:
- stopwords(): includes ~ 175
- tm(): options for nearly 600 words

## [1] "i"      "me"     "my"     "myself" "we"     "our"

## [1] "a"         "a's"       "able"      "about"     "above"     "according"

Multiword sequences: if you collect data in a practical way, you would simply use the finalized “clean” version of the data to combine for analysis
- Stopwords may allow of fuzzy matching for ontology labels
Bag of Words approach: tidytext presents many great functions for grouping and combining lemmas to create cue-feature lists

Idiosyncratic responses: by providing the complete feature lists, one may decide what level of response is necessary for their project

Correlation with MEN dataset ratings is rs .54 (multiword) and .70 (bag of worlds)
Positive agreement with previous analysis of this dataset

Google Statistics of DOOM
- Practical primer code explanation
- https://www.youtube.com/embed/FbznaCOXbcU
- Natural Language Processing tutorials in R and Python posted weekly