This repository contains analyses and data from the following work:

Grewal, K. and Xu, Y. (to appear) Chaining and historical adjective extension. In Proceedings of the 42nd Annual Meeting of the Cognitive Science Society.

The data/ folder contains all historical adjective-noun cooccurrence counts grouped by decade, as well as adjectives and nouns used in our experiments. The following python code (executed from the scripts/ folder) prints the raw co-occurrence count between "artificial" and "computer" during the 1980s.

from util import *

# load the adjective set FRQ-200
adjectives = load_adjectives('frq200')

# retrieve the set of nouns to consider in the 1970s given our choice of 
# adjective set
noun2ind = pickle.load(open('../data/wordnet/nouns_wn.pkl', 'rb'))
ind2noun = {i: a for a, i in noun2ind.iteritems()}
nouns_t_i = get_noun_indices_t(t)
nouns_t = [ind2noun[i] for i in nouns_t_i]

# C is a 3D numpy array where the (t, i, j) entry gives the raw cooccurrence 
# count between noun i and adjective j in decade t (offset by 1970). 
C = load_cooc(1970, 1990)

# print the raw co-occurrence count between "computer" and "artificial" in the 
# 1980s
i, j = nouns_t.index('computer'), adjectives.index('artificial')
print(C[1, i, j])

The scripts/ folder contains the exemplar, prototype, progenitor, k-nearest neighbors, and baseline models. One of these models can be executed as

$ python scripts/exemplar.py --frq200 --precision

where --frq200 specified the "FRQ-200" adjective set, and --precision specifies we want to evaluate models based on predictive accuracy. Instead, you can use --jsd to specify the Jensen-Shannon divergence between the predicted and empirical posterior distributions over adjectives. In addition, the options --future and --map can be used to consider all future pairings when evaluating models and to use kernel parameters learned via MAP estimates, respectively.

Note that the prototype and progenitor models are both wrapped into scripts/prototype.py. Also, you will need to download diachronic Word2Vec embeddings on your own and place the .pkl and .npy files in data/diachronic_embeddings. The embeddings can be found here.