This code is a development of the original word learning model FAS10, which introduces several in-the-moment learning mechanisms. We formulate such mutual exclusivity biases brought by these mechanisms as word competitions, referent competitions and no competition.
A script called scripts/SAMPLE_RUN.sh can be run and is a good starting point to get familiar with the program.
To fully replicate the experiment results mentioned in the paper Fig.2, you can run the scripts/prepare_four_alignment.sh. It automatically runs all 4 alignment methods with training time 20k, and generate all four learning curves in one graph. The graphs are under folder plot/ when completed.
Configuration settings such alignment method, training time, similarity measure can be adjusted in config.ini.
It is worthy to mention that the most important parameter inside configuration file isalignment-method which controls the alignment mechanism. You can switch to different word-referent alignment model by inputting different integers to the variable.
If it's 0, the word learning model aligns word and feature directly as in FAS10.
If it's 1, the model runs referent competition (i.e. ref-comp) aligning each word with all the referents in the scene.
If it's 2, the model swithes to word competition (i.e. word-comp), in other words all words are competing for a referent during the alignemnts.
If it's 3, then then there's no competition among words or referents (i.e. no-comp).
Training time is set to 20k by default since the learning curve stays robust afterwards.
Under data/ folder, a list of training data as well as the gold lexicon used in the experiments are provided.
data/all_features_included.all is the gold lexicon. For a given word, the probability distribution of a set of semantic features representing its gold-standard meaning is uniform.
In each training data, every utterance is paired with a scene. The scene consists of a set of ALL meaning features for ALL words in the utterance. Different set of semantic features belonging to each word in the sentence are separated by semicolons.
all_features_included.dev is the training data to generate the overall learning performance (Fig.2 in paper).
mlu_*.txt are datas to test the effect of Mean Length of Utterance. In particular, mlu_long.txt only consists of utterances of length greater than 4 whereas mul_short.txt are made up by utterances of length less than 4.
We also experimented on the effect of referential certainty, the datas are the following: rw_one.txt, rw_two.txt, rw_three.txt.
seaborn, matplotlib, nltk, numpy, scipy
http://www.cs.toronto.edu/~aida/papers/nematzadeh_etal_17_cogsci_alignments.pdf