paynesa / transitional-probabilities

A computational model of the SLM of Saffran, Aslin, and Newport (1996), which segments speech based on transitional probabilities.

Statistical Learning via Transitional Probabilities

Overview

This code provides two slightly different implementations of statistical learning via transitional probabilities as originally proposed by Saffran, Aslin, and Newport (1996) and modelled computationally by Yang and Gambell (2005) .

The transitional probability of two syllables AB is given by: TP(AB) = freq(AB)/freq(A). Both models provided here calculate these probabilities and then apply them to the input data in an attempt to segment it into words at local minima in the TP. Precision, recall, and f-score are then reported by comparing the input data to the learner's proposed segmentation on a per-utterance basis. I provide both the SLM as modelled by Yang and Gambell, and a slightly modified version wherein prediction of word boundaries is carried out separately for each utterance.

Running the Code

run_slm.py provides the traditional implementation of the SLM model using transitional probabilities as specified in Yang and Gambell (2005).

separate_utterance_slm.py provides a slightly modified version of the Yang and Gambell model that predicts word boundaries separately for each utterance rather than predicting over the entire data and then evaluating each utterance separately.

Both models take in the following parameters:

  path                  path of the file to perform SLM learning on

optional arguments:
  -h, --help            show this help message and exit
  -boundary {W,S}       boundary of interest (default=W)
  -keep_accents         keep accents when calculating TPs (default=ignore accents)

Finally, mother.speech.txt provides the annotated CHILDES data used by Yang and Gambell (2005) and provided courtesy of Charles Yang.

To replicate Yang and Gambell (2005), run python3 run_slm.py mother.speech.txt.