For detailed explanations of the software, why we need it, what it can do etc., please read the pdf in this repo (will be available soon).
SoftStress can learn a stress pattern with a Maximum Entropy learner (Goldwater and Johnson 2003) using the Gradient Descent as its optimization algorithm, combined with Expectation Maximization (Dempster et al. 1977) to deal with the presence of hidden structure (Tesar and Smolensky 2000, Jarosz 2013). The learner is first presented in Pater et al. 2012, and the source code for the learner came from Brandon Prickett's github.
It can also compute the minimal set of weights that represent a stress pattern via Linear Programming, introduced in Potts et al 2010 and implemented in OT-Help. The current software can handle larger and more complicated problems.
You can investigate learnability/representability of any of the 61 attested stress patterns [click to see these patterns], without any input files, using five predefined constraint sets:
Foot: OriginalTS, RevisedTS, RevisedTS_nonfinmain
Grid: OriginalGordon, RevisedGordon
For the definitions of these constraints, see the pdf manual.
If the stress pattern you want to investigate is not in the list of 61 patterns, you can also upload your own input file (see the Example input file format)
Some code snippets... click on the link to the google colab demo, to try them yourself
To see the pattern (winners for each UR),
read_winners(filename, QI_or_QS)
To learn a pattern
final_weights, learned_when = learn_language(filename, QI_or_QS, Foot_or_Grid, Constraint_set)
print_result_pretty(filename, QI_or_QS, Foot_or_Grid, Constraint_set, final_weights, learned_when, con_suffix = myConstraintsName)
To solve a pattern (to see if a pattern is representable with the constraint set)
solutions = solve_language(filename, QI_or_QS, Foot_or_Grid, Constraint_set)
print_solutions_pretty(filename, QI_or_QS, Foot_or_Grid, Constraint_set, solutions, con_suffix = myConstraintsName)
If you want to contribute to the database/program by adding more constraints or adding a stress pattern that's not one of the 61 patterns, please let me know: seungsuklee[at]umass[dot]edu
This research and software development was supported by NSF BCS-2140826 awarded to the University of Massachusetts Amherst.