http://www.sigmorphon.org/conll2017/

This is a very simple baseline algorithm for solving tasks 1 and 2. The system has a strong bias toward only accounting for prefix and suffix changes when inflecting a form from a given lemma.

When run with the -o option, the program writes its outputs into files. The default behavior is to just run the algorithm on all tasks and languages and print out accuracy on the dev set.

The default relative path to the train/dev files (./../) can be changed with the -p flag.

As a first step, the system aligns input/output training examples using Levenshtein distance. For example, the example

schielen        geschielt       V.PTCP;PST

is aligned as follows:

__schielen
geschielt_

The system now assumes that each input-output pair can be broken up into a prefixation part, a stem part, and a suffixation part, based on where the inputs or outputs have initial or trailing zeroes:

Pr   St   Su
__|schiele|n
ge|schielt|_

After this, the system extracts a set of prefix-changing rules based on the Pr pairings, and a set of suffix-changing rules based on St+Su pairings.

In this example, the following suffix-changing rules are extracted:

n$ > $
en$ > t$
len$ > lt$
elen$ > elt$
ielen$ > ielt$
hielen$ > hielt$
chielen$ > chielt$
schielen$ > schielt$

Likewise, the only prefix-changing rule extracted is the following:

$ > $ge

For languages with few prefix changes, the only prefix rule will often be $ > $, i.e. no change.

All these rules are associated with the complete MSD of an example, in this case V.PTCP;PST.

At generation time, the longest suffix rule that applies to a lemma form to be inflected is used. For example, if we're asked to inflect kaufen into V.PTCP;PST, we may find that, for example, en$ > t$ is the longest-matching suffix rule among all rules for V.PTCP;PST, which transforms:

$kaufen$ > $kauft$

and, following this, we find the most frequently seen prefix-changing rule for the target MSD in question that is applicable, in this case $ > $ge and apply that:

$kaufen$ > $gekauft$

If no rule has been associated with a particular MSD combination, the lemma form is simply repeated.

In task 2 (the paradigm cell-filling task), all the forms are generated from the lemma form. That is, the system does not take advantage of the given forms in a partially filled table in any way.

There is a heuristic to decide if a language is largely prefixing or largely suffixing. This is done by simply counting how often there is a prefix-change vs. suffix-change in going from the lemma form to the inflected form. Whenever a language is found to be largely prefixing, the system works with reversed strings throughout.