Next word prediction as logical inference

The file arrow.pl implements a Prolog specification of the arrow.py neural net trainer and inferencer.

Both show the retrieval of the context and the actual sentence matvching the query.

As our tokens are lower case words, they need to be an exact subsequence occurring in one of the sentences for sucessful retrival.

to fetch files from guttenberg.org to data folder do:

python guttenberg.py

The try out for the Prolog-based QA:

swipl -s arrow.pl

?- query(war_and_peace,napoleon).
...
?-query(dracula,'the castle').
...

After training with arrow.py on the files in data:

./train.sh the_ayes
...
./train.sh dracula
...
./train.sh war_and_peace
...

and for inference with arrow.py on the trained checkpoints:

./generate.sh the_ayes
...
./generate.sh dracula
...
./generate.sh war_and_peace
...

---

Note: If you want ot work with other *.txt files, place them in folder data and (using to_sent_file.py) convert data/<file>.txt to data/<file>_sents.txt

Enjoy,

Paul Tarau

Jan 1, 2026