webis-de / ir_axioms Goto Github PK

Document ir_axiom's different variants for obtaining term similarities in documentation/similarity.md

We had to adapt some axiom definitions to make them applicable to queries of arbitrary length.
We should document and/or formalize the specific changes in documentation/axioms.md

At the moment, we need the text of the documents (e.g., provided by ir_datasets).
The goal of this ticket is, that ir_axioms works standalone from an PyTerrier Index.

The STMC1, STMC2, and REG axioms depend on similarities. It might be promising to have similarity variants that are based on word embeddings fine-tuned on test collections like Robust 04.
It is only required to have some word embeddings that PyMagnitude can handle, the rest is already implemented.

To implement new embedding variants, take a look at the fastText implementation here:

Define a new similarity mixin extending MagnitudeTermSimilarityMixin and then implement variants of STMC1, STMC2, and REG like this:

Hi everyone! First of all, this is really nice work, that could potentially put axioms back into fashion =)
However, I've noticed that at least when using the system naively, there is no support for parallel execution (via threads of processes). So, if you want to extract all triples of preferences from, say, the training set of MsMarco, it is unfeasible to be done.

To be more specific, when talking about the axiomatic preferences modules, at least in theory, it should be possible to start a pool of workers and a queue of jobs, where each job is a query (and its corresponding ranking).

Alternatively, I've seen that you mention that ir-axioms is compatible with SLURM commands. How one would go to start a (series of) slurm jobs using it?

Thanks!

To improve a combined axiom's performance we could simplify complex axiom expressions using pattern matching with MatchPy. The current approach only "optimizes" by overloading the operator functions, e.g., __and__ in AndAxiom, but cannot resolve deep nested axiom expressions.

We could backport some functionality, so that ir_axioms also works with Python 3.7.
This would allow users to experiment with ir_axioms in Google Colab

Define a simple command line interface (CLI) to use axioms with existing run files, qrels etc.
We should then support most of the use cases from the CLI.

We could generate a documentation website using Sphynx.