Skip to content

Conclusions

DRAFT (WIP)

TextGraphs library provides a highly configurable and extensible open source Python library for the integration and evaluation of several LLM components. This has been built with attention to allowing for concurrency and parallelism for high-performance computing on distributed systems.

TODO:

  • leverage co-reference
  • leverage closure constrained by domain/range
  • general => specific, uncertain => confident

The state of relation extraction is arguably immature. While the papers in this area compare against benchmarks, their training datasets mostly have been built from Wikidata sources, and inferred relations result in labels not IRIs. This precludes downstream use of the inferred relations for semantic inference. Ultimately, how can better training data be developed -- e.g., for relation extraction -- to improve large models used in constructing/augmenting knowledge graphs?

Questions for Follow Up Research

Many existing projects produce results which are descriptive, but not computable. However, given recent innovations, such as DPO, there appear to be many opportunities for reworking the training datasets used in NRE and RE models, following the pattern of Notus

R1: we have demonstrated how to leverage LLM components while emphasizing HITL (domain experts) and quality of results

R2: we have suggested areas where investments in data quality may provide substantial gains

One key take-away from this project is that the model deployments are relatively haphazard across a wide spectrum of performance: some of the open source dependencies use efficient frameworks such as Hugging Face transformers to load models, while others use ad-hoc approaches which are much less performant.

Granted, use of LLMs and other deep learning models is expected to increase computational requirements substantially. Given the integration of APIs, the compute, memory, and network requirements for running the TextGraphs library in product can be quite large. Software engineering optimizations can reduce these requirements substantially through use of hardware acceleration, localized services, proxy/caching, and concurrency.

However, a more effective approach would be to make investments in data quality (training datasets, benchmarks, evals, etc.) for gains within the core technologies used here: NER, RE, etc. Data-first iterations on the model dependencies can alleviate much of this problem.

R3: we have proposed a rubric for evaluating/rating ML open source w.r.t. production use cases

This project integrates available open source projects across a wide range of NLP topics. Perspectives were gained from evaluating many open source LLM projects related to NLP components, and the state of readiness for their use in production libraries overall.

Note that reproducibility rates are abysmally low for open source which accompanies machine learning research papers. Few project install correctly, and fewer still run without exceptions. Even among the better available OSS project for a given research topic (e.g., graph embeddings, relation extraction) tend to not have been maintained for years. Of the projects which run, few reproduce their published results, and most are oriented toward command-line (CLI) use to prove specific benchmarks claims. These tend to be difficult to rework into production-quality libraries, due to concerns about performance, security, licensing, etc.

As an outcome of this inquiry, this project presents a rubric for evaluating research papers and their associated code, based on reproducibility and eventual usefulness in software implementations.

The views expressed are those of the authors and do not reflect the official policy or position of the funding organizations.