NL4Opt Competition: Formulating Optimization Problems Based on Their Natural Language Descriptions

Abstract: The Natural Language for Optimization (NL4Opt) Competition was created to investigate methods of extracting the meaning and formulation of an optimization problem based on its text description. Specifically, the goal of the competition is to increase the accessibility and usability of optimization solvers by allowing non-experts to interface with them using natural language. We separate this challenging goal into two sub-tasks: (1) recognize and label the semantic entities that correspond to the components of the optimization problem; (2) generate a meaning representation (i.e., a logical form) of the problem from its detected problem entities. The first task aims to reduce ambiguity by detecting and tagging the entities of the optimization problems. The second task creates an intermediate representation of the linear programming (LP) problem that is converted into a format that can be used by commercial solvers. In this report, we present the LP word problem dataset and shared tasks for the NeurIPS 2022 competition. Furthermore, we investigate and compare the performance of the ChatGPT LLM against the winning solutions. Through this competition, we hope to bring interest towards the development of novel machine learning applications and datasets for optimization modeling.

• The paper introduces a contest to translate natural language descriptions into solver-ready optimization formulations.
• Participants used techniques like ensemble learning and transformer models, achieving high F1 scores and accuracy metrics.
• The study highlights the potential of pre-trained large language models while addressing challenges in input variability and domain adaptation.

NL4Opt Competition: Formulating Optimization Problems Based on Their Natural Language Descriptions

Introduction

The NL4Opt Competition centered around converting natural language descriptions of optimization problems into mathematical formulations suitable for commercial solvers. This was achieved by introducing two sub-tasks: entity recognition and problem formulation generation. The initiative aimed to make optimization tools accessible to non-experts, thus broadening their application across different domains.

Competition Structure

Sub-task 1: Entity Recognition

The first sub-task required participants to identify semantic entities within a problem description. These entities correspond to different components of the optimization problem, such as constraints and variables. The goal was to unambiguously tag these components to facilitate further processing.

Sub-task 2: Generation of Problem Formulation

Participants generated a logical representation of optimization problems using identified entities from the first sub-task. The task involved transforming the natural language descriptions into a machine-readable form, aligning closely with formats accepted by optimization solvers.

Achievements and Winning Strategies

The competition saw diverse methods with several teams leveraging ensemble learning and data augmentation techniques to boost performance. For instance, the successful application of transformers like XLM-R, DeBERTa, and BART, often combined with CRF for sequence tagging, were notable among winning strategies.

1. Sub-task 1 Results: The top team, Infrrd AI Lab, achieved an F1 score of 0.939, utilizing ensemble learning with model variations and data augmentation strategies. Other ranked teams employed similar approaches with advanced ensemble techniques and adversarial training methods.
2. Sub-task 2 Results: UIUC-NLP achieved a high accuracy through input enrichment and the usage of large transformer models, demonstrating the potential for innovative input representation and tagging strategies in improving generation accuracy.

Dataset and Evaluation

The NL4Opt dataset comprised 1101 annotated problems from various domains, ensuring coverage across differing problem structures. Evaluation metrics were primarily based on precision and recall for the entity recognition task and accuracy of the generated formulations with reference to ground-truth declarations.

Insights and Future Directions

LLMs: The post-competition evaluation revealed that models like ChatGPT outperformed many participant-submitted methods in problem formulation tasks, despite not being tuned on the dataset. This indicates the potential of pre-trained large models in natural language processing tasks beyond their original scope.

Challenges and Considerations: Key challenges included managing unstructured multi-sentence input, ensuring robust model training with limited annotated resources, and generalizing across problem domains. Overcoming these obstacles paves the way for practical deployment in various industrial applications.

Conclusion

The NL4Opt Competition highlighted machine learning's potential to enhance the accessibility of optimization tools through natural language interfaces. The innovations in entity recognition and problem formulation generation demonstrated significant progress, yet challenges like input variability and domain adaptation remain. Future research should aim to further explore these aspects, advancing the integration of AI and optimization in real-world scenarios.