'I Commit' ('מתחייב אני'): A Computational Framework for Monitoring Political Commitments in Israeli Public Discourse

Political promises are central to democratic representation, yet not all promises bind equally: specific, direct, unhedged commitments create clear expectations against which conduct can be judged, while vague, conditional, or attributed ones leave room for reinterpretation and blame-shifting. Pledge trackers evaluate whether promises are kept, overlooking how commitments are formulated. We combine annotated promises with labeled Knesset speeches, Facebook posts, and X posts to build a two-stage pipeline for detection and tagging (markers, specificity, hedging, attribution, beneficiaries), benchmarking fine-tuned Hebrew transformers (DictaBERT, HeBERT) against multilingual LLMs. The corpus and classifier turn long-standing questions into traceable, large-scale comparisons: do representatives calibrate commitments between the plenary floor and a tweet, across parties and policy domains, or for in-group versus out-group audiences? The project provides the first open resource for studying Hebrew commitment-making, showing that how commitments are made bears on what we should expect from representation, and may foreshadow which promises hold.