Many approaches to English financial text analysis still rely on keyword or rule-based extraction, with limited trust in sentiment models despite advances in contextual understanding. Past studies have explored concepts such as aspect-based sentiment analysis and named entity recognition, yet none address how entities appear implicitly through context rather than direct mentions, or provide a dataset that brings these elements together. This gap limits how well models capture the links between entities, aspect, and sentiment. We introduce FinER-ABSA, a benchmark that integrates implicit and explicit entity recognition with aspect-based sentiment in financial text. Experiments on seven open-source large language models under zero- and few-shot settings show that even the best systems still miss key aspects of implicit reasoning. In the few-shot case (K= 3), Llama-3.3-70B reached an F1 of 0.7623 for implicit entities, suggesting that while models can detect signals, their consistency remains far from the level of reliability required for financial analysis or decision-making. These insights emerge only through FinER-ABSA, which makes such gaps measurable and advances financial Natural Language Processing (NLP) toward deeper contextual understanding and enables systems that better extract comprehensive insights from market-moving information in an industry where such precision is critical.

Instruction-following machine translation (IF-MT) requires respecting prompt-level rules on terminology, formatting, and register. Rule compliance typically trades off against translation quality, a tension that general-purpose IF data augmentation methods do not address. We propose Reference-Grounded Data Curation, a two-phase pipeline that extracts every supervised constraint from a reference translation that already satisfies it, ensuring feasibility by construction. Phase 1 applies Instruction-Following Difficulty (IFD) scoring to retain the hardest-but-learnable instances from an English-Thai parallel pool. Phase 2 extracts constraints from each reference target and keeps only generations satisfying every constraint, yielding the 1.97M-record Grounded dataset. We fine-tune open-weight bases on Grounded to produce ChindaMT, a Thai-English translation family at 4B, 2B, and 0.8B parameters. Under length-controlled pairwise judging, ChindaMT outperforms or matches every same-size baseline at every tier on both plain translation and under explicit rules, reaching up to a 68.4% win rate against the strongest baseline. The recipe transfers cleanly across Qwen generations. We release model weights, the Grounded dataset, and evaluation suites.