Large language models (LLMs) have demonstrated significant capability to generalize across a large number of NLP tasks. For industry applications, it is imperative to assess the performance of the LLM on unlabeled production data from time to time to validate for a real-world setting. Human labeling to assess model error requires considerable expense and time delay. Here we demonstrate that ensemble disagreement scores work well as a proxy for human labeling for language models in zero-shot, few-shot, and fine-tuned settings, per our evaluation on keyphrase extraction (KPE) task. We measure fidelity of the results by comparing to true error measured from human labeled ground truth. We contrast with the alternative of using another LLM as a source of machine labels, or ‘silver labels’. Results across various languages and domains show disagreement scores provide a better estimation of model performance with mean average error (MAE) as low as 0.4% and on average 13.8% better than using silver labels.

In today’s interconnected and multilingual world, code-mixing of languages on social media is a common occurrence. While many Natural Language Processing (NLP) tasks like sentiment analysis are mature and well designed for monolingual text, techniques to apply these tasks to code-mixed text still warrant exploration. This paper describes our feature engineering approach to sentiment analysis in code-mixed social media text for SemEval-2020 Task 9: SentiMix. We tackle this problem by leveraging a set of hand-engineered lexical, sentiment, and metadata fea- tures to design a classifier that can disambiguate between “positive”, “negative” and “neutral” sentiment. With this model we are able to obtain a weighted F1 score of 0.65 for the “Hinglish” task and 0.63 for the “Spanglish” tasks.