Pre-trained language models derive substantial linguistic and factual knowledge from the massive corpora on which they are trained, and prompt engineering seeks to align these models to specific tasks. Unfortunately, existing prompt engineering methods require significant amounts of labeled data, access to model parameters, or both. We introduce a new method for selecting prompt templates without labeled examples and without direct access to the model. Specifically, over a set of candidate templates, we choose the template that maximizes the mutual information between the input and the corresponding model output. Across 8 datasets representing 7 distinct NLP tasks, we show that when a template has high mutual information, it also has high accuracy on the task. On the largest model, selecting prompts with our method gets 90% of the way from the average prompt accuracy to the best prompt accuracy and requires no ground truth labels.

Multilingual transfer techniques often improve low-resource machine translation (MT). Many of these techniques are applied without considering data characteristics. We show in the context of Haitian-to-English translation that transfer effectiveness is correlated with amount of training data and relationships between knowledge-sharing languages. Our experiments suggest that for some languages beyond a threshold of authentic data, back-translation augmentation methods are counterproductive, while cross-lingual transfer from a sufficiently related language is preferred. We complement this finding by contributing a rule-based French-Haitian orthographic and syntactic engine and a novel method for phonological embedding. When used with multilingual techniques, orthographic transformation makes statistically significant improvements over conventional methods. And in very low-resource Jamaican MT, code-switching with a transfer language for orthographic resemblance yields a 6.63 BLEU point advantage.

A variety of NLP applications use word2vec skip-gram, GloVe, and fastText word embeddings. These models learn two sets of embedding vectors, but most practitioners use only one of them, or alternately an unweighted sum of both. This is the first study to systematically explore a range of linear combinations between the first and second embedding sets. We evaluate these combinations on a set of six NLP benchmarks including IR, POS-tagging, and sentence similarity. We show that the default embedding combinations are often suboptimal and demonstrate 1.0-8.0% improvements. Notably, GloVes default unweighted sum is its least effective combination across tasks. We provide a theoretical basis for weighting one set of embeddings more than the other according to the algorithm and task. We apply our findings to improve accuracy in applications of cross-lingual alignment and navigational knowledge by up to 15.2%.

Understanding and identifying humor has been increasingly popular, as seen by the number of datasets created to study humor. However, one area of humor research, humor generation, has remained a difficult task, with machine generated jokes failing to match human-created humor. As many humor prediction datasets claim to aid in generative tasks, we examine whether these claims are true. We focus our experiments on the most popular dataset, included in the 2020 SemEval’s Task 7, and teach our model to take normal text and “translate” it into humorous text. We evaluate our model compared to humorous human generated headlines, finding that our model is preferred equally in A/B testing with the human edited versions, a strong success for humor generation, and is preferred over an intelligent random baseline 72% of the time. We also show that our model is assumed to be human written comparable with that of the human edited headlines and is significantly better than random, indicating that this dataset does indeed provide potential for future humor generation systems.