This report describes our PromptShots submissions to a shared task on Evaluating the Rationales of Amateur Investors (ERAI). We participated in both pairwise comparison and unsupervised ranking tasks. For pairwise comparison, we employed instruction-based models based on T5-small and OpenAI InstructGPT language models. Surprisingly, we observed OpenAI InstructGPT language model few-shot trained on Chinese data works best in our submissions, ranking 3rd on the maximal loss (ML) pairwise accuracy. This model works better than training on the Google translated English data by a large margin, where the English few-shot trained InstructGPT model even performs worse than an instruction-based T5-small model finetuned on the English data. However, all instruction-based submissions do not perform well on the maximal potential profit (MPP) pairwise accuracy where there are more data and learning signals. The Chinese few-shot trained InstructGPT model still performs best in our setting. For unsupervised ranking, we utilized many language models, including many financial-specific ones, and Bayesian lexicons unsupervised-learned on both Chinese and English words using a method-of-moments estimator. All our submissions rank best in the MPP ranking, from 1st to 3rd. However, they all do not perform well for ML scoring. Therefore, both MPP and ML scores need different treatments since we treated MPP and ML using the same formula. Our only difference is the treatment of market sentiment lexicons.

In this report, we describe our ClassBases submissions to a shared task on multilingual protest event detection. For the multilingual protest news detection, we participated in subtask-1, subtask-2 and subtask-4 which are document classification, sentence classification and token classification. In subtask-1, we compare XLM-RoBERTa-base, mLUKE-base and XLM-RoBERTa-large on finetuning in a sequential classification setting. We always use a combination of the training data from every language provided to train our multilingual models. We found that larger models seem to work better and entity knowledge helps but at a non-negligible cost. For subtask-2, we only submitted an mLUKE-base system for sentence classification. For subtask-4, we only submitted an XLM-RoBERTa-base for token classification system for sequence labeling. For automatically replicating manually created event datasets, we participated in COVID-related protest events from the New York Times news corpus. We created a system to process the crawled data into a dataset of protest events.

The goal of Shared Task 2 is evaluating state-of-the-art event detection systems by comparing the spatio-temporal distribution of the events they detect with existing event databases. The task focuses on some usability requirements of event detection systems in real worldscenarios. Namely, it aims to measure the ability of such a system to: (i) detect socio-political event mentions in news and social media, (ii) properly find their geographical locations, (iii) de-duplicate reports extracted from multiple sources referring to the same actual event. Building an annotated corpus for training and evaluating jointly these sub-tasks is highly time consuming. One possible way to indirectly evaluate a system’s output without an annotated corpus available is to measure its correlation with human-curated event data sets. In the last three years, the COVID-19 pandemic became motivation for restrictions and anti-pandemic measures on a world scale. This has triggered a wave of reactions and citizen actions in many countries. Shared Task 2 challenges participants to identify COVID-19 related protest actions from large unstructureddata sources both from mainstream and social media. We assess each system’s ability to model the evolution of protest events both temporally and spatially by using a number of correlation metrics with respect to a comprehensive and validated data set of COVID-related protest events (Raleigh et al., 2010).

In this report, we describe our Transformers for euphemism detection baseline (TEDB) submissions to a shared task on euphemism detection 2022. We cast the task of predicting euphemism as text classification. We considered Transformer-based models which are the current state-of-the-art methods for text classification. We explored different training schemes, pretrained models, and model architectures. Our best result of 0.816 F1-score (0.818 precision and 0.814 recall) consists of a euphemism-detection-finetuned TweetEval/TimeLMs-pretrained RoBERTa model as a feature extractor frontend with a KimCNN classifier backend trained end-to-end using a cosine annealing scheduler. We observed pretrained models on sentiment analysis and offensiveness detection to correlate with more F1-score while pretraining on other tasks, such as sarcasm detection, produces less F1-scores. Also, putting more word vector channels does not improve the performance in our experiments.

In this report, we describe our transformers for text classification baseline (TTCB) submissions to a shared task on implicit and underspecified language 2021. We cast the task of predicting revision requirements in collaboratively edited instructions as text classification. We considered transformer-based models which are the current state-of-the-art methods for text classification. We explored different training schemes, loss functions, and data augmentations. Our best result of 68.45% test accuracy (68.84% validation accuracy), however, consists of an XLNet model with a linear annealing scheduler and a cross-entropy loss. We do not observe any significant gain on any validation metric based on our various design choices except the MiniLM which has a higher validation F1 score and is faster to train by a half but also a lower validation accuracy score.

This paper presents a new task, the grounding of spatio-temporal identifying descriptions in videos. Previous work suggests potential bias in existing datasets and emphasizes the need for a new data creation schema to better model linguistic structure. We introduce a new data collection scheme based on grammatical constraints for surface realization to enable us to investigate the problem of grounding spatio-temporal identifying descriptions in videos. We then propose a two-stream modular attention network that learns and grounds spatio-temporal identifying descriptions based on appearance and motion. We show that motion modules help to ground motion-related words and also help to learn in appearance modules because modular neural networks resolve task interference between modules. Finally, we propose a future challenge and a need for a robust system arising from replacing ground truth visual annotations with automatic video object detector and temporal event localization.