There're many NLP problem relies on long documents. For example, Multi-Hop QA need to take many documents as input to answer the given question. Summarization may also need to take the whole document as input to extract the key part of the paragraph. Also, in the real word, existing data may be in the format of long sequence. A key challenge of long sequence NLP is that many popular algorithms such as RNN/LSTM and Transformers are focused on short sequence.

Eradicating hunger and malnutrition is a key development goal of the 21st century. We address the problem of optimally identifying seed varieties to reliably increase crop yield within a risk-sensitive decision-making framework. Specifically, we introduce a novel hierarchical machine learning mechanism for predicting crop yield (the yield of different seed varieties of the same crop). We integrate this prediction mechanism with a weather forecasting model, and propose three different approaches for decision making under uncertainty to select seed varieties for planting so as to balance yield maximization and risk.We apply our model to the problem of soybean variety selection given in the 2016 Syngenta Crop Challenge. Our prediction model achieves a median absolute error of 3.74 bushels per acre and thus provides good estimates for input into the decision models.Our decision models identify the selection of soybean varieties that appropriately balance yield and risk as a function of the farmer's risk aversion level. More generally, our models support farmers in decision making about which seed varieties to plant.

The advent and ubiquity of mass-market portable computational devices has opened up new opportunities for the development of assistive technologies for disabilities, especially within the domain of augmentative and alternative communications devices. Word prediction can facilitate everyday communication on mobile devices by reducing the physical interactions required to produce dialogue with them. To support personalized word prediction, a text prediction system should learn from the user’s own data to update the initial learned likelihoods that provide high quality "out of the box" performance. Within this lies an inherent trade-off: a larger corpus of initial training data can yield better default performance, but may also increase the amount of user data required for personalization of the system to be effective. We investigate a learning approach employing hierarchical modeling of phrases expected to offer sufficient "out of the box" performance relative to other learning approaches, while reducing the amount of initial training data required to facilitate on-line personalization of the text prediction system. The key insight of the proposed approach is the separation of stopwords, which primarily play syntactical roles in phrases, from keywords, which provide context and meaning in the phrase. This allows the abstraction of a phrase from an ordered list of all words to an ordered list of keywords. Thus the proposed hierarchical modeling of phrases employs two layers: keywords and stopwords. A third level abstracting keywords to a single topic is also considered, combining the power of both topic modeling and trigrams to make predictions within and between layers.