In Achieving Open Vocabulary Neural Machine Translation with Hybrid Word-Character Models by Minh-Thang Luong and Christopher D. Manning from Stanford (PDF, ) , a model is presented that works as a normal word-based sequence-to-sequence model, as long as you feed it words in the vocabulary. When the model encounters an OOV term, the system employs a second sequence model working on character level. This model computes a representation for any word that is expressible in the given set of characters, and experimental results show that the representations computed in this way share many of the properties of neural word embeddings computed on word-level (read more on Word embeddings on Wikipedia) . The system shows large improvements in BLEU scores, especially when used with a small word vocabulary, on the task of translating between Czech and English.
It is helpful if you know a little about the ESL students' backgrounds and interests, since this will enable you to make connections to their personal lives. At the ESL placement interview the ESL teacher finds out this information and then sends it out to all concerned by e-mail. Little things can be important, such as spelling the child's name correctly and learning how to pronounce it with some accuracy. It is also helpful in class to seat ESL students with native-speakers who are sympathetic and encouraging. You can also devise group activities in which the ESL student's contribution is essential to the successful completion of the task. (See next question!)