Machine Translation

Intelligent John collected French expressions during a tour of France and needs a computer translator. The contestant prepares a sentence-pair corpus and completes a GRU-based encoder–decoder whose decoder architecture and trained weights are provided.

Task 1 (10 points): print the first 10 sentence pairs (2); implement a Lang class with name, word2index, word2count, index2word (pre-filled with SOS, EOS and PAD) and n_words starting from 3, with add_sentence and add_word (4); implement read_langs(lang1, lang2, reverse) using the given normalisation helpers (4). Task 2 (15 points, 3 each): create reversed French–English Lang objects; keep only pairs whose English side starts with one of the given eng_prefixes ("i am", "he is", "you are", ...); drop sentences longer than MAX_LENGTH = 10 words; update the vocabularies; implement prepare_data. Task 3 (6 points): compute what percentage of words makes up 80% of each corpus (3) and bar-plot the 100 most frequent English and French words (3).

Task 4 (12 points): implement the encoder from the figure (4); instantiate both models and load encoder.pth and decoder.pth (1); convert the pairs to index tensors — left-padded input with SOS/EOS, right-padded decoder input starting with SOS, and a target shifted one position left and padded to MAX_LENGTH (7). Task 5 (7 points): generate tokens with the given predict function (1) and translate the first 10 French sentences greedily, stopping at EOS (6).