dataset.py

import torch
import torch.nn.functional as F
from transformers import AutoTokenizer

from tokenizer import NLTKWordTokenizer
from lemma_rule import apply_lemma_rule


class Dataset:
    def __init__(self):
        self.word_tokenizer = NLTKWordTokenizer()
        self.subword_tokenizer = AutoTokenizer.from_pretrained("ltg/norbert3-large")

    def prepare_input(self, sentence: str):
        word_spans = list(self.word_tokenizer.span_tokenize(sentence))
        forms = [sentence[start:end] for start, end in word_spans]

        subwords, alignment = [self.subword_tokenizer.convert_tokens_to_ids("[CLS]")], [0]
        for i, word in enumerate(forms):
            space_before = (i == 0) or sentence[word_spans[i - 1][1]] == " "

            # very very ugly hack ;(
            encoding = self.subword_tokenizer(f"| {word}" if space_before else f"|{word}", add_special_tokens=False)
            subwords += encoding.input_ids[1:]
            alignment += (len(encoding.input_ids) - 1) * [i + 1]

        subwords.append(self.subword_tokenizer.convert_tokens_to_ids("[SEP]"))
        alignment.append(alignment[-1] + 1)

        subwords = torch.tensor([subwords])
        alignment = torch.tensor([alignment])
        alignment = F.one_hot(alignment, num_classes=len(forms) + 2).float()

        return forms, subwords, alignment
    
    def decode_output(self, forms, lemma_p, upos_p, xpos_p, feats_p, dep_p, ne_p, head_p):
        lemmas = [apply_lemma_rule(form, self.lemma_vocab[lemma_p[0, i, :].argmax().item()]) for i, form in enumerate(forms)]
        upos = [self.upos_vocab[upos_p[0, i, :].argmax().item()] for i in range(len(forms))]
        xpos = [self.xpos_vocab[xpos_p[0, i, :].argmax().item()] for i in range(len(forms))]
        feats = [self.feats_vocab[feats_p[0, i, :].argmax().item()] for i in range(len(forms))]
        heads = [head_p[0, i].item() for i in range(len(forms))]
        deprel = [self.arc_dep_vocab[dep_p[0, i, :].argmax().item()] for i in range(len(forms))]
        ne = [self.ne_vocab[ne_p[0, i, :].argmax().item()] for i in range(len(forms))]

        return lemmas, upos, xpos, feats, heads, deprel, ne

    # save state dict
    def state_dict(self):
        return {
            "forms_vocab": self.forms_vocab,
            "lemma_vocab": self.lemma_vocab,
            "upos_vocab": self.upos_vocab,
            "xpos_vocab": self.xpos_vocab,
            "feats_vocab": self.feats_vocab,
            "arc_dep_vocab": self.arc_dep_vocab,
            "ne_vocab": self.ne_vocab
        }

    # load state dict
    def load_state_dict(self, state_dict):
        self.forms_vocab = state_dict["forms_vocab"]
        self.lemma_vocab = state_dict["lemma_vocab"]
        self.upos_vocab = state_dict["upos_vocab"]
        self.xpos_vocab = state_dict["xpos_vocab"]
        self.feats_vocab = state_dict["feats_vocab"]
        self.arc_dep_vocab = state_dict["arc_dep_vocab"]
        self.ne_vocab = state_dict["ne_vocab"]

        self.lemma_indexer = {i: n for n, i in enumerate(self.lemma_vocab)}
        self.upos_indexer = {i: n for n, i in enumerate(self.upos_vocab)}
        self.xpos_indexer = {i: n for n, i in enumerate(self.xpos_vocab)}
        self.feats_indexer = {i: n for n, i in enumerate(self.feats_vocab)}
        self.ne_indexer = {i: n for n, i in enumerate(self.ne_vocab)}
        self.arc_dep_indexer = {i: n for n, i in enumerate(self.arc_dep_vocab)}