Upload 4 files

BertForJointParsing.py

@@ -0,0 +1,310 @@
+from dataclasses import dataclass
+import math
+from operator import itemgetter
+import torch
+from torch import nn
+from typing import Any, Dict, List, Literal, Optional, Tuple, Union
+from transformers import BertPreTrainedModel, BertModel, BertTokenizerFast
+from transformers.models.bert.modeling_bert import BertOnlyMLMHead
+from transformers.utils import ModelOutput
+try:
+    from .BertForSyntaxParsing import BertSyntaxParsingHead, SyntaxLabels, SyntaxLogitsOutput, parse_logits as syntax_parse_logits
+    from .BertForPrefixMarking import BertPrefixMarkingHead, parse_logits as prefix_parse_logits, encode_sentences_for_bert_for_prefix_marking
+    from .BertForMorphTagging import BertMorphTaggingHead, MorphLogitsOutput, MorphLabels, parse_logits as morph_parse_logits 
+except ImportError:
+    from BertForSyntaxParsing import BertSyntaxParsingHead, SyntaxLabels, SyntaxLogitsOutput, parse_logits as syntax_parse_logits
+    from BertForPrefixMarking import BertPrefixMarkingHead, parse_logits as prefix_parse_logits, encode_sentences_for_bert_for_prefix_marking
+    from BertForMorphTagging import BertMorphTaggingHead, MorphLogitsOutput, MorphLabels, parse_logits as morph_parse_logits 
+    
+import warnings
+
+@dataclass
+class JointParsingOutput(ModelOutput):
+    loss: Optional[torch.FloatTensor] = None
+    # logits will contain the optional predictions for the given labels
+    logits: Optional[Union[SyntaxLogitsOutput, None]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+    # if no labels are given, we will always include the syntax logits separately
+    syntax_logits: Optional[SyntaxLogitsOutput] = None
+    ner_logits: Optional[torch.FloatTensor] = None
+    prefix_logits: Optional[torch.FloatTensor] = None
+    lex_logits: Optional[torch.FloatTensor] = None
+    morph_logits: Optional[MorphLogitsOutput] = None
+
+# wrapper class to wrap a torch.nn.Module so that you can store a module in multiple linked
+# properties without registering the parameter multiple times
+class ModuleRef:
+    def __init__(self, module: torch.nn.Module):
+        self.module = module
+
+    def forward(self, *args, **kwargs):
+        return self.module.forward(*args, **kwargs)
+
+    def __call__(self, *args, **kwargs):
+        return self.module(*args, **kwargs)
+
+class BertForJointParsing(BertPreTrainedModel):
+    _tied_weights_keys = ["predictions.decoder.bias", "cls.predictions.decoder.weight"]
+
+    def __init__(self, config, do_syntax=None, do_ner=None, do_prefix=None, do_lex=None, do_morph=None, syntax_head_size=64):
+        super().__init__(config)
+
+        self.bert = BertModel(config, add_pooling_layer=False)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        # create all the heads as None, and then populate them as defined
+        self.syntax, self.ner, self.prefix, self.lex, self.morph = (None,)*5
+
+        if do_syntax is not None: 
+            config.do_syntax = do_syntax
+            config.syntax_head_size = syntax_head_size
+        if do_ner is not None: config.do_ner = do_ner
+        if do_prefix is not None: config.do_prefix = do_prefix
+        if do_lex is not None: config.do_lex = do_lex
+        if do_morph is not None: config.do_morph = do_morph
+        
+        # add all the individual heads
+        if config.do_syntax:
+            self.syntax = BertSyntaxParsingHead(config)
+        if config.do_ner:
+            self.num_labels = config.num_labels
+            self.classifier = nn.Linear(config.hidden_size, config.num_labels) # name it same as in BertForTokenClassification 
+            self.ner = ModuleRef(self.classifier)
+        if config.do_prefix:
+            self.prefix = BertPrefixMarkingHead(config)
+        if config.do_lex:
+            self.cls = BertOnlyMLMHead(config) # name it the same as in BertForMaskedLM
+            self.lex = ModuleRef(self.cls)
+        if config.do_morph:
+            self.morph = BertMorphTaggingHead(config)
+        
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder if self.lex is not None else None
+
+    def set_output_embeddings(self, new_embeddings):
+        if self.lex is not None:
+            self.cls.predictions.decoder = new_embeddings
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        prefix_class_id_options: Optional[torch.Tensor] = None,
+        labels: Optional[Union[SyntaxLabels, MorphLabels, torch.Tensor]] = None,
+        labels_type: Optional[Literal['syntax', 'ner', 'prefix', 'lex', 'morph']] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        compute_syntax_mst: Optional[bool] = None
+    ):
+        if return_dict is False:
+            warnings.warn("Specified `return_dict=False` but the flag is ignored and treated as always True in this model.")
+        
+        if labels is not None and labels_type is None:
+            raise ValueError("Cannot specify labels without labels_type")
+        
+        if labels_type == 'seg' and prefix_class_id_options is None:
+            raise ValueError('Cannot calculate prefix logits without prefix_class_id_options')
+        
+        if compute_syntax_mst is not None and self.syntax is None:
+            raise ValueError("Cannot compute syntax MST when the syntax head isn't loaded")
+
+
+        bert_outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=True,
+        )
+        
+        # calculate the extended attention mask for any child that might need it
+        extended_attention_mask = None
+        if attention_mask is not None:
+            extended_attention_mask = self.get_extended_attention_mask(attention_mask, input_ids.size())
+
+        # extract the hidden states, and apply the dropout
+        hidden_states = self.dropout(bert_outputs[0])
+
+        logits = None    
+        syntax_logits = None
+        ner_logits = None
+        prefix_logits = None
+        lex_logits = None
+        morph_logits = None
+
+        # Calculate the syntax
+        if self.syntax is not None and (labels is None or labels_type == 'syntax'):
+            # apply the syntax head
+            loss, syntax_logits = self.syntax(hidden_states, extended_attention_mask, labels, compute_syntax_mst)
+            logits = syntax_logits
+
+        # Calculate the NER
+        if self.ner is not None and (labels is None or labels_type == 'ner'):
+            ner_logits = self.ner(hidden_states)
+            logits = ner_logits
+            if labels is not None:
+                loss_fct = nn.CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        # Calculate the segmentation
+        if self.prefix is not None and (labels is None or labels_type == 'prefix'):
+            loss, prefix_logits = self.prefix(hidden_states, prefix_class_id_options, labels)
+            logits = prefix_logits
+        
+        # Calculate the lexeme
+        if self.lex is not None and (labels is None or labels_type == 'lex'):
+            lex_logits = self.lex(hidden_states)
+            logits = lex_logits
+            if labels is not None:
+                loss_fct = nn.CrossEntropyLoss()  # -100 index = padding token
+                loss = loss_fct(lex_logits.view(-1, self.config.vocab_size), labels.view(-1))
+
+        if self.morph is not None and (labels is None or labels_type == 'morph'):
+            loss, morph_logits = self.morph(hidden_states, labels)
+            logits = morph_logits
+
+        # no labels => logits = None
+        if labels is None: logits = None
+
+        return JointParsingOutput(
+            loss,
+            logits,
+            hidden_states=bert_outputs.hidden_states,
+            attentions=bert_outputs.attentions,
+            # all the predicted logits section
+            syntax_logits=syntax_logits,
+            ner_logits=ner_logits,
+            prefix_logits=prefix_logits,
+            lex_logits=lex_logits,
+            morph_logits=morph_logits
+        )
+
+    def predict(self, sentences: Union[str, List[str]], tokenizer: BertTokenizerFast, padding='longest', truncation=True, compute_syntax_mst=True, per_token_ner=False):
+        is_single_sentence = isinstance(sentences, str)
+        if is_single_sentence:
+            sentences = [sentences]
+            
+        # predict the logits for the sentence
+        if self.prefix is not None:
+            inputs = encode_sentences_for_bert_for_prefix_marking(tokenizer, sentences, padding)
+        else:
+            inputs = tokenizer(sentences, padding=padding, truncation=truncation, return_tensors='pt')
+            
+        # Copy the tensors to the right device, and parse!
+        inputs = {k:v.to(self.device) for k,v in inputs.items()}
+        output = self.forward(**inputs, return_dict=True, compute_syntax_mst=compute_syntax_mst)
+        
+        final_output = [dict(text=sentence, tokens=[dict(token=t) for t in combine_token_wordpieces(ids, tokenizer)]) for sentence, ids in zip(sentences, inputs['input_ids'])]
+        # Syntax logits: each sentence gets a dict(tree: List[dict(word,dep_head,dep_head_idx,dep_func)], root_idx: int)
+        if output.syntax_logits is not None:
+            for sent_idx,parsed in enumerate(syntax_parse_logits(inputs, sentences, tokenizer, output.syntax_logits)):
+                merge_token_list(final_output[sent_idx]['tokens'], parsed['tree'], 'syntax')
+                final_output[sent_idx]['root_idx'] = parsed['root_idx']
+                
+        # Prefix logits: each sentence gets a list([prefix_segment, word_without_prefix]) - **WITH CLS & SEP**
+        if output.prefix_logits is not None:
+            for sent_idx,parsed in enumerate(prefix_parse_logits(inputs, sentences, tokenizer, output.prefix_logits)):
+                merge_token_list(final_output[sent_idx]['tokens'], map(tuple, parsed[1:-1]), 'seg')
+            
+        # Lex logits each sentence gets a list(tuple(word, lexeme))
+        if output.lex_logits is not None:
+            for sent_idx, parsed in enumerate(lex_parse_logits(inputs, sentences, tokenizer, output.lex_logits)):
+                merge_token_list(final_output[sent_idx]['tokens'], map(itemgetter(1), parsed), 'lex')
+                
+        # morph logits each sentences get a dict(text=str, tokens=list(dict(token, pos, feats, prefixes, suffix, suffix_feats?)))
+        if output.morph_logits is not None:
+            for sent_idx,parsed in enumerate(morph_parse_logits(inputs, sentences, tokenizer, output.morph_logits)):
+                merge_token_list(final_output[sent_idx]['tokens'], parsed['tokens'], 'morph')
+            
+        # NER logits each sentence gets a list(tuple(word, ner))
+        if output.ner_logits is not None:
+            for sent_idx,parsed in enumerate(ner_parse_logits(inputs, sentences, tokenizer, output.ner_logits, self.config.id2label)):
+                if per_token_ner:
+                    merge_token_list(final_output[sent_idx]['tokens'], map(itemgetter(1), parsed), 'ner')
+                final_output[sent_idx]['ner_entities'] = aggregate_ner_tokens(parsed) 
+        
+        if is_single_sentence:
+            final_output = final_output[0]
+        return final_output
+
+def aggregate_ner_tokens(predictions):
+    entities = []
+    prev = None
+    for word,pred in predictions:
+        # O does nothing
+        if pred == 'O': prev = None
+        # B- || I-entity != prev (different entity or none)
+        elif pred.startswith('B-') or pred[2:] != prev:
+            prev = pred[2:]
+            entities.append(([word], prev))
+        else: entities[-1][0].append(word)
+    
+    return [dict(phrase=' '.join(words), label=label) for words,label in entities]
+        
+
+def merge_token_list(src, update, key):
+    for token_src, token_update in zip(src, update):
+        token_src[key] = token_update
+        
+def combine_token_wordpieces(input_ids: torch.Tensor, tokenizer: BertTokenizerFast):
+    ret = []
+    for token in tokenizer.convert_ids_to_tokens(input_ids):
+        if token in [tokenizer.cls_token, tokenizer.sep_token, tokenizer.pad_token]: continue
+        if token.startswith('##'):
+            ret[-1] += token[2:]
+        else: ret.append(token)
+    return ret
+
+def ner_parse_logits(inputs: Dict[str, torch.Tensor], sentences: List[str], tokenizer: BertTokenizerFast, logits: torch.Tensor, id2label: Dict[int, str]):
+    input_ids = inputs['input_ids']
+    
+    predictions = torch.argmax(logits, dim=-1)
+    batch_ret = []
+    for batch_idx in range(len(sentences)):
+        ret = []
+        batch_ret.append(ret)
+        for tok_idx in range(input_ids.shape[1]):
+            token_id = input_ids[batch_idx, tok_idx]
+            # ignore cls, sep, pad
+            if token_id in [tokenizer.cls_token_id, tokenizer.sep_token_id, tokenizer.pad_token_id]: continue
+
+            token = tokenizer._convert_id_to_token(token_id)
+            # wordpieces should just be appended to the previous word
+            if token.startswith('##'):
+                ret[-1] = (ret[-1][0] + token[2:], ret[-1][1])
+                continue
+            ret.append((token, id2label[predictions[batch_idx, tok_idx].item()]))
+    return batch_ret
+
+def lex_parse_logits(inputs: Dict[str, torch.Tensor], sentences: List[str], tokenizer: BertTokenizerFast, logits: torch.Tensor):
+    input_ids = inputs['input_ids']
+    
+    predictions = torch.argmax(logits, dim=-1)
+    batch_ret = []
+    for batch_idx in range(len(sentences)):
+        ret = []
+        batch_ret.append(ret)
+        for tok_idx in range(input_ids.shape[1]):
+            token_id = input_ids[batch_idx, tok_idx]
+            # ignore cls, sep, pad
+            if token_id in [tokenizer.cls_token_id, tokenizer.sep_token_id, tokenizer.pad_token_id]: continue
+
+            token = tokenizer._convert_id_to_token(token_id)
+            # wordpieces should just be appended to the previous word
+            if token.startswith('##'):
+                ret[-1] = (ret[-1][0] + token[2:], ret[-1][1])
+                continue
+            ret.append((token, tokenizer._convert_id_to_token(predictions[batch_idx, tok_idx])))
+    return batch_ret

BertForMorphTagging.py

@@ -0,0 +1,212 @@
+from collections import OrderedDict
+from operator import itemgetter
+from transformers.utils import ModelOutput
+import torch
+from torch import nn
+from typing import Dict, List, Tuple, Optional
+from dataclasses import dataclass
+from transformers import BertPreTrainedModel, BertModel, BertTokenizerFast
+
+ALL_POS = ['DET', 'NOUN', 'VERB', 'CCONJ', 'ADP', 'PRON', 'PUNCT', 'ADJ', 'ADV', 'SCONJ', 'NUM', 'PROPN', 'AUX', 'X', 'INTJ', 'SYM']
+ALL_PREFIX_POS = ['SCONJ', 'DET', 'ADV', 'CCONJ', 'ADP', 'NUM']
+ALL_SUFFIX_POS = ['none', 'ADP_PRON', 'PRON']
+ALL_FEATURES = [
+    ('Gender', ['none', 'Masc', 'Fem', 'Fem,Masc']),
+    ('Number', ['none', 'Sing', 'Plur', 'Plur,Sing', 'Dual', 'Dual,Plur']),
+    ('Person', ['none', '1', '2', '3', '1,2,3']),
+    ('Tense', ['none', 'Past', 'Fut', 'Pres', 'Imp'])
+]
+
+@dataclass
+class MorphLogitsOutput(ModelOutput):
+    prefix_logits: torch.FloatTensor = None
+    pos_logits: torch.FloatTensor = None
+    features_logits: List[torch.FloatTensor] = None
+    suffix_logits: torch.FloatTensor = None
+    suffix_features_logits: List[torch.FloatTensor] = None
+
+    def detach(self):
+        return MorphLogitsOutput(self.prefix_logits.detach(), self.pos_logits.detach(), [logits.deatch() for logits in self.features_logits], self.suffix_logits.detach(), [logits.deatch() for logits in self.suffix_features_logits])
+
+
+@dataclass
+class MorphTaggingOutput(ModelOutput):
+    loss: Optional[torch.FloatTensor] = None
+    logits: Optional[MorphLogitsOutput] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+@dataclass
+class MorphLabels(ModelOutput):
+    prefix_labels: Optional[torch.FloatTensor] = None
+    pos_labels: Optional[torch.FloatTensor] = None
+    features_labels: Optional[List[torch.FloatTensor]] = None
+    suffix_labels: Optional[torch.FloatTensor] = None
+    suffix_features_labels: Optional[List[torch.FloatTensor]] = None
+
+    def detach(self):
+        return MorphLabels(self.prefix_labels.detach(), self.pos_labels.detach(), [labels.detach() for labels in self.features_labels], self.suffix_labels.detach(), [labels.detach() for labels in self.suffix_features_labels])
+    
+    def to(self, device):
+        return MorphLabels(self.prefix_labels.to(device), self.pos_labels.to(device), [feat.to(device) for feat in self.features_labels], self.suffix_labels.to(device), [feat.to(device) for feat in self.suffix_features_labels])
+
+class BertMorphTaggingHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+
+        self.num_prefix_classes = len(ALL_PREFIX_POS)
+        self.num_pos_classes = len(ALL_POS)
+        self.num_suffix_classes = len(ALL_SUFFIX_POS)
+        self.num_features_classes = list(map(len, map(itemgetter(1), ALL_FEATURES)))
+        # we need a classifier for prefix cls and POS cls
+        # the prefix will use BCEWithLogits for multiple labels cls
+        self.prefix_cls = nn.Linear(config.hidden_size, self.num_prefix_classes)
+        # and pos + feats will use good old cross entropy for single label
+        self.pos_cls = nn.Linear(config.hidden_size, self.num_pos_classes)
+        self.features_cls = nn.ModuleList([nn.Linear(config.hidden_size, len(features)) for _, features in ALL_FEATURES])
+        # and suffix + feats will also be cross entropy
+        self.suffix_cls = nn.Linear(config.hidden_size, self.num_suffix_classes)
+        self.suffix_features_cls = nn.ModuleList([nn.Linear(config.hidden_size, len(features)) for _, features in ALL_FEATURES])
+
+    def forward(
+            self, 
+            hidden_states: torch.Tensor,
+            labels: Optional[MorphLabels] = None):
+        # run each of the classifiers on the transformed output
+        prefix_logits = self.prefix_cls(hidden_states)
+        pos_logits = self.pos_cls(hidden_states)
+        suffix_logits = self.suffix_cls(hidden_states)
+        features_logits = [cls(hidden_states) for cls in self.features_cls]
+        suffix_features_logits = [cls(hidden_states) for cls in self.suffix_features_cls]
+
+        loss = None
+        if labels is not None:
+            # step 1: prefix labels loss
+            loss_fct = nn.BCEWithLogitsLoss(weight=(labels.prefix_labels != -100).float())
+            loss = loss_fct(prefix_logits, labels.prefix_labels)
+            # step 2: pos labels loss
+            loss_fct = nn.CrossEntropyLoss()
+            loss += loss_fct(pos_logits.view(-1, self.num_pos_classes), labels.pos_labels.view(-1))
+            # step 2b: features
+            for feat_logits,feat_labels,num_features in zip(features_logits, labels.features_labels, self.num_features_classes):
+                loss += loss_fct(feat_logits.view(-1, num_features), feat_labels.view(-1))
+            # step 3: suffix logits loss
+            loss += loss_fct(suffix_logits.view(-1, self.num_suffix_classes), labels.suffix_labels.view(-1))
+            # step 3b: suffix features
+            for feat_logits,feat_labels,num_features in zip(suffix_features_logits, labels.suffix_features_labels, self.num_features_classes):
+                loss += loss_fct(feat_logits.view(-1, num_features), feat_labels.view(-1))
+
+        return loss, MorphLogitsOutput(prefix_logits, pos_logits, features_logits, suffix_logits, suffix_features_logits)
+
+class BertForMorphTagging(BertPreTrainedModel):
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.bert = BertModel(config, add_pooling_layer=False)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.morph = BertMorphTaggingHead(config)
+        
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        labels: Optional[MorphLabels] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ):
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        bert_outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = bert_outputs[0]
+        hidden_states = self.dropout(hidden_states)
+
+        loss, logits = self.morph(hidden_states, labels)
+        
+        if not return_dict:
+            return (loss,logits) + bert_outputs[2:]
+        
+        return MorphTaggingOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=bert_outputs.hidden_states,
+            attentions=bert_outputs.attentions,
+        )
+
+    def predict(self, sentences: List[str], tokenizer: BertTokenizerFast, padding='longest'):
+        # tokenize the inputs and convert them to relevant device
+        inputs = tokenizer(sentences, padding=padding, truncation=True, return_tensors='pt')
+        inputs = {k:v.to(self.device) for k,v in inputs.items()}
+        # calculate the logits
+        logits = self.forward(**inputs, return_dict=True).logits
+        return parse_logits(inputs, sentences, tokenizer, logits)
+    
+def parse_logits(inputs: Dict[str, torch.Tensor], sentences: List[str], tokenizer: BertTokenizerFast, logits: MorphLogitsOutput):
+    prefix_logits, pos_logits, feats_logits, suffix_logits, suffix_feats_logits = \
+                logits.prefix_logits, logits.pos_logits, logits.features_logits, logits.suffix_logits, logits.suffix_features_logits
+
+    prefix_predictions = (prefix_logits > 0.5).int() # Threshold at 0.5 for multi-label classification
+    pos_predictions = pos_logits.argmax(axis=-1)
+    suffix_predictions = suffix_logits.argmax(axis=-1)
+    feats_predictions = [logits.argmax(axis=-1) for logits in feats_logits]
+    suffix_feats_predictions = [logits.argmax(axis=-1) for logits in suffix_feats_logits]
+
+    # create the return dictionary 
+    # for each sentence, return a dict object with the following files { text, tokens }
+    # Where tokens is a list of dicts, where each dict is: 
+    #       { pos: str, feats: dict, prefixes: List[str], suffix: str | bool, suffix_feats: dict | None}
+    special_tokens = set([tokenizer.pad_token, tokenizer.cls_token, tokenizer.sep_token])
+    ret = []
+    for sent_idx,sentence in enumerate(sentences):
+        input_id_strs = tokenizer.convert_ids_to_tokens(inputs['input_ids'][sent_idx])
+        # iterate through each token in the sentence, ignoring special tokens
+        tokens = []
+        for token_idx,token_str in enumerate(input_id_strs):
+            if not token_str in special_tokens: 
+                if token_str.startswith('##'):
+                    tokens[-1]['token'] += token_str[2:]
+                    continue
+                tokens.append(dict(
+                    token=token_str,
+                    pos=ALL_POS[pos_predictions[sent_idx, token_idx]],
+                    feats=get_features_dict_from_predictions(feats_predictions, (sent_idx, token_idx)),
+                    prefixes=[ALL_PREFIX_POS[idx] for idx,i in enumerate(prefix_predictions[sent_idx, token_idx]) if i > 0],
+                    suffix=get_suffix_or_false(ALL_SUFFIX_POS[suffix_predictions[sent_idx, token_idx]]),
+                ))
+                if tokens[-1]['suffix']:
+                    tokens[-1]['suffix_feats'] = get_features_dict_from_predictions(suffix_feats_predictions, (sent_idx, token_idx))
+        ret.append(dict(text=sentence, tokens=tokens))
+    return ret
+    
+def get_suffix_or_false(suffix):
+    return False if suffix == 'none' else suffix
+
+def get_features_dict_from_predictions(predictions, idx):
+    ret = {}
+    for (feat_idx, (feat_name, feat_values)) in enumerate(ALL_FEATURES):
+        val = feat_values[predictions[feat_idx][idx]]
+        if val != 'none':
+            ret[feat_name] = val
+    return ret
+
+

BertForPrefixMarking.py

@@ -0,0 +1,248 @@
+from transformers.utils import ModelOutput
+import torch
+from torch import nn
+from typing import Dict, List, Tuple, Optional
+from dataclasses import dataclass
+from transformers import BertPreTrainedModel, BertModel, BertTokenizerFast
+
+# define the classes, and the possible prefixes for each class
+POSSIBLE_PREFIX_CLASSES =  [ ['לכש', 'כש', 'מש', 'בש', 'לש'], ['מ'], ['ש'], ['ה'], ['ו'], ['כ'], ['ל'], ['ב'] ]
+# map each individual prefix to it's class number
+PREFIXES_TO_CLASS = {w:i for i,l in enumerate(POSSIBLE_PREFIX_CLASSES) for w in l}
+# keep a list of all the prefixes, sorted by length, so that we can decompose
+# a given prefixes and figure out the classes
+ALL_PREFIX_ITEMS = list(sorted(PREFIXES_TO_CLASS.keys(), key=len, reverse=True))
+TOTAL_POSSIBLE_PREFIX_CLASSES = len(POSSIBLE_PREFIX_CLASSES)    
+
+def get_prefixes_from_str(s, greedy=False):
+    # keep trimming prefixes from the string
+    while len(s) > 0 and s[0] in PREFIXES_TO_CLASS:
+        # find the longest string to trim
+        next_pre = next((pre for pre in ALL_PREFIX_ITEMS if s.startswith(pre)), None)
+        if next_pre is None:
+            return
+        yield next_pre
+        # if the chosen prefix is more than one letter, there is always an option that the 
+        # prefix is actually just the first letter of the prefix - so offer that up as a valid prefix
+        # as well. We will still jump to the length of the longer one, since if the next two/three
+        # letters are a prefix, they have to be the longest one
+        if not greedy and len(next_pre) > 1:
+            yield next_pre[0]
+        s = s[len(next_pre):]
+
+def get_prefix_classes_from_str(s, greedy=False):
+    for pre in get_prefixes_from_str(s, greedy):
+        yield PREFIXES_TO_CLASS[pre]
+
+@dataclass
+class PrefixesClassifiersOutput(ModelOutput):
+    loss: Optional[torch.FloatTensor] = None
+    logits: Optional[torch.FloatTensor] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+class BertPrefixMarkingHead(nn.Module):
+    def __init__(self, config) -> None:
+        super().__init__()
+        self.config = config
+
+        # an embedding table containing an embedding for each prefix class + 1 for NONE
+        # we will concatenate either the embedding/NONE for each class - and we want the concatenate
+        # size to be the hidden_size
+        prefix_class_embed = config.hidden_size // TOTAL_POSSIBLE_PREFIX_CLASSES
+        self.prefix_class_embeddings = nn.Embedding(TOTAL_POSSIBLE_PREFIX_CLASSES + 1, prefix_class_embed)
+        
+        # one layer for transformation, apply an activation, then another N classifiers for each prefix class
+        self.transform = nn.Linear(config.hidden_size + prefix_class_embed * TOTAL_POSSIBLE_PREFIX_CLASSES, config.hidden_size)
+        self.activation = nn.Tanh()
+        self.classifiers = nn.ModuleList([nn.Linear(config.hidden_size, 2) for _ in range(TOTAL_POSSIBLE_PREFIX_CLASSES)])
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            prefix_class_id_options: torch.Tensor,
+            labels: Optional[torch.Tensor] = None) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
+        
+        # encode the prefix_class_id_options
+        # If input_ids is batch x seq_len
+        # Then sequence_output is batch x seq_len x hidden_dim
+        # So prefix_class_id_options is batch x seq_len x TOTAL_POSSIBLE_PREFIX_CLASSES
+        # Looking up the embeddings should give us batch x seq_len x TOTAL_POSSIBLE_PREFIX_CLASSES x hidden_dim / N
+        possible_class_embed = self.prefix_class_embeddings(prefix_class_id_options)
+        # then flatten the final dimension - now we have batch x seq_len x hidden_dim_2
+        possible_class_embed = possible_class_embed.reshape(possible_class_embed.shape[:-2] + (-1,))
+
+        # concatenate the new class embed into the sequence output before the transform
+        pre_transform_output = torch.cat((hidden_states, possible_class_embed), dim=-1) # batch x seq_len x (hidden_dim + hidden_dim_2)
+        pre_logits_output = self.activation(self.transform(pre_transform_output))# batch x seq_len x hidden_dim
+
+        # run each of the classifiers on the transformed output
+        logits = torch.cat([cls(pre_logits_output).unsqueeze(-2) for cls in self.classifiers], dim=-2)
+
+        loss = None
+        if labels is not None:
+            loss_fct = nn.CrossEntropyLoss()
+            loss = loss_fct(logits.view(-1, 2), labels.view(-1))
+        
+        return (loss, logits)
+        
+
+
+class BertForPrefixMarking(BertPreTrainedModel):
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.bert = BertModel(config, add_pooling_layer=False)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.prefix = BertPrefixMarkingHead(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        prefix_class_id_options: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ):
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        bert_outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = bert_outputs[0]
+        hidden_states = self.dropout(hidden_states)
+
+        loss, logits = self.prefix.forward(hidden_states, prefix_class_id_options, labels)
+        if not return_dict:
+            return (loss,logits,) + bert_outputs[2:]
+
+        return PrefixesClassifiersOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=bert_outputs.hidden_states,
+            attentions=bert_outputs.attentions,
+        )
+    
+    def predict(self, sentences: List[str], tokenizer: BertTokenizerFast, padding='longest'):
+        # step 1: encode the sentences through using the tokenizer, and get the input tensors + prefix id tensors
+        inputs = encode_sentences_for_bert_for_prefix_marking(tokenizer, sentences, padding)
+        inputs = {k:v.to(self.device) for k,v in inputs.items()}
+
+        # run through bert
+        logits = self.forward(**inputs, return_dict=True).logits
+        return parse_logits(inputs, sentences, tokenizer, logits)
+
+def parse_logits(inputs: Dict[str, torch.Tensor], sentences: List[str], tokenizer: BertTokenizerFast, logits: torch.FloatTensor):
+    # extract the predictions by argmaxing the final dimension (batch x sequence x prefixes x prediction)
+    logit_preds = torch.argmax(logits, axis=3)
+
+    ret = []
+
+    for sent_idx,sent_ids in enumerate(inputs['input_ids']):
+        tokens = tokenizer.convert_ids_to_tokens(sent_ids)
+        ret.append([])
+        for tok_idx,token in enumerate(tokens):
+            # If we've reached the pad token, then we are at the end
+            if token == tokenizer.pad_token: continue
+            if token.startswith('##'): continue
+
+            # combine the next tokens in? only if it's a breakup
+            next_tok_idx = tok_idx + 1
+            while next_tok_idx < len(tokens) and tokens[next_tok_idx].startswith('##'):
+                token += tokens[next_tok_idx][2:]
+                next_tok_idx += 1
+
+            prefix_len = get_predicted_prefix_len_from_logits(token, logit_preds[sent_idx, tok_idx])
+        
+            if not prefix_len:
+                ret[-1].append([token])
+            else:
+                ret[-1].append([token[:prefix_len], token[prefix_len:]])
+    return ret
+
+def encode_sentences_for_bert_for_prefix_marking(tokenizer: BertTokenizerFast, sentences: List[str], padding='longest', truncation=True):
+    inputs = tokenizer(sentences, padding=padding, truncation=truncation, return_tensors='pt')
+
+    # create our prefix_id_options array which will be like the input ids shape but with an addtional
+    # dimension containing for each prefix whether it can be for that word
+    prefix_id_options = torch.full(inputs['input_ids'].shape + (TOTAL_POSSIBLE_PREFIX_CLASSES,), TOTAL_POSSIBLE_PREFIX_CLASSES, dtype=torch.long)
+
+    # go through each token, and fill in the vector accordingly
+    for sent_idx, sent_ids in enumerate(inputs['input_ids']):
+        tokens = tokenizer.convert_ids_to_tokens(sent_ids)
+        for tok_idx, token in enumerate(tokens):
+            # if the first letter isn't a valid prefix letter, nothing to talk about
+            if len(token) < 2 or not token[0] in PREFIXES_TO_CLASS: continue
+
+            # combine the next tokens in? only if it's a breakup
+            next_tok_idx = tok_idx + 1
+            while next_tok_idx < len(tokens) and tokens[next_tok_idx].startswith('##'):
+                token += tokens[next_tok_idx][2:]
+                next_tok_idx += 1
+
+            # find all the possible prefixes - and mark them as 0 (and in the possible mark it as it's value for embed lookup)
+            for pre_class in get_prefix_classes_from_str(token):
+                prefix_id_options[sent_idx, tok_idx, pre_class] = pre_class
+        
+    inputs['prefix_class_id_options'] = prefix_id_options
+    return inputs
+
+def get_predicted_prefix_len_from_logits(token, token_logits):
+    # Go through each possible prefix, and check if the prefix is yes - and if
+    # so increase the counter of the matched length, otherwise break out. That will solve cases
+    # of predicting prefix combinations that don't exist on the word.
+    # For example, if we have the word ושכשהלכתי and the model predict ו & כש, then we will only
+    # take the vuv because in order to get the כש we need the ש as well.
+    # Two extra items:
+    # 1] Don't allow the same prefix multiple times
+    # 2] Always check that the word starts with that prefix - otherwise it's bad 
+    #    (except for the case of multi-letter prefix, where we force the next to be last)
+    cur_len, skip_next, last_check, seen_prefixes = 0, False, False, set()
+    for prefix in get_prefixes_from_str(token):
+        # Are we skipping this prefix? This will be the case where we matched כש, don't allow ש
+        if skip_next:
+            skip_next = False
+            continue
+        # check for duplicate prefixes, we don't allow two of the same prefix
+        # if it predicted two of the same, then we will break out
+        if prefix in seen_prefixes: break
+        seen_prefixes.add(prefix)
+
+        # check if we predicted this prefix
+        if token_logits[PREFIXES_TO_CLASS[prefix]].item():
+            cur_len += len(prefix)
+            if last_check: break
+            skip_next = len(prefix) > 1
+        # Otherwise, we predicted no. If we didn't, then this is the end of the prefix
+        # and time to break out. *Except* if it's a multi letter prefix, then we allow
+        # just the next letter - e.g., if כש doesn't match, then we allow כ, but then we know
+        # the word continues with a ש, and if it's not כש, then it's not כ-ש- (invalid)
+        elif len(prefix) > 1:
+            last_check = True
+        else:
+            break
+
+    return cur_len

BertForSyntaxParsing.py

@@ -0,0 +1,285 @@
+import math
+from transformers.utils import ModelOutput
+import torch
+from torch import nn
+from typing import Dict, List, Tuple, Optional, Union
+from dataclasses import dataclass
+from transformers import BertPreTrainedModel, BertModel, BertTokenizerFast
+
+ALL_FUNCTION_LABELS = ["nsubj", "punct", "mark", "case", "fixed", "obl", "det", "amod", "acl:relcl", "nmod", "cc", "conj", "root", "compound", "cop", "compound:affix", "advmod", "nummod", "appos", "nsubj:pass", "nmod:poss", "xcomp", "obj", "aux", "parataxis", "advcl", "ccomp", "csubj", "acl", "obl:tmod", "csubj:pass", "dep", "dislocated", "nmod:tmod", "nmod:npmod", "flat", "obl:npmod", "goeswith", "reparandum", "orphan", "list", "discourse", "iobj", "vocative", "expl", "flat:name"]
+
+@dataclass
+class SyntaxLogitsOutput(ModelOutput):
+    dependency_logits: torch.FloatTensor = None
+    function_logits: torch.FloatTensor = None
+    dependency_head_indices: torch.LongTensor = None
+
+    def detach(self):
+        return SyntaxTaggingOutput(self.dependency_logits.detach(), self.function_logits.detach(), self.dependency_head_indices.detach())
+
+@dataclass
+class SyntaxTaggingOutput(ModelOutput):
+    loss: Optional[torch.FloatTensor] = None
+    logits: Optional[SyntaxLogitsOutput] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+@dataclass
+class SyntaxLabels(ModelOutput):
+    dependency_labels: Optional[torch.LongTensor] = None
+    function_labels: Optional[torch.LongTensor] = None
+
+    def detach(self):
+        return SyntaxLabels(self.dependency_labels.detach(), self.function_labels.detach())
+    
+    def to(self, device):
+        return SyntaxLabels(self.dependency_labels.to(device), self.function_labels.to(device))
+
+class BertSyntaxParsingHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+
+        # the attention query & key values
+        self.head_size = config.syntax_head_size# int(config.hidden_size / config.num_attention_heads * 2)
+        self.query = nn.Linear(config.hidden_size, self.head_size)
+        self.key = nn.Linear(config.hidden_size, self.head_size)
+        # the function classifier gets two encoding values and predicts the labels
+        self.num_function_classes = len(ALL_FUNCTION_LABELS)
+        self.cls = nn.Linear(config.hidden_size * 2, self.num_function_classes)
+
+    def forward(
+            self, 
+            hidden_states: torch.Tensor, 
+            extended_attention_mask: Optional[torch.Tensor],
+            labels: Optional[SyntaxLabels] = None,
+            compute_mst: bool = False) -> Tuple[torch.Tensor, SyntaxLogitsOutput]:
+        
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        query_layer = self.query(hidden_states)
+        key_layer = self.key(hidden_states)
+        attention_scores = torch.bmm(query_layer, key_layer.transpose(-1, -2)) / math.sqrt(self.head_size)
+
+        # add in the attention mask
+        if extended_attention_mask is not None:
+            if extended_attention_mask.ndim == 4:
+                extended_attention_mask = extended_attention_mask.squeeze(1)
+            attention_scores += extended_attention_mask# batch x seq x seq
+
+        # At this point take the hidden_state of the word and of the dependency word, and predict the function
+        # If labels are provided, use the labels.
+        if self.training and labels is not None:
+            # Note that the labels can have -100, so just set those to zero with a max
+            dep_indices = labels.dependency_labels.clamp_min(0)
+        # Otherwise - check if he wants the MST or just the argmax
+        elif compute_mst:
+            dep_indices = compute_mst_tree(attention_scores)
+        else: 
+            dep_indices = torch.argmax(attention_scores, dim=-1)
+        
+        # After we retrieved the dependency indicies, create a tensor of teh batch indices, and and retrieve the vectors of the heads to calculate the function
+        batch_indices = torch.arange(dep_indices.size(0)).view(-1, 1).expand(-1, dep_indices.size(1)).to(dep_indices.device)
+        dep_vectors = hidden_states[batch_indices, dep_indices, :] # batch x seq x dim
+
+        # concatenate that with the last hidden states, and send to the classifier output
+        cls_inputs = torch.cat((hidden_states, dep_vectors), dim=-1)
+        function_logits = self.cls(cls_inputs)
+
+        loss = None
+        if labels is not None:
+            loss_fct = nn.CrossEntropyLoss()
+            # step 1: dependency scores loss - this is applied to the attention scores
+            loss = loss_fct(attention_scores.view(-1, hidden_states.size(-2)), labels.dependency_labels.view(-1))
+            # step 2: function loss
+            loss += loss_fct(function_logits.view(-1, self.num_function_classes), labels.function_labels.view(-1))
+        
+        return (loss, SyntaxLogitsOutput(attention_scores, function_logits, dep_indices))
+
+
+class BertForSyntaxParsing(BertPreTrainedModel):
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.bert = BertModel(config, add_pooling_layer=False)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.syntax = BertSyntaxParsingHead(config)
+        
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        labels: Optional[SyntaxLabels] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        compute_syntax_mst: Optional[bool] = None,
+    ):
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        bert_outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        
+        extended_attention_mask = None
+        if attention_mask is not None:
+            extended_attention_mask = self.get_extended_attention_mask(attention_mask, input_ids.size())
+        # apply the syntax head
+        loss, logits = self.syntax(self.dropout(bert_outputs[0]), extended_attention_mask, labels, compute_syntax_mst)
+        
+        if not return_dict:
+            return (loss,(logits.dependency_logits, logits.function_logits)) + bert_outputs[2:]
+        
+        return SyntaxTaggingOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=bert_outputs.hidden_states,
+            attentions=bert_outputs.attentions,
+        )
+
+    def predict(self, sentences: Union[str, List[str]], tokenizer: BertTokenizerFast, compute_mst=True):
+        if isinstance(sentences, str):
+            sentences = [sentences]
+            
+        # predict the logits for the sentence
+        inputs = tokenizer(sentences, padding='longest', truncation=True, return_tensors='pt')
+        inputs = {k:v.to(self.device) for k,v in inputs.items()}
+        logits = self.forward(**inputs, return_dict=True, compute_syntax_mst=compute_mst).logits
+        return parse_logits(inputs, sentences, tokenizer, logits)
+
+def parse_logits(inputs: Dict[str, torch.Tensor], sentences: List[str], tokenizer: BertTokenizerFast, logits: SyntaxLogitsOutput):
+    outputs = []        
+    for i in range(len(sentences)):
+        deps = logits.dependency_head_indices[i].tolist()
+        funcs = logits.function_logits.argmax(-1)[i].tolist()
+        toks = tokenizer.convert_ids_to_tokens(inputs['input_ids'][i])[1:-1] # ignore cls and sep
+        
+        # first, go through the tokens and create a mapping between each dependency index and the index without wordpieces
+        # wordpieces. At the same time, append the wordpieces in
+        idx_mapping = {-1:-1} # default root 
+        real_idx = -1
+        for i in range(len(toks)):
+            if not toks[i].startswith('##'):
+                real_idx += 1
+            idx_mapping[i] = real_idx
+            
+        # build our tree, keeping tracking of the root idx
+        tree = []
+        root_idx = 0
+        for i in range(len(toks)):
+            if toks[i].startswith('##'):
+                tree[-1]['word'] += toks[i][2:]
+                continue 
+            
+            dep_idx = deps[i + 1] - 1 # increase 1 for cls, decrease 1 for cls
+            dep_head = 'root' if dep_idx == -1 else toks[dep_idx]
+            dep_func = ALL_FUNCTION_LABELS[funcs[i + 1]]
+
+            if dep_head == 'root': root_idx = len(tree)
+            tree.append(dict(word=toks[i], dep_head_idx=idx_mapping[dep_idx], dep_func=dep_func))
+        # append the head word
+        for d in tree:
+            d['dep_head'] = tree[d['dep_head_idx']]['word']
+        
+        outputs.append(dict(tree=tree, root_idx=root_idx))
+    return outputs
+
+
+def compute_mst_tree(attention_scores: torch.Tensor):
+    # attention scores should be 3 dimensions - batch x seq x seq (if it is 2 - just unsqueeze)
+    if attention_scores.ndim == 2: attention_scores = attention_scores.unsqueeze(0)
+    if attention_scores.ndim != 3 or attention_scores.shape[1] != attention_scores.shape[2]:
+        raise ValueError(f'Expected attention scores to be of shape batch x seq x seq, instead got {attention_scores.shape}')
+    
+    batch_size, seq_len, _ = attention_scores.shape
+    # start by softmaxing so the scores are comparable
+    attention_scores = attention_scores.softmax(dim=-1)
+
+    # set the values for the CLS and sep to all by very low, so they never get chosen as a replacement arc
+    attention_scores[:, 0, :] = -10000
+    attention_scores[:, -1, :] = -10000
+    attention_scores[:, :, -1] = -10000 # can never predict sep
+    
+    # find the root, and make him super high so we never have a conflict
+    root_cands = torch.argsort(attention_scores[:, :, 0], dim=-1)
+    batch_indices = torch.arange(batch_size, device=root_cands.device)
+    attention_scores[batch_indices.unsqueeze(1), root_cands, 0] = -10000
+    attention_scores[batch_indices, root_cands[:, -1], 0] = 10000
+    
+    # we start by getting the argmax for each score, and then computing the cycles and contracting them
+    sorted_indices = torch.argsort(attention_scores, dim=-1, descending=True)
+    indices = sorted_indices[:, :, 0].clone() # take the argmax
+    
+    # go through each batch item and make sure our tree works
+    for batch_idx in range(batch_size):
+        # We have one root - detect the cycles and contract them. A cycle can never contain the root so really
+        # for every cycle, we look at all the nodes, and find the highest arc out of the cycle for any values. Replace that and tada
+        has_cycle, cycle_nodes = detect_cycle(indices[batch_idx])
+        while has_cycle:
+            base_idx, head_idx = choose_contracting_arc(indices[batch_idx], sorted_indices[batch_idx], cycle_nodes, attention_scores[batch_idx])
+            indices[batch_idx, base_idx] = head_idx
+            # find the next cycle
+            has_cycle, cycle_nodes = detect_cycle(indices[batch_idx])
+            
+    return indices
+
+def detect_cycle(indices: torch.LongTensor):
+    # Simple cycle detection algorithm
+    # Returns a boolean indicating if a cycle is detected and the nodes involved in the cycle
+    visited = set()
+    for node in range(1, len(indices) - 1): # ignore the CLS/SEP tokens
+        if node in visited:
+            continue
+        current_path = set()
+        while node not in visited:
+            visited.add(node)
+            current_path.add(node)
+            node = indices[node].item()
+            if node == 0: break # roots never point to anything
+            if node in current_path:
+                return True, current_path  # Cycle detected
+    return False, None
+
+def choose_contracting_arc(indices: torch.LongTensor, sorted_indices: torch.LongTensor, cycle_nodes: set, scores: torch.FloatTensor):
+    # Chooses the highest-scoring, non-cycling arc from a graph. Iterates through 'cycle_nodes' to find
+    # the best arc based on 'scores', avoiding cycles and zero node connections.
+    # For each node, we only look at the next highest scoring non-cycling arc 
+    best_base_idx, best_head_idx = -1, -1
+    score = float('-inf')
+    
+    # convert the indices to a list once, to avoid multiple conversions (saves a few seconds)
+    currents = indices.tolist()
+    for base_node in cycle_nodes:
+        # we don't want to take anything that has a higher score than the current value - we can end up in an endless loop
+        # Since the indices are sorted, as soon as we find our current item, we can move on to the next. 
+        current = currents[base_node]
+        found_current = False
+        
+        for head_node in sorted_indices[base_node].tolist():
+            if head_node == current:
+                found_current = True
+                continue
+            if not found_current or head_node in cycle_nodes or head_node == 0: 
+                continue
+            
+            current_score = scores[base_node, head_node].item()
+            if current_score > score:
+                best_base_idx, best_head_idx, score = base_node, head_node, current_score
+            break
+    
+    return best_base_idx, best_head_idx