This repository contains the code for EMNLP-2021 paper "Word-Level Coreference Resolution"

Here is the following error i met: Epoch 1: bc/cnn/00/cnn_0001 c_loss: 2.11580 s_loss: 0.57502: 14% 394/2802 [01:04<04:54, 8.18docs/s] It seems the training process stopped. can u tell me why? thanks.

Hello, I have a new question about convert_ to_ heads.py file, in which some span and clusters will be deleted. Is this the case as follows? In those cases "A" and "A & B" are different spans with the same head word, "A". In our implementation such cases were simply discarded from the training set, because they were few and we were able to perform well, even though we couldn't predict any of such cases during inference. like u said in #2 thanks.

Hello, thank you for your great work of open source. I want to process Chinese datasets according to your process, but in convert_ to_ jsonlines.py. Py this step reports an error, do you know why? Thanks.

Hi,

I can't find the ArabicGrammaticalStructure class from the nlp.stanford. It works for english data but not for Arabic .

Converting constituents to dependencies... development: 0% 0/44 [00:00<?, ?docs/s]Exception in thread "main" java.lang.IllegalArgumentException: No head rule defined for PV+PVSUFF using class edu.stanford.nlp.trees.SemanticHeadFinder in PV+PVSUFF-39 at edu.stanford.nlp.trees.AbstractCollinsHeadFinder.determineNonTrivialHead(AbstractCollinsHeadFinder.java:222) at edu.stanford.nlp.trees.SemanticHeadFinder.determineNonTrivialHead(SemanticHeadFinder.java:348) at edu.stanford.nlp.trees.AbstractCollinsHeadFinder.determineHead(AbstractCollinsHeadFinder.java:179) at edu.stanford.nlp.trees.TreeGraphNode.percolateHeads(TreeGraphNode.java:476) at edu.stanford.nlp.trees.TreeGraphNode.percolateHeads(TreeGraphNode.java:474) at edu.stanford.nlp.trees.TreeGraphNode.percolateHeads(TreeGraphNode.java:474) at edu.stanford.nlp.trees.TreeGraphNode.percolateHeads(TreeGraphNode.java:474) at edu.stanford.nlp.trees.TreeGraphNode.percolateHeads(TreeGraphNode.java:474) at edu.stanford.nlp.trees.GrammaticalStructure.(GrammaticalStructure.java:94) at edu.stanford.nlp.trees.EnglishGrammaticalStructure.(EnglishGrammaticalStructure.java:86) at edu.stanford.nlp.trees.EnglishGrammaticalStructure.(EnglishGrammaticalStructure.java:66) at edu.stanford.nlp.parser.lexparser.EnglishTreebankParserParams.getGrammaticalStructure(EnglishTreebankParserParams.java:2271) at edu.stanford.nlp.trees.GrammaticalStructure$TreeBankGrammaticalStructureWrapper$GsIterator.primeGs(GrammaticalStructure.java:1361) at edu.stanford.nlp.trees.GrammaticalStructure$TreeBankGrammaticalStructureWrapper$GsIterator.(GrammaticalStructure.java:1348) at edu.stanford.nlp.trees.GrammaticalStructure$TreeBankGrammaticalStructureWrapper.iterator(GrammaticalStructure.java:1325) at edu.stanford.nlp.trees.GrammaticalStructure.main(GrammaticalStructure.java:1604) development: 0% 0/44 [00:00<?, ?docs/s] Traceback (most recent call last): File "convert_to_jsonlines.py", line 392, in convert_con_to_dep(args.tmp_dir, conll_filenames) File "convert_to_jsonlines.py", line 195, in convert_con_to_dep subprocess.run(cmd, check=True, stdout=out) File "/home/souid/anaconda3/envs/wl-coref/lib/python3.7/subprocess.py", line 512, in run output=stdout, stderr=stderr) subprocess.CalledProcessError: Command '['java', '-cp', 'downloads/stanford-parser.jar', 'edu.stanford.nlp.trees.EnglishGrammaticalStructure', '-basic', '-keepPunct', '-conllx', '-treeFile', 'temp/data/conll-2012/v4/data/development/data/arabic/annotations/nw/ann/00/ann_0010.v4_gold_conll']' returned non-zero exit status 1.

I ran the preparation scripts successfully.

Downloaded the roberta checkpoint from dropbox link, and placed it in data folder.

Ran the command: python calculate_conll.py roberta test 20

I noticed some errors due to subprocess because I was using python3.6 instead of python3.7.

Error was: unexpected keyword argument 'capture_output'

Fixed the issue with this

But then I got an error: 'NoneType' object has no attribute 'group' origin of error --> line 15

I ran the perl script directly in bash: perl reference-coreference-scorers/scorer.pl all data/conll_logs/roberta_test_e20.gold.conll data/conll_logs/roberta_test_e20.pred.conll none

MUC came out to be 86 (f1) but while calculating b3, I got this error: Found too many repeated mentions (> 10) in the response, so refusing to score. Please fix the output

I think it is because of this error only that the line 15 above was throwing that error (because output was empty).

How to proceed forward now? How to evaluate the results?

Hi! Thank you for posting the model. Could you please provide how to make inference from the box? If I understood correctly, model from dropbox has already been fitted, so we should be able to run it, but by the design model require original data and building of optimisers to run

class CorefModel:
    Attributes:
        config (coref.config.Config): the model's configuration,
            see config.toml for the details
        epochs_trained (int): number of epochs the model has been trained for
        trainable (Dict[str, torch.nn.Module]): trainable submodules with their
            names used as keys
        training (bool): used to toggle train/eval modes

    Submodules (in the order of their usage in the pipeline):
        tokenizer (transformers.AutoTokenizer)
        bert (transformers.AutoModel)
        we (WordEncoder)
        rough_scorer (RoughScorer)
        pw (PairwiseEncoder)
        a_scorer (AnaphoricityScorer)
        sp (SpanPredictor)
    """
    def __init__(self,
                 config_path: str,
                 section: str,
                 epochs_trained: int = 0):
        """
        A newly created model is set to evaluation mode.

        Args:
            config_path (str): the path to the toml file with the configuration
            section (str): the selected section of the config file
            epochs_trained (int): the number of epochs finished
                (useful for warm start)
        """
        self.config = CorefModel._load_config(config_path, section)
        self.epochs_trained = epochs_trained
        self._docs: Dict[str, List[Doc]] = {}
        self._build_model()
        self._build_optimizers()
        self._set_training(False)
        self._coref_criterion = CorefLoss(self.config.bce_loss_weight)
        self._span_criterion = torch.nn.CrossEntropyLoss(reduction="sum")

So maybe there is an option to run it without all this stuff?

Based on my observation in this code base, training use the following features, e.g: cased_words, sent_id, speaker, pos, deprel, head, clusters.

then converted into: cased_words, sent_id, speaker, pos, deprel, head, head2span, word_clusters, span_clusters.

while in inference data example, the feature used only cased_words, sent_id, and optionally speaker information.

My questions is.

1. how we get the pos, deprel, head, and clusters data from in inference mode? It is derived from cased_words or not?
2. in training mode, is the speaker, pos, deprel, head, clusters data is used as well?

Thank you

Hi, thanks so much for your work! I have a question regarding convert_to_heads.py script. I'm trying to make RoBERTa learn coreference resolution, but my data is in .conllu format. I have quite hard time trying to preprocess data/modify some of your code to make it work. Can you share some insights/thoughts on that? I would be very much obliged.

Cheers

Hello, I'd like to ask about the .jsonlines file executived through convert_ to_ jsonlines. py, Can some attributes in the jsonlines file be successfully trained after being discarded? Such as speaker, pos.

Hello, great work.

I had two questions:

1. what sent_id in the sample input file supposed to refer to??

2. If I want to make an inference for dialogue like tc genre, what should be the conversation format ??

Currently, when running this source code. I have an error cuda running out of memory. Since single GPU have only 32GB memory.

but, in another side, I have access to server which have 8 GPU (each of them having 32GB memory). Can I run this training experiment with the paralel mode?

If it can, how to achieve that?

thanks in advance..

CUDA-enabled machine (48 GB to train, 4 GB to evaluate)

@vdobrovolskii friendly ping Are 48GB really needed to train? Can't we train longer (how long) with less ? couldn't your project leverage FP16, FP8 and other optimizations ? You can get them out of the box if you use roberta from the Transformers library https://github.com/huggingface/transformers Also there is accelerate https://huggingface.co/docs/accelerate/index

I have a 3070 with 8GB of GDDR6 :/