Thank you for your implementation of HMNet! I have a few questions about testing the model on my own data.

Here's what I've done so far:

I've put my data in the format of data_list:

[{'dialogues': [{'pos_sentence': 'uh punct nnp cc nnp rb in dt nn punct uh punct ex vbz cd punct',
    'role': 'speaker1',
    'sentence': "Yeah. Bob and him right at the moment. Ok, there's one."},
   {'pos_sentence': 'prp vbp punct prp md vb prp nn cc punct rb prp vbz punct',
    'role': 'speaker2',
    'sentence': 'I mean - I can start my video but, there it is.'},
   {'pos_sentence': 'uh punct', 'role': 'speaker1', 'sentence': 'Cool.'},
...]

I then slightly modified the AMIDataset class and afterwards loaded my data into test_dataloader.

# For input dialogues
PAD = 0
BOS = 1
EOS = 2
UNK = 3
# For target summaries
BEGIN = 4
END = 5


class AttrDict(dict):
    """ Access dictionary keys like attribute
        https://stackoverflow.com/questions/4984647/accessing-dict-keys-like-an-attribute
    """
    def __init__(self, *av, **kav):
        dict.__init__(self, *av, **kav)
        self.__dict__ = self


class TranscriptDataset(Dataset):
    def __init__(self, hparams, data_list, type='test', vocab_word=None,
                 vocab_role=None, vocab_pos=None, max_vocab_size=50000):
        super().__init__()
        self.hparams = hparams
        self.data_list = data_list

        if (vocab_word == None) and (vocab_role == None):
            print('building vocab')
            counter, role_counter, pos_counter = self.build_counter()
            self.vocab_word = self.build_vocab(counter, max_vocab_size, type='word')
            self.vocab_role = self.build_vocab(role_counter, max_vocab_size, type='role')
            self.vocab_pos = self.build_vocab(pos_counter, max_vocab_size, type='pos')
        else:
            self.vocab_word = vocab_word
            self.vocab_role = vocab_role
            self.vocab_pos = vocab_pos

    def __len__(self):
        return len(self.data_list)

    def __getitem__(self, index):
        """

        :param index:
        :return:
            Input Examples (vocab_ids)
                dialogues: list of (role, utterance)
                labels: reference summaries for texts
        """
        dialogues = self.data_list[index]['dialogues']
        labels = self.data_list[index]['labels']

        dialogues_ids = [] # (num_turn, seq_len)
        pos_ids = [] # (num_turn, seq_len)
        role_ids = [] # (num_turns)

        label_tokens = self.tokenize(labels)

        labels_ids = self.tokens2ids(label_tokens, self.vocab_word.token2id,
                                     is_reference=True) #(seq_len)

        for turn_idx, dialogue in enumerate(dialogues):
            # print('turn_idx: ', turn_idx)
            if turn_idx >= self.hparams.max_length:
                break
            sentence = dialogue['sentence']
            sentence_count = len(sentence.split('.')) - 1
            if sentence_count < 2 or len(sentence) < 4:
                continue
            sentence = sentence.replace('. .', '.').replace(', ,', ',')

            tokens = self.tokenize(sentence)
            pos_tokens = self.tokenize(dialogue['pos_sentence'])
            role_tokens = self.tokenize(dialogue['role'])

            if len(tokens) >= self.hparams.max_length - 2:
                tokens = tokens[:self.hparams.max_length - 2]
                pos_tokens = pos_tokens[:self.hparams.max_length - 2]

            token_ids = self.tokens2ids(tokens, self.vocab_word.token2id)
            pos_token_ids = self.tokens2ids(pos_tokens, self.vocab_pos.token2id)
            role_token_ids = self.tokens2ids(role_tokens, self.vocab_role.token2id, is_role=True)

            dialogues_ids.append(token_ids)
            pos_ids.append(pos_token_ids)
            role_ids.append(role_token_ids)

        padded_dialogues, dialogues_lens, src_masks = self.pad_sequence(dialogues_ids)
        padded_pos_ids, _, _ = self.pad_sequence(pos_ids)

        data = dict()
        data['dialogues'] = dialogues
        data['labels'] = labels
        data['dialogues_ids'] = padded_dialogues
        data['pos_ids'] = padded_pos_ids
        data['dialogues_lens'] = torch.tensor(dialogues_lens).long()
        data['src_masks'] = src_masks
        data['role_ids'] = torch.tensor(role_ids).long()
        data['labels_ids'] = torch.tensor(labels_ids).long()
        return data

    def pad_sequence(self, seqs):
        lens = [len(seq) for seq in seqs]
        max_seq_length = max(lens)
        padded_seqs = torch.zeros(len(seqs), max_seq_length).long()
        for i, seq in enumerate(seqs):
            end_idx = lens[i]
            padded_seqs[i, :end_idx] = torch.LongTensor(seq[:end_idx])

        src_masks = _gen_seq_bias_mask(lens, max_seq_length)

        return padded_seqs, lens, src_masks

    def tokenize(self, sentence):
        return sentence.split()

    def build_counter(self):
        counter = Counter()
        role_counter = Counter()
        pos_counter = Counter()

        role_words = []
        sentence_words = []
        pos_words = []

        augmented_data_list = [] # add dev vocab
        augmented_data_list.extend(self.data_list)

        dev_examples = torch.load(self.hparams.data_dir + 'dev_corpus')

        print('[%s] %d examples is loaded' % ('Dev', len(dev_examples)))
        for key, value in dev_examples.items():
            texts = value['texts']
            labels = value['labels']
            dialogues = []
            for each in texts:
                role = each[1]
                sentence = ' '.join(word_pos.split('/')[0] for word_pos in each[2].split())
                sentence = sentence.strip().lower()
                pos_sentence = ' '.join(word_pos.split('/')[1] for word_pos in each[2].split())
                pos_sentence = pos_sentence.strip().lower()
                dialogues.append({'role': role, 'sentence': sentence, 'pos_sentence': pos_sentence})
            augmented_data_list.append({'labels': labels, 'dialogues': dialogues})

        for data in augmented_data_list:
            dialogues = data['dialogues']
            for dialogue in dialogues:
                sentence_words.append(self.tokenize(dialogue['sentence']))
                pos_words.append(self.tokenize(dialogue['pos_sentence']))
                role_words.append([dialogue['role']])
            labels = data['labels']
            sentence_words.append(self.tokenize(labels))

        for words in sentence_words:
            counter.update(words)

        for role in role_words:
            role_counter.update(role)

        for pos in pos_words:
            pos_counter.update(pos)
        print('role_counter: ', role_counter)
        print('pos_counter: ', pos_counter)

        return counter, role_counter, pos_counter

    def build_vocab(self, counter, max_vocab_size, type='word'):
        vocab = AttrDict()
        if type == 'word':
            print("\n===== Building [Word Vocab] =========")
        elif type == 'role':
            print("\n===== Building [Role Vocab] =========")
        elif type == 'pos':
            print("\n===== Building [POS Vocab] =========")

        vocab.token2id = {'<PAD>': PAD, '<BOS>': BOS, '<EOS>': EOS,
                          '<UNK>': UNK, '<BEGIN>': BEGIN, '<END>': END}
        preset_vocab_size = len(vocab.token2id)
        print('preset_vocab_size: ', preset_vocab_size)
        vocab.token2id.update(
            {token: _id + preset_vocab_size for _id, (token, count) in
             tqdm(enumerate(counter.most_common(max_vocab_size)))})
        vocab.id2token = {v: k for k, v in tqdm(vocab.token2id.items())}
        print('Vocab size: ', len(vocab.token2id))
        print('==========================================')
        return vocab

    def tokens2ids(self, tokens, token2id, is_reference=False, is_role=False):
        seq = []
        if is_reference is False and is_role is False:
            # For input dialogues.
            seq.append(BOS)
            seq.extend([token2id.get(token, UNK) for token in tokens])
            seq.append(EOS)
        elif is_reference:
            # For target summaries.
            seq.append(BEGIN)
            seq.extend([token2id.get(token, UNK) for token in tokens])
            seq.append(END)

        if is_role:
            seq.extend([token2id.get(token, UNK) for token in tokens])
        return seq

test = TranscriptDataset(hparams=hparams,
                         data_list=data_list)

test_dataloader = DataLoader(
            test,
            batch_size=hparams.batch_size,
            num_workers=hparams.workers,
            drop_last=False
        )

I instantiate the Predictor class, using the vocab_role and vocab_pos from my test data. I use the vocab_word that was saved from model training otherwise I get an error about different tensor sizes:

predictor = Predictor(hparams,
                      model=None,
                      vocab_word=torch.load(hparams.vocab_word_path), 
                      vocab_role=test.vocab_role,
                      vocab_pos=test.vocab_pos,
                      checkpoint=None,
                    summary_writer = SummaryWriter('./', comment="Unmt"))

Finally, I try to generate summaries using some of the code from the evaluate() method within Predictor():

cand_list = []
for batch_idx, batch in enumerate(tqdm(test_dataloader)):
    data = batch
    dialogues_ids = data['dialogues_ids'].to(predictor.device)
    pos_ids = data['pos_ids'].to(predictor.device)
    labels_ids = data['labels_ids'].to(predictor.device)  # [batch, tgt_seq_len]
    src_masks = data['src_masks'].to(predictor.device)
    role_ids = data['role_ids'].to(predictor.device)
    generated_summaries = predictor.inference(inputs=dialogues_ids, src_masks=src_masks,
                                                    role_ids=role_ids, pos_ids=pos_ids)
    cand_list.append(generated_summaries)

This is where I am running into some issues.

It is generating a summary, but the summary is completely irrelevant to my data. The summary it generates seems to be the summary from the training data:

the prototype of the project manager opened the project finance of the meeting || the user interface designer then discussed the group then then recapped the interface specialist on the product and discussed the marketing expert discussed the industrial designer and the team discussed the team and vegetable designer heshe || na || the remote || he discussed the remote control || the team then discussed their marketing expert then on the team on the interface designer was examined the prototype || he instructed the team presents the prototype and discussed their designers to create the project process on the project || the project budget || he ended the team suggested the group had into the group report || the group decided to incorporate the group and the group discussed components and a remote control functions and the remote controls to prepare the remote and the user requirements for the product || they will work on their project requirements || ...

I am trying to figure out where the issue is. Even if I change the test data to a different meeting transcript that I have, it still generates that same above summary.

Mar 05 '21 02:03 spencernelsonucla

Hello, does this model's inference on custom data not work?

Apr 01 '21 02:04 Sperao

Hello. Sorry for late reply. First of all, thank you for interesting my code. For custom inference, you should write own code based on my evaluation code. However, since the dataset I used is consist of small dataset, I suggest you to use bart(huggingface) model which is pretrained for generation task. Thank you. 2021년 4월 1일 (목) 오전 11:18, Sperao @.***>님이 작성:

Did you solve this problem？

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/JudeLee19/HMNet-End-to-End-Abstractive-Summarization-for-Meetings/issues/12#issuecomment-811588563, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABSYFRNCRXMOIT6WRUG7VZ3TGPJYTANCNFSM4YULGZ5A .

-- Best,

Dongyub (Jude) Lee

Apr 01 '21 02:04 JudeLee19

Thank you for your implementation of HMNet! I have a few questions about testing the model on my own data.

Here's what I've done so far:

I've put my data in the format of data_list:

[{'dialogues': [{'pos_sentence': 'uh punct nnp cc nnp rb in dt nn punct uh punct ex vbz cd punct',
    'role': 'speaker1',
    'sentence': "Yeah. Bob and him right at the moment. Ok, there's one."},
   {'pos_sentence': 'prp vbp punct prp md vb prp nn cc punct rb prp vbz punct',
    'role': 'speaker2',
    'sentence': 'I mean - I can start my video but, there it is.'},
   {'pos_sentence': 'uh punct', 'role': 'speaker1', 'sentence': 'Cool.'},
...]

I then slightly modified the AMIDataset class and afterwards loaded my data into test_dataloader.

# For input dialogues
PAD = 0
BOS = 1
EOS = 2
UNK = 3
# For target summaries
BEGIN = 4
END = 5


class AttrDict(dict):
    """ Access dictionary keys like attribute
        https://stackoverflow.com/questions/4984647/accessing-dict-keys-like-an-attribute
    """
    def __init__(self, *av, **kav):
        dict.__init__(self, *av, **kav)
        self.__dict__ = self


class TranscriptDataset(Dataset):
    def __init__(self, hparams, data_list, type='test', vocab_word=None,
                 vocab_role=None, vocab_pos=None, max_vocab_size=50000):
        super().__init__()
        self.hparams = hparams
        self.data_list = data_list

        if (vocab_word == None) and (vocab_role == None):
            print('building vocab')
            counter, role_counter, pos_counter = self.build_counter()
            self.vocab_word = self.build_vocab(counter, max_vocab_size, type='word')
            self.vocab_role = self.build_vocab(role_counter, max_vocab_size, type='role')
            self.vocab_pos = self.build_vocab(pos_counter, max_vocab_size, type='pos')
        else:
            self.vocab_word = vocab_word
            self.vocab_role = vocab_role
            self.vocab_pos = vocab_pos

    def __len__(self):
        return len(self.data_list)

    def __getitem__(self, index):
        """

        :param index:
        :return:
            Input Examples (vocab_ids)
                dialogues: list of (role, utterance)
                labels: reference summaries for texts
        """
        dialogues = self.data_list[index]['dialogues']
        labels = self.data_list[index]['labels']

        dialogues_ids = [] # (num_turn, seq_len)
        pos_ids = [] # (num_turn, seq_len)
        role_ids = [] # (num_turns)

        label_tokens = self.tokenize(labels)

        labels_ids = self.tokens2ids(label_tokens, self.vocab_word.token2id,
                                     is_reference=True) #(seq_len)

        for turn_idx, dialogue in enumerate(dialogues):
            # print('turn_idx: ', turn_idx)
            if turn_idx >= self.hparams.max_length:
                break
            sentence = dialogue['sentence']
            sentence_count = len(sentence.split('.')) - 1
            if sentence_count < 2 or len(sentence) < 4:
                continue
            sentence = sentence.replace('. .', '.').replace(', ,', ',')

            tokens = self.tokenize(sentence)
            pos_tokens = self.tokenize(dialogue['pos_sentence'])
            role_tokens = self.tokenize(dialogue['role'])

            if len(tokens) >= self.hparams.max_length - 2:
                tokens = tokens[:self.hparams.max_length - 2]
                pos_tokens = pos_tokens[:self.hparams.max_length - 2]

            token_ids = self.tokens2ids(tokens, self.vocab_word.token2id)
            pos_token_ids = self.tokens2ids(pos_tokens, self.vocab_pos.token2id)
            role_token_ids = self.tokens2ids(role_tokens, self.vocab_role.token2id, is_role=True)

            dialogues_ids.append(token_ids)
            pos_ids.append(pos_token_ids)
            role_ids.append(role_token_ids)

        padded_dialogues, dialogues_lens, src_masks = self.pad_sequence(dialogues_ids)
        padded_pos_ids, _, _ = self.pad_sequence(pos_ids)

        data = dict()
        data['dialogues'] = dialogues
        data['labels'] = labels
        data['dialogues_ids'] = padded_dialogues
        data['pos_ids'] = padded_pos_ids
        data['dialogues_lens'] = torch.tensor(dialogues_lens).long()
        data['src_masks'] = src_masks
        data['role_ids'] = torch.tensor(role_ids).long()
        data['labels_ids'] = torch.tensor(labels_ids).long()
        return data

    def pad_sequence(self, seqs):
        lens = [len(seq) for seq in seqs]
        max_seq_length = max(lens)
        padded_seqs = torch.zeros(len(seqs), max_seq_length).long()
        for i, seq in enumerate(seqs):
            end_idx = lens[i]
            padded_seqs[i, :end_idx] = torch.LongTensor(seq[:end_idx])

        src_masks = _gen_seq_bias_mask(lens, max_seq_length)

        return padded_seqs, lens, src_masks

    def tokenize(self, sentence):
        return sentence.split()

    def build_counter(self):
        counter = Counter()
        role_counter = Counter()
        pos_counter = Counter()

        role_words = []
        sentence_words = []
        pos_words = []

        augmented_data_list = [] # add dev vocab
        augmented_data_list.extend(self.data_list)

        dev_examples = torch.load(self.hparams.data_dir + 'dev_corpus')

        print('[%s] %d examples is loaded' % ('Dev', len(dev_examples)))
        for key, value in dev_examples.items():
            texts = value['texts']
            labels = value['labels']
            dialogues = []
            for each in texts:
                role = each[1]
                sentence = ' '.join(word_pos.split('/')[0] for word_pos in each[2].split())
                sentence = sentence.strip().lower()
                pos_sentence = ' '.join(word_pos.split('/')[1] for word_pos in each[2].split())
                pos_sentence = pos_sentence.strip().lower()
                dialogues.append({'role': role, 'sentence': sentence, 'pos_sentence': pos_sentence})
            augmented_data_list.append({'labels': labels, 'dialogues': dialogues})

        for data in augmented_data_list:
            dialogues = data['dialogues']
            for dialogue in dialogues:
                sentence_words.append(self.tokenize(dialogue['sentence']))
                pos_words.append(self.tokenize(dialogue['pos_sentence']))
                role_words.append([dialogue['role']])
            labels = data['labels']
            sentence_words.append(self.tokenize(labels))

        for words in sentence_words:
            counter.update(words)

        for role in role_words:
            role_counter.update(role)

        for pos in pos_words:
            pos_counter.update(pos)
        print('role_counter: ', role_counter)
        print('pos_counter: ', pos_counter)

        return counter, role_counter, pos_counter

    def build_vocab(self, counter, max_vocab_size, type='word'):
        vocab = AttrDict()
        if type == 'word':
            print("\n===== Building [Word Vocab] =========")
        elif type == 'role':
            print("\n===== Building [Role Vocab] =========")
        elif type == 'pos':
            print("\n===== Building [POS Vocab] =========")

        vocab.token2id = {'<PAD>': PAD, '<BOS>': BOS, '<EOS>': EOS,
                          '<UNK>': UNK, '<BEGIN>': BEGIN, '<END>': END}
        preset_vocab_size = len(vocab.token2id)
        print('preset_vocab_size: ', preset_vocab_size)
        vocab.token2id.update(
            {token: _id + preset_vocab_size for _id, (token, count) in
             tqdm(enumerate(counter.most_common(max_vocab_size)))})
        vocab.id2token = {v: k for k, v in tqdm(vocab.token2id.items())}
        print('Vocab size: ', len(vocab.token2id))
        print('==========================================')
        return vocab

    def tokens2ids(self, tokens, token2id, is_reference=False, is_role=False):
        seq = []
        if is_reference is False and is_role is False:
            # For input dialogues.
            seq.append(BOS)
            seq.extend([token2id.get(token, UNK) for token in tokens])
            seq.append(EOS)
        elif is_reference:
            # For target summaries.
            seq.append(BEGIN)
            seq.extend([token2id.get(token, UNK) for token in tokens])
            seq.append(END)

        if is_role:
            seq.extend([token2id.get(token, UNK) for token in tokens])
        return seq

test = TranscriptDataset(hparams=hparams,
                         data_list=data_list)

test_dataloader = DataLoader(
            test,
            batch_size=hparams.batch_size,
            num_workers=hparams.workers,
            drop_last=False
        )

I instantiate the Predictor class, using the vocab_role and vocab_pos from my test data. I use the vocab_word that was saved from model training otherwise I get an error about different tensor sizes:

predictor = Predictor(hparams,
                      model=None,
                      vocab_word=torch.load(hparams.vocab_word_path), 
                      vocab_role=test.vocab_role,
                      vocab_pos=test.vocab_pos,
                      checkpoint=None,
                    summary_writer = SummaryWriter('./', comment="Unmt"))

Finally, I try to generate summaries using some of the code from the evaluate() method within Predictor():

cand_list = []
for batch_idx, batch in enumerate(tqdm(test_dataloader)):
    data = batch
    dialogues_ids = data['dialogues_ids'].to(predictor.device)
    pos_ids = data['pos_ids'].to(predictor.device)
    labels_ids = data['labels_ids'].to(predictor.device)  # [batch, tgt_seq_len]
    src_masks = data['src_masks'].to(predictor.device)
    role_ids = data['role_ids'].to(predictor.device)
    generated_summaries = predictor.inference(inputs=dialogues_ids, src_masks=src_masks,
                                                    role_ids=role_ids, pos_ids=pos_ids)
    cand_list.append(generated_summaries)

This is where I am running into some issues.

It is generating a summary, but the summary is completely irrelevant to my data. The summary it generates seems to be the summary from the training data:

the prototype of the project manager opened the project finance of the meeting || the user interface designer then discussed the group then then recapped the interface specialist on the product and discussed the marketing expert discussed the industrial designer and the team discussed the team and vegetable designer heshe || na || the remote || he discussed the remote control || the team then discussed their marketing expert then on the team on the interface designer was examined the prototype || he instructed the team presents the prototype and discussed their designers to create the project process on the project || the project budget || he ended the team suggested the group had into the group report || the group decided to incorporate the group and the group discussed components and a remote control functions and the remote controls to prepare the remote and the user requirements for the product || they will work on their project requirements || ...

I am trying to figure out where the issue is. Even if I change the test data to a different meeting transcript that I have, it still generates that same above summary.

Did you solve this problem？

Apr 13 '21 09:04 Sperao

Hello, Sperao.

I believe the reason of the poor quality of generated summary is caused by the very small dataset. I recommend you to use the BART-large model which is trained on the large-scale dataset with various training objectives.

Best,

Dongyub (Jude) Lee

2021년 4월 13일 (화) 오후 6:23, Sperao @.***>님이 작성:

Thank you for your implementation of HMNet! I have a few questions about testing the model on my own data.

Here's what I've done so far:

I've put my data in the format of data_list:

[{'dialogues': [{'pos_sentence': 'uh punct nnp cc nnp rb in dt nn punct uh punct ex vbz cd punct',

'role': 'speaker1',

'sentence': "Yeah. Bob and him right at the moment. Ok, there's one."},

{'pos_sentence': 'prp vbp punct prp md vb prp nn cc punct rb prp vbz punct',

'role': 'speaker2',

'sentence': 'I mean - I can start my video but, there it is.'},

{'pos_sentence': 'uh punct', 'role': 'speaker1', 'sentence': 'Cool.'},

...]

I then slightly modified the AMIDataset class and afterwards loaded my data into test_dataloader.

For input dialogues

PAD = 0

BOS = 1

EOS = 2

UNK = 3

For target summaries

BEGIN = 4

END = 5

class AttrDict(dict):

""" Access dictionary keys like attribute

    https://stackoverflow.com/questions/4984647/accessing-dict-keys-like-an-attribute

"""

def __init__(self, *av, **kav):

    dict.__init__(self, *av, **kav)

    self.__dict__ = self

class TranscriptDataset(Dataset):

def __init__(self, hparams, data_list, type='test', vocab_word=None,

             vocab_role=None, vocab_pos=None, max_vocab_size=50000):

    super().__init__()

    self.hparams = hparams

    self.data_list = data_list



    if (vocab_word == None) and (vocab_role == None):

        print('building vocab')

        counter, role_counter, pos_counter = self.build_counter()

        self.vocab_word = self.build_vocab(counter, max_vocab_size, type='word')

        self.vocab_role = self.build_vocab(role_counter, max_vocab_size, type='role')

        self.vocab_pos = self.build_vocab(pos_counter, max_vocab_size, type='pos')

    else:

        self.vocab_word = vocab_word

        self.vocab_role = vocab_role

        self.vocab_pos = vocab_pos



def __len__(self):

    return len(self.data_list)



def __getitem__(self, index):

    """



    :param index:

    :return:

        Input Examples (vocab_ids)

            dialogues: list of (role, utterance)

            labels: reference summaries for texts

    """

    dialogues = self.data_list[index]['dialogues']

    labels = self.data_list[index]['labels']



    dialogues_ids = [] # (num_turn, seq_len)

    pos_ids = [] # (num_turn, seq_len)

    role_ids = [] # (num_turns)



    label_tokens = self.tokenize(labels)



    labels_ids = self.tokens2ids(label_tokens, self.vocab_word.token2id,

                                 is_reference=True) #(seq_len)



    for turn_idx, dialogue in enumerate(dialogues):

        # print('turn_idx: ', turn_idx)

        if turn_idx >= self.hparams.max_length:

            break

        sentence = dialogue['sentence']

        sentence_count = len(sentence.split('.')) - 1

        if sentence_count < 2 or len(sentence) < 4:

            continue

        sentence = sentence.replace('. .', '.').replace(', ,', ',')



        tokens = self.tokenize(sentence)

        pos_tokens = self.tokenize(dialogue['pos_sentence'])

        role_tokens = self.tokenize(dialogue['role'])



        if len(tokens) >= self.hparams.max_length - 2:

            tokens = tokens[:self.hparams.max_length - 2]

            pos_tokens = pos_tokens[:self.hparams.max_length - 2]



        token_ids = self.tokens2ids(tokens, self.vocab_word.token2id)

        pos_token_ids = self.tokens2ids(pos_tokens, self.vocab_pos.token2id)

        role_token_ids = self.tokens2ids(role_tokens, self.vocab_role.token2id, is_role=True)



        dialogues_ids.append(token_ids)

        pos_ids.append(pos_token_ids)

        role_ids.append(role_token_ids)



    padded_dialogues, dialogues_lens, src_masks = self.pad_sequence(dialogues_ids)

    padded_pos_ids, _, _ = self.pad_sequence(pos_ids)



    data = dict()

    data['dialogues'] = dialogues

    data['labels'] = labels

    data['dialogues_ids'] = padded_dialogues

    data['pos_ids'] = padded_pos_ids

    data['dialogues_lens'] = torch.tensor(dialogues_lens).long()

    data['src_masks'] = src_masks

    data['role_ids'] = torch.tensor(role_ids).long()

    data['labels_ids'] = torch.tensor(labels_ids).long()

    return data



def pad_sequence(self, seqs):

    lens = [len(seq) for seq in seqs]

    max_seq_length = max(lens)

    padded_seqs = torch.zeros(len(seqs), max_seq_length).long()

    for i, seq in enumerate(seqs):

        end_idx = lens[i]

        padded_seqs[i, :end_idx] = torch.LongTensor(seq[:end_idx])



    src_masks = _gen_seq_bias_mask(lens, max_seq_length)



    return padded_seqs, lens, src_masks



def tokenize(self, sentence):

    return sentence.split()



def build_counter(self):

    counter = Counter()

    role_counter = Counter()

    pos_counter = Counter()



    role_words = []

    sentence_words = []

    pos_words = []



    augmented_data_list = [] # add dev vocab

    augmented_data_list.extend(self.data_list)



    dev_examples = torch.load(self.hparams.data_dir + 'dev_corpus')



    print('[%s] %d examples is loaded' % ('Dev', len(dev_examples)))

    for key, value in dev_examples.items():

        texts = value['texts']

        labels = value['labels']

        dialogues = []

        for each in texts:

            role = each[1]

            sentence = ' '.join(word_pos.split('/')[0] for word_pos in each[2].split())

            sentence = sentence.strip().lower()

            pos_sentence = ' '.join(word_pos.split('/')[1] for word_pos in each[2].split())

            pos_sentence = pos_sentence.strip().lower()

            dialogues.append({'role': role, 'sentence': sentence, 'pos_sentence': pos_sentence})

        augmented_data_list.append({'labels': labels, 'dialogues': dialogues})



    for data in augmented_data_list:

        dialogues = data['dialogues']

        for dialogue in dialogues:

            sentence_words.append(self.tokenize(dialogue['sentence']))

            pos_words.append(self.tokenize(dialogue['pos_sentence']))

            role_words.append([dialogue['role']])

        labels = data['labels']

        sentence_words.append(self.tokenize(labels))



    for words in sentence_words:

        counter.update(words)



    for role in role_words:

        role_counter.update(role)



    for pos in pos_words:

        pos_counter.update(pos)

    print('role_counter: ', role_counter)

    print('pos_counter: ', pos_counter)



    return counter, role_counter, pos_counter



def build_vocab(self, counter, max_vocab_size, type='word'):

    vocab = AttrDict()

    if type == 'word':

        print("\n===== Building [Word Vocab] =========")

    elif type == 'role':

        print("\n===== Building [Role Vocab] =========")

    elif type == 'pos':

        print("\n===== Building [POS Vocab] =========")



    vocab.token2id = {'<PAD>': PAD, '<BOS>': BOS, '<EOS>': EOS,

                      '<UNK>': UNK, '<BEGIN>': BEGIN, '<END>': END}

    preset_vocab_size = len(vocab.token2id)

    print('preset_vocab_size: ', preset_vocab_size)

    vocab.token2id.update(

        {token: _id + preset_vocab_size for _id, (token, count) in

         tqdm(enumerate(counter.most_common(max_vocab_size)))})

    vocab.id2token = {v: k for k, v in tqdm(vocab.token2id.items())}

    print('Vocab size: ', len(vocab.token2id))

    print('==========================================')

    return vocab



def tokens2ids(self, tokens, token2id, is_reference=False, is_role=False):

    seq = []

    if is_reference is False and is_role is False:

        # For input dialogues.

        seq.append(BOS)

        seq.extend([token2id.get(token, UNK) for token in tokens])

        seq.append(EOS)

    elif is_reference:

        # For target summaries.

        seq.append(BEGIN)

        seq.extend([token2id.get(token, UNK) for token in tokens])

        seq.append(END)



    if is_role:

        seq.extend([token2id.get(token, UNK) for token in tokens])

    return seq

test = TranscriptDataset(hparams=hparams,

                     data_list=data_list)

test_dataloader = DataLoader(

        test,

        batch_size=hparams.batch_size,

        num_workers=hparams.workers,

        drop_last=False

    )

I instantiate the Predictor class, using the vocab_role and vocab_pos from my test data. I use the vocab_word that was saved from model training otherwise I get an error about different tensor sizes:

predictor = Predictor(hparams,

                  model=None,

                  vocab_word=torch.load(hparams.vocab_word_path),

                  vocab_role=test.vocab_role,

                  vocab_pos=test.vocab_pos,

                  checkpoint=None,

                summary_writer = SummaryWriter('./', comment="Unmt"))

Finally, I try to generate summaries using some of the code from the evaluate() method within Predictor():

cand_list = []

for batch_idx, batch in enumerate(tqdm(test_dataloader)):

data = batch

dialogues_ids = data['dialogues_ids'].to(predictor.device)

pos_ids = data['pos_ids'].to(predictor.device)

labels_ids = data['labels_ids'].to(predictor.device)  # [batch, tgt_seq_len]

src_masks = data['src_masks'].to(predictor.device)

role_ids = data['role_ids'].to(predictor.device)

generated_summaries = predictor.inference(inputs=dialogues_ids, src_masks=src_masks,

                                                role_ids=role_ids, pos_ids=pos_ids)

cand_list.append(generated_summaries)

This is where I am running into some issues.

It is generating a summary, but the summary is completely irrelevant to my data. The summary it generates seems to be the summary from the training data:

the prototype of the project manager opened the project finance of the meeting || the user interface designer then discussed the group then then recapped the interface specialist on the product and discussed the marketing expert discussed the industrial designer and the team discussed the team and vegetable designer heshe || na || the remote || he discussed the remote control || the team then discussed their marketing expert then on the team on the interface designer was examined the prototype || he instructed the team presents the prototype and discussed their designers to create the project process on the project || the project budget || he ended the team suggested the group had into the group report || the group decided to incorporate the group and the group discussed components and a remote control functions and the remote controls to prepare the remote and the user requirements for the product || they will work on their project requirements || ...

I am trying to figure out where the issue is. Even if I change the test data to a different meeting transcript that I have, it still generates that same above summary.

Did you solve this problem？

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/JudeLee19/HMNet-End-to-End-Abstractive-Summarization-for-Meetings/issues/12#issuecomment-818589428, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABSYFRPOQ2IJF5RFR6YNKLDTIQEP7ANCNFSM4YULGZ5A .

Apr 13 '21 09:04 JudeLee19

Hi @spencernelsonucla could you please your own data_list structure for me, I am trying to run your shared codes but it throws an error because there is not labels in the data_list.

Jun 03 '21 09:06 alifarid0991

HMNet-End-to-End-Abstractive-Summarization-for-Meetings
HMNet-End-to-End-Abstractive-Summarization-for-Meetings copied to clipboard

Inference on custom data not working

For input dialogues

For target summaries

HMNet-End-to-End-Abstractive-Summarization-for-Meetings HMNet-End-to-End-Abstractive-Summarization-for-Meetings copied to clipboard

Inference on custom data not working

For input dialogues

For target summaries

HMNet-End-to-End-Abstractive-Summarization-for-Meetings
HMNet-End-to-End-Abstractive-Summarization-for-Meetings copied to clipboard