Improving Textual coherence

• GPU-optimised PyTorch implementation of a neural local discriminative model (LCD) with a much smaller negative sampling space that can efficiently learn against incorrect orderings inspired by Xu et al. (2019).

• Improves accuracy on insertion tasks by 16% than reported in the paper.

Preprocessing

Permute the original documents or paragraphs to obtain the negative samples for evaluation.

python preprocess.py

Training and Evaluation

To train and evaluate a model, run

python run_model.py --sent_encoder <sent_encoder>

where sent_encoder can be "average_glove" or "sbert" (sbert by default).

Literature Review

Main

• A Cross-Domain Transferable Neural Coherence Model
• Neural Net Models for Open-Domain Discourse Coherence

Others (mainly for scoring methods)

**: Uses sigmoid for scoring

• https://aclanthology.org/D14-1218.pdf
• https://aclanthology.org/P17-1121/
• https://arxiv.org/abs/1909.00349
• https://aclanthology.org/D18-1464/ **
• https://aclanthology.org/2020.coling-main.194/
• https://ojs.aaai.org/index.php/AAAI/article/view/12045 **
• https://aclanthology.org/W16-3407.pdf
• https://link.springer.com/chapter/10.1007/978-3-030-01716-3_32 **
• https://aclanthology.org/P11-1100.pdf
• https://aclanthology.org/C14-1089.pdf
• https://arxiv.org/abs/2005.10389
• https://aclanthology.org/P08-2011.pdf
• https://arxiv.org/abs/1804.06898 **
• https://dl.acm.org/doi/abs/10.1145/3132847.3133047

Evaluation

The following methods are known of

• Discrimination
• Insertion
• Paragraph Reconstruction
• Readability Assessment

The first two are used.

Experiments

Comparison of baseline results

• Results with the recommended hyperparameters

Output function	Encoder	Discrimination	Insertion
None	Avg_Glove	0.92537	0.29847
Sigmoid	Avg_Glove	0.80393	0.21469
TanH	Avg_Glove	0.10488	0.72060
None	SBERT	0.93851	0.33034

"hparams": {
    "input_dropout": 0.6,
    "hidden_layers": 1,
    "hidden_dropout": 0.3,
    "margin": 5.0,
    "weight_decay": 0.0,
    "dpout_model": 0.0
}

Analysis of Hyperparameter tuning

Hyparameters tuned:

• input_dropout: [0.5, 0.6, 0.7]
• hidden_layers: [1, 2]
• hidden_dropout: [0.2, 0.3, 0.4]
• margin: [4.0, 5.0, 6.0]
• weight_decay: [0.0, 0.1]
• dpout_model: [0.0, 0.05, 0.1]

Discrimination validation

Results for best 20 are taken

Scores

• Discrimination

  Description	Value
  mean	0.92649
  std	0.00107
  min	0.925
  max	0.9291
  25%	0.92585
  50%	0.92635
  75%	0.9269

• Insertion

  Description	Value
  mean	0.30562
  std	0.00156
  min	0.3036
  max	0.3091
  25%	0.304275
  50%	0.3054
  75%	0.306825

Trends

• input_dropout: 0.5 > 0.6
• hidden_layers: 2 > 1
• margin: 6.0 > 4.0 > 5.0
• weight_decay: 0.0
• dpout_model: 0.0 > 0.1 > 0.05
• hidden_dropout
  • Discrimination: 0.3 > 0.2 > 0.4
  • Insertion: 0.3 > 0.4 > 0.2

Best model

• Same model performs best in both tasks

"hparams": {
    "input_dropout": 0.5,
    "hidden_layers": 2,
    "hidden_dropout": 0.3,
    "margin": 6.0,
    "weight_decay": 0.0,
    "dpout_model": 0.0
}

Insertion validation

Results for best 20 are taken

Scores

• Discrimination

  Description	Value
  mean	0.926410
  std	0.001753
  min	0.923900
  max	0.930300
  25%	0.925100
  50%	0.925900
  75%	0.927425

• Insertion

  Description	Value
  mean	0.3041
  std	0.0029
  min	0.301
  max	0.3124
  25%	0.3024
  50%	0.3033
  75%	0.305

Trends

• input_dropout: 0.5
• hidden_layers: 1 > 2
• hidden_dropout: 0.4 > 0.2 > 0.3
• margin: 6.0 > 5.0 > 4.0
• weight_decay: 0.0
• dpout_model: 0.05 > 0.1 > 0.0

Best model

• Same model performs best in both tasks

"hparams": {
    "input_dropout": 0.5,
    "hidden_layers": 1,
    "hidden_dropout": 0.4,
    "margin": 6.0,
    "weight_decay": 0.0,
    "dpout_model": 0.05
}

Final Results

(After Hyperparameter tuning)

"Discr" refers to score for Discrimination task and "Ins" refers to score for Insertion task. "Avg" refers to average score for both tasks.

Validation Task	Bidirectional	Output function	Encoder	Discr	Ins	Avg
discrimination	False	None	average_glove	0.9204	0.3028	0.6116
discrimination	False	None	sbert	0.9232	0.3139	0.61855
discrimination	False	tanh	average_glove	0.0952	0.8153	0.45525
discrimination	False	tanh	sbert	0.2989	0.2798	0.28935
discrimination	False	sigmoid	average_glove	0.6148	0.4915	0.55315
discrimination	False	sigmoid	sbert	0.2711	0.3196	0.29535
discrimination	True	None	average_glove	0.9259	0.3091	0.61749
discrimination	True	None	sbert	0.9268	0.313	0.6199
discrimination	True	tanh	average_glove	0.0001	1.0	0.50005
discrimination	True	tanh	sbert	0.807	0.3134	0.5602
discrimination	True	sigmoid	average_glove	0.713	0.3809	0.54695
discrimination	True	sigmoid	sbert	0.8113	0.2484	0.52985
insertion	False	None	average_glove	0.9185	0.2993	0.6089
insertion	False	None	sbert	0.9226	0.3164	0.61949
insertion	False	tanh	average_glove	0.0	1.0	0.5
insertion	False	tanh	sbert	0.5872	0.4079	0.49755
insertion	False	sigmoid	average_glove	0.0581	0.9511	0.50459
insertion	False	sigmoid	sbert	0.6704	0.1794	0.4249
insertion	True	None	average_glove	0.9191	0.2929	0.606
insertion	True	None	sbert	0.9245	0.3239	0.6242
insertion	True	tanh	average_glove	0.7525	0.3163	0.5344
insertion	True	tanh	sbert	0.7852	0.3211	0.55315
insertion	True	sigmoid	average_glove	0.6717	0.3519	0.5118
insertion	True	sigmoid	sbert	0.8125	0.3122	0.56235

References

A Cross-Domain Transferable Neural Coherence Model

Dataset

• [Link] to download the dataset.
• Name of dataset is wsj_bigram
• Save in ./data/ directory relative to run_model.py file.

Checkpoints

[Link] for checkpoints of the trained baseline models.

Code Structure

• Main model is in models/coherence_models.py
• The NN used in the model is at models/discriminator_models.py