PSYCHIC

PSYCHIC (Pre-trained SYmbolic CHecker In Context) is a model that is a fine-tuned version of distilbert-base-uncased on the DBLP-QuAD dataset. It achieves the following results on the evaluation set:

• Loss: 0.0000

Model description

The model is trained to learn specific tokens from a question and its context to better determine the answer from the context. It is fine-tuned on the Extractive QA task from which it should return the answer to a knowledge graph question in the form of a SPARQL query. The advantage of PSYCHIC is that it leverages neuro-symbolic capabilities to validate query structures as well as LLM capacities to learn from context tokens.

Intended uses & limitations

This model is intended to be used with a question-context pair to determine the answer in the form of a SPARQL query.

Training and evaluation data

The DBLP-QuAD dataset is used for training and evaluation.

Example

Here's an example of the model capabilities:

• input:

  • question: Was the paper 'Stabilizing Client/Server Protocols without the Tears' not not published by Mohamed G. Gouda?
  • context: [CLS] DOUBLE_NEGATION [SEP] TC51 [SEP] sparql: ASK { https://dblp.org/rec/conf/forte/Gouda95 https://dblp.org/rdf/schema#authoredBy https://dblp.org/pid/g/MohamedGGouda } [SEP] [https://dblp.org/pid/g/MohamedGGouda, https://dblp.org/rec/conf/forte/Gouda95]

• output: sparql: ASK { https://dblp.org/rec/conf/forte/Gouda95 https://dblp.org/rdf/schema#authoredBy https://dblp.org/pid/g/MohamedGGouda } [SEP] [https://dblp.org/pid/g/MohamedGGouda, https://dblp.org/rec/conf/forte/Gouda95]

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss
0.001	1.0	1000	0.0001
0.0005	2.0	2000	0.0000
0.0002	3.0	3000	0.0000

Framework versions

• Transformers 4.33.1
• Pytorch 2.0.1+cu118
• Datasets 2.14.5
• Tokenizers 0.13.3