Poly-encoders: Transformer Architectures and Pre-training Strategies for Fast and Accurate Multi-sentence Scoring
Paper: https://arxiv.org/abs/1905.01969
The use of deep pre-trained bidirectional transformers has led to remarkable progress in anumber of applications (Devlin et al., 2018). For tasks that make pairwise comparisons between sequences, matching a given input with a corresponding label, two approaches are common: Cross-encoders performing full self-attention over the pair and Bi-encoders encoding the pair separately. The former often performs better, but is too slow for practical use. In this work, we develop a new transformer architecture, the Poly-encoder, that learns global rather than token level self-attention features. We perform a detailed comparison of all three approaches, including what pre-training and fine-tuning strategies work best. We show our models achieve state-of-the-art results on three existing tasks; that Poly-encoders are faster than Cross-encoders and more accurate than Bi-encoders; and that the best results are obtained by pre-training on large datasets similar to the downstream tasks.
Below we give details about available code and models:
- pretrained transformers on Reddit and Wikipedia + Toronto Books that can be used as a base for pretraining
- pretrained models fine tuned on ConvAI2 for the bi-encoder and poly-encoder The polyencoder scores 89+ on convai2 valid set and is fast enough to interact in real time with 100k candidates (which are provided.)
Run this command: (assumes your model zoo is in the default ./data/models)
parlai interactive -m transformer/polyencoder \
-mf zoo:pretrained_transformers/model_poly/model \
--encode-candidate-vecs true \
--eval-candidates fixed \
--fixed-candidates-path data/models/pretrained_transformers/convai_trainset_cands.txt
Example output:
Enter Your Message: your persona: i love to drink fancy tea.\nyour persona: i have a big library at home.\nyour persona: i'm a museum tour guide.\nhi how are you doing ?
[Polyencoder]: i am alright . i am back from the library .
Enter Your Message: oh, what do you do for a living?
[Polyencoder]: i work at the museum downtown . i love it there .
Enter Your Message: what is your favorite drink?
[Polyencoder]: i am more of a tea guy . i get my tea from china .
Note the polyencoder gives 89+ hits@1/20 on convai2, however, it expects data that is close to the dataset. If you do not include the multiple 'your persona: ...\n' at the beginning it will answer nonsense.
Execute this to train a biencoder scoring 86+ on Convai2 valid set (requires 8 x GPU 32GB., If you don't have this, reduce the batch size )
parlai train_model \
--init-model zoo:pretrained_transformers/bi_model_huge_reddit/model \
--batchsize 512 -t convai2 \
--model transformer/biencoder --eval-batchsize 6 \
--warmup_updates 100 --lr-scheduler-patience 0 \
--lr-scheduler-decay 0.4 -lr 5e-05 --data-parallel True \
--history-size 20 --label-truncate 72 --text-truncate 360 \
--num-epochs 10.0 --max_train_time 200000 -veps 0.5 -vme 8000 \
--validation-metric accuracy --validation-metric-mode max \
--save-after-valid True --log_every_n_secs 20 --candidates batch \
--dict-tokenizer bpe --dict-lower True --optimizer adamax \
--output-scaling 0.06 \
--variant xlm --reduction-type mean --share-encoders False \
--learn-positional-embeddings True --n-layers 12 --n-heads 12 \
--ffn-size 3072 --attention-dropout 0.1 --relu-dropout 0.0 --dropout 0.1 \
--n-positions 1024 --embedding-size 768 --activation gelu \
--embeddings-scale False --n-segments 2 --learn-embeddings True \
--share-word-embeddings False --dict-endtoken __start__ --fp16 True \
--model-file <YOUR MODEL FILE>
Execute this to train a poly-encoder scoring 89+ on Convai2 valid set (requires 8 x GPU 32GB., If you don't have this, reduce the batch size )
parlai train_model \
--init-model zoo:pretrained_transformers/poly_model_huge_reddit/model \
-t convai2 \
--model transformer/polyencoder --batchsize 256 --eval-batchsize 10 \
--warmup_updates 100 --lr-scheduler-patience 0 --lr-scheduler-decay 0.4 \
-lr 5e-05 --data-parallel True --history-size 20 --label-truncate 72 \
--text-truncate 360 --num-epochs 8.0 --max_train_time 200000 -veps 0.5 \
-vme 8000 --validation-metric accuracy --validation-metric-mode max \
--save-after-valid True --log_every_n_secs 20 --candidates batch --fp16 True \
--dict-tokenizer bpe --dict-lower True --optimizer adamax --output-scaling 0.06 \
--variant xlm --reduction-type mean --share-encoders False \
--learn-positional-embeddings True --n-layers 12 --n-heads 12 --ffn-size 3072 \
--attention-dropout 0.1 --relu-dropout 0.0 --dropout 0.1 --n-positions 1024 \
--embedding-size 768 --activation gelu --embeddings-scale False --n-segments 2 \
--learn-embeddings True --polyencoder-type codes --poly-n-codes 64 \
--poly-attention-type basic --dict-endtoken __start__ \
--model-file <YOUR MODEL FILE>
Execute this to train a cross-encoder scoring 90+ on Convai2 valid set (requires 8 x GPU 32GB., If you don't have this, reduce the batch size )
parlai train_model \
--init-model zoo:pretrained_transformers/cross_model_huge_reddit/model \
-t convai2 \
--model transformer/crossencoder --batchsize 16 --eval-batchsize 10 \
--warmup_updates 1000 --lr-scheduler-patience 0 --lr-scheduler-decay 0.4 \
-lr 5e-05 --data-parallel True --history-size 20 --label-truncate 72 \
--text-truncate 360 --num-epochs 12.0 --max_train_time 200000 -veps 0.5 \
-vme 2500 --validation-metric accuracy --validation-metric-mode max \
--save-after-valid True --log_every_n_secs 20 --candidates inline --fp16 True \
--dict-tokenizer bpe --dict-lower True --optimizer adamax --output-scaling 0.06 \
--variant xlm --reduction-type first --share-encoders False \
--learn-positional-embeddings True --n-layers 12 --n-heads 12 --ffn-size 3072 \
--attention-dropout 0.1 --relu-dropout 0.0 --dropout 0.1 --n-positions 1024 \
--embedding-size 768 --activation gelu --embeddings-scale False --n-segments 2 \
--learn-embeddings True --dict-endtoken __start__ \
--model-file <YOUR MODEL FILE>
To fine-tune on the DSTC7 task, simply replace -t convai2 with -t dstc7:DSTC7TeacherAugmentedSampled in the above commands.
To fine-tune on the UbuntuV2 task, simply replace -t convai2 with -t ubuntu in the above commands.