Pre-training techniques for entity extraction in low resource domains
Embodiments of the present invention provide systems, methods, and computer storage media for pre-training entity extraction models to facilitate domain adaptation in resource-constrained domains. In an example embodiment, a first machine learning model is used to encode sentences of a source domain corpus and a target domain corpus into sentence embeddings. The sentence embeddings of the target domain corpus are combined into a target corpus embedding. Training sentences from the source domain corpus within a threshold of similarity to the target corpus embedding are selected. A second machine learning model is trained on the training sentences selected from the source domain corpus.
1. A method comprising:
using a first machine learning model to encode sentences of a source domain corpus and a target domain corpus into sentence embeddings;
combining the sentence embeddings of the target domain corpus into a target corpus embedding;
selecting training sentences from the source domain corpus within a threshold of similarity to the target corpus embedding; and
training a second machine learning model on the training sentences selected from the source domain corpus.
2. The method of claim 1 , wherein selecting the training sentences within the threshold of similarity comprises determining cosine similarity between the target corpus embedding and the sentence embeddings of the sentences from the source domain corpus.
3. The method of claim 1 , wherein selecting the training sentences within the threshold of similarity comprises using the target corpus embedding as a query vector to identify a set of the sentence embeddings of the source domain corpus within the threshold of similarity to the target corpus embedding.
4. The method of claim 1 , wherein the first machine learning model is trained on the target domain corpus prior to using the first machine learning model to encode the sentences.
5. The method of claim 1 ,
wherein a first instance of the first machine learning model is trained on the target domain corpus;
the method further comprising copying the first instance to create a second instance of the first machine learning model;
wherein the second instance of the first machine learning model is trained on a labeled subset of the target domain corpus;
the method further comprising using the first and second instances of the first machine learning model to encode the sentences.
6. The method of claim 1 , wherein the first machine learning model and the second machine learning model include a same type of language model.
7. The method of claim 1 , further comprising:
using the second machine learning model to extract one or more token values of one or more corresponding entity classes from one or more contracts; and
causing a document dashboard to present a representation of the one or more token values and the one or more corresponding entity classes.
8. The method of claim 1 , further comprising:
using the second machine learning model to extract one or more values of one or more payment terms from one or more contracts; and
causing a document dashboard to present a representation of the one or more token values and the one or more corresponding entity classes.
9. The method of claim 1 , wherein the second machine learning model is trained on the training sentences using masked language modeling.
10. The method of claim 1 , wherein after training the second machine learning model on the training sentences selected from the source domain corpus, the second machine learning model is trained on a labeled subset of the target domain corpus.
11. A non-transitory computer-readable medium storing computer-executable instructions that, when executed by a processing device, cause the processing device to:
use a first machine learning model to predict entities from sentences of a source domain corpus;
select as training sentences a set of the sentences that have a maximum number of the predicted entities; and
train a second machine learning model on the training sentences selected from the source domain corpus.
12. The non-transitory computer-readable medium of claim 11 , wherein the computer-executable instructions, when executed by the processing device, further cause the processing device to train the second machine learning model on a labeled subset of a target domain corpus after training the second machine learning model on the training sentences selected from the source domain corpus.
13. The non-transitory computer-readable medium of claim 11 , wherein the computer-executable instructions, when executed by the processing device, further cause the processing device to train the first machine learning model on a target domain corpus using unsupervised learning prior to using the first machine learning model to predict the entities.
14. The non-transitory computer-readable medium of claim 11 , wherein the computer-executable instructions, when executed by the processing device, further cause the processing device to train the first machine learning model on a labeled subset of a target domain corpus prior to using the first machine learning model to predict the entities.
15. The non-transitory computer-readable medium of claim 11 , wherein the computer-executable instructions, when executed by the processing device, further cause the processing device to train the first machine learning model on a target domain corpus using unsupervised learning and on a labeled subset of the target domain corpus using supervised learning prior to using the first machine learning model to predict the entities.
16. The non-transitory computer-readable medium of claim 11 , wherein training the second machine learning model on the training sentences selected from the source domain corpus comprises using masked language modeling.
17. The non-transitory computer-readable medium of claim 11 , wherein the computer-executable instructions, when executed by the processing device, further cause the processing device to:
use the second machine learning model to extract one or more token values of one or more corresponding entity classes from one or more contracts; and
cause a document dashboard to present a representation of the one or more token values and the one or more corresponding entity classes.
18. The non-transitory computer-readable medium of claim 11 , wherein the computer-executable instructions, when executed by the processing device, further cause the processing device to:
use the second machine learning model to extract one or more values of one or more payment terms from one or more contracts; and
trigger a payment according to the one or more values of the one or more payment terms.
19. A system comprising:
a memory component; and
one or more hardware processors coupled to the memory component, the one or more hardware processors to perform operations comprising:
receiving a first training sentence from a target domain corpus into a machine learning model; and
training the machine learning model based on the first training sentence using unsupervised learning to generate a trained machine learning model that is used to select training sentences from a source domain corpus based on predicted entities from sentences of the source domain corpus.
20. The system of claim 19 , the operations further comprising training the machine learning model on a labeled subset of the target domain corpus using supervised learning to generate the trained machine learning model.