Automatically tagging topics in posts during composition thereof
An online system predicts topics for content items. The online system provides one or more topic labels for a user to apply concurrently while a user is composing a post, in response to requests periodically received from the user's device. A request includes information such as content composed by the user and contextual information. The online system employs machine learning techniques to analyze content composed by a user and contextual information thereby to predict topic labels. Different machine learning models for classifying individual topic labels, identifying relevant topic labels, and/or detecting changes in existing topic predictions are developed. Some machine learning models predict topics for full content and some predict topics for partial content. The online system trains the machine learning models to ensure accurate topic predictions are provided timely. The online system employs various machine learning model training methods such as active training and gradient training.
1. A computer-implemented method comprising:
receiving, at an online system, a plurality of predetermined topic labels, each predetermined topic label corresponding to one or more topics and configured to tag content items to identify the content items by topics;
receiving, at the online system, a request for a topic prediction of a content item from a client device, the request being sent concurrently when the content item is being composed, and the request comprising content of the content item and contextual information describing a context of the content item, the contextual information including a composition stage based on a time period spent composing the content item;
predicting one or more topics of the content item in response to the request, the predicting comprising:
providing the content and the contextual information to a machine learning model corresponding to the composition stage included in the contextual information selected from a plurality of machine learning models each machine learning model corresponding ton one or more composition stages, the one or more machine learning models determining a likelihood of a predetermined topic label being relevant to the content item, and
selecting a set of predetermined topic labels from the predetermined topic labels, a likelihood of a selected predetermined topic label being relevant to the content item being greater than a threshold likelihood; and
providing for display, by the online system to the client device, the set of predetermined topic labels.
2. The computer-implemented method of claim 1 , wherein the one or more machine learning models comprise a classifier corresponding to a predetermined topic label, the classifier configured to classify a likelihood of the corresponding predetermined topic label being relevant to the content item.
3. The computer-implemented method of claim 2 , wherein the classifier is a binary classifier.
4. The computer-implemented method of claim 1 , wherein the one or more machine learning models comprise a ensemble classifier corresponding to multiple topic labels, the ensemble classifier configured to classify likelihoods of the multiple topic labels being relevant to the content item.
5. The computer-implemented method of claim 1 , wherein the contextual information comprises an existing topic prediction, and the one or more machine learning models comprise a machine learning model configured to detect if the existing topic prediction will change.
6. The computer-implemented method of claim 1 , wherein the predicting further comprises: providing the content and the contextual information to a machine learning model
configured to detect if an existing topic prediction will change; and wherein the content and the contextual information is provided to the one or more
machine learning models responsive to a determination that an existing topic prediction will change.
7. The computer-implemented method of claim 1 , wherein the request for topics is received periodically.
8. The computer-implemented method of claim 1 , wherein the content includes all content input by the user since the user started composing the content item.
9. The computer-implemented method of claim 1 , wherein the content includes content update by the user since a most recent request was sent.
10. The computer-implemented method of claim 1 , wherein further comprising determining the threshold likelihood according to a score function.
11. The computer-implemented method of claim 1 , further comprising: training the one or more machine learning models by training data, the training data comprising content items labeled with topic labels.
12. The computer-implemented method of claim 11 , wherein the training comprises:
identifying disagreements between output topic predictions of multiple classifiers; identifying training content items and associated contextual information corresponding to
the output topic predictions having disagreements; and
providing additional training data to re-train the multiple classifiers, the additional training data comprising content items and contextual information labeled with topic labels and sharing one or more features as the training content and contextual information.
13. The computer-implemented method of claim 11 , wherein the training comprises: for a machine learning model:
identifying output topic predictions associated with confidence scores lower than a threshold confidence score;
identifying training content items and associated contextual information corresponding to the identified output topic predictions;
identifying a feature of the training content item and associated contextual information that if labeled causes the output topic predictions to change the most; and
providing additional training data to re-train the machine learning model, the additional training data comprising content items and contextual information labeled with topic labels and having the identified feature.
14. The computer-implemented method of claim 1 , wherein the contextual information comprises at least one of user information related to the user, a user action of the user, content information of the content item, composer information of a composer that is used by the user to compose the content item.
15. The computer-implemented method of claim 14 , wherein the user information related to the user comprises at least one of an identity of the user, a current location of the user, an interest of the user, and a group associated with the user.
16. The computer-implemented method of claim 14 , wherein the content information comprises a character count, a word count, a composing order, an existing topic label, a topic label that has been applied, and a topic label that has been removed.
17. The computer-implemented method of claim 14 , wherein the content information of the content item comprises metadata associated with a media content included the content.
18. The computer-implemented method of claim 14 , wherein the user action comprises at least one of an interaction of the user with a topic label and incorporating another content into the content.
19. The computer-implemented method of claim 14 , further comprising providing, by the online system to the client device the composer for the user to compose the content item, wherein the composer information comprises a composer location of the composer within the online system.
20. The computer-implemented method of claim 1 , further comprising retrieving additional contextual information and providing the additional contextual information to the one or more machine learning models.
21. The computer-implemented method of claim 1 , further comprising ranking the set of predetermined topic labels according to associated likelihood of a selected predetermined topic label being relevant to the content item, wherein the set of predetermined topic labels are presented according a ranked order.
22. A non-transitory computer readable storage medium having instructions encoded thereon that, when executed by a processor, cause the processor to perform:
receiving, at an online system, a plurality of predetermined topic labels, each predetermined topic label corresponding to one or more topics and configured to tag content items to identify the content items by topics;
receiving, at the online system, a request for a topic prediction of a content item from a client device, the request being sent concurrently when the content item is being composed, and the request comprising content of the content item and contextual information describing a context of the content item, the contextual information including a composition stage based on a time period spent composing the content item;
predicting one or more topics of the content item in response to the request, the predicting comprising:
providing the content and the contextual information to a machine learning model corresponding to the composition stage included in the contextual information selected from a plurality of machine learning models each machine learning model corresponding ton one or more composition stages, the one or more machine learning models determining a likelihood of a predetermined topic label being relevant to the content item, and
selecting a set of predetermined topic labels from the predetermined topic labels, a likelihood of a selected predetermined topic label being relevant to the content item being greater than a threshold likelihood; and
providing for display, by the online system to the client device, the set of predetermined topic labels.