Machine-learning based generation of text style variations for digital content items
An online system generates a set of content item variations for a reference content item that include different styles of text for the content item. The different styles of text are generated by applying machine-learned style transfer models, for example, neural network based models to reference text of the reference content item. The text variations retain the textual content of the reference text but are synthesized with different styles. The online system can provide the content item variations to users on an online experimental platform to collect user interaction information that may indicate how users respond to different styles of text. The online system or the content providers can effectively target users with content items that include the style of text the users respond to based on the collected information.
1. A computer-implemented method, comprising:
receiving, by an online system, a request including a reference content item, wherein the reference content item includes an image and a reference text that invites a viewer of the reference content item to perform a desired response;
applying, by the online system, a machine-learned model to the reference text to generate a set of text variations corresponding to a set of text styles, wherein each text variation stylizes the textual content of the reference text in a respective text style in the set of text styles, wherein the machine-learned model includes an encoder and a set of decoders each assigned to a respective text style, and applying the machine-learned model further comprises:
applying the encoder to an encoded version of the reference text to generate a text content embedding, the text content embedding characterizing textual content of the reference text, and
for each decoder, applying the decoder to the text content embedding to generate a text variation that styles the reference text in the respective text style assigned to the decoder;
generating, by the online system, a set of content item variations, each content item variation including the image and a respective text variation; and
providing, to one or more client devices, the set of content item variations for display on one or more client devices.
2. The computer-implemented method of claim 1 , further comprising:
obtaining, by the online system, information on whether users of the one or more client devices interacted with the set of content item variations, and
using the obtained information to determine targeting criteria for targeting the set of content item variations.
3. The computer-implemented method of claim 1 , wherein the encoder is configured as a recurrent neural network (RNN) architecture configured to receive a sequence of tokens for the reference text to generate the text content embedding.
4. The computer-implemented method of claim 1 , wherein each decoder is configured as a recurrent neural network (RNN) architecture configured to receive the text content embedding and generate a sequence of tokens as the text variation for the decoder.
5. The computer-implemented method of claim 1 , wherein the respective text variation for a content item variation is placed at a position of the reference text on the image of the reference content item.
6. The computer-implemented method of claim 1 , wherein a set of parameters of the machine-learned model are trained in conjunction with an adversarial network, the adversarial network coupled to receive an output generated by applying the machine-learned model to an input text, and generate one or more likelihood predictions that indicate whether the input text is in one or more text styles.
7. The computer-implemented method of claim 1 , further comprising:
training, by the online system, a plurality of machine-learned models, each machine-learned model for use with requests associated with a respective affiliation, and
responsive to receiving the request, identifying an affiliation associated with the request and selecting a machine-learned model for use based on the identified affiliation.
8. The computer-implemented method of claim 7 , wherein the respective affiliation is a brand of a product advertised in the reference content item or a manufacturer of the product advertised in the reference content item.
9. The computer-implemented method of claim 1 , wherein the reference content item is a sponsored content item describing a product.
10. A non-transitory computer-readable storage medium storing computer program instructions executable to perform operations, the operations comprising:
receiving, by an online system, a request including a reference content item, wherein the reference content item includes an image and a reference text that invites a viewer of the reference content item to perform a desired response;
applying, by the online system, a machine-learned model to the reference text to generate a set of text variations corresponding to a set of text styles, wherein each text variation stylizes the textual content of the reference text in a respective text style in the set of text styles, wherein the machine-learned model includes an encoder and a set of decoders each assigned to a respective text style, and applying the machine-learned model further comprises:
applying the encoder to an encoded version of the reference text to generate a text content embedding, the text content embedding characterizing textual content of the reference text, and
for each decoder, applying the decoder to the text content embedding to generate a text variation that styles the reference text in the respective text style assigned to the decoder;
generating, by the online system, a set of content item variations, each content item variation including the image and a respective text variation; and
providing, to one or more client devices, the set of content item variations for display on one or more client devices.
11. The non-transitory computer-readable storage medium of claim 10 , further comprising:
obtaining, by the online system, information on whether users of the one or more client devices interacted with the set of content item variations, and
using the obtained information to determine targeting criteria for targeting the set of content item variations.
12. The non-transitory computer-readable storage medium of claim 11 , wherein the encoder is configured as a recurrent neural network (RNN) architecture configured to receive a sequence of tokens for the reference text to generate the text content embedding.
13. The non-transitory computer-readable storage medium of claim 11 , wherein each decoder is configured as a recurrent neural network (RNN) architecture configured to receive the text content embedding and generate a sequence of tokens as the text variation for the decoder.
14. The non-transitory computer-readable storage medium of claim 10 , wherein the respective text variation for a content item variation is placed at a position of the reference text on the image of the reference content item.
15. The non-transitory computer-readable storage medium of claim 10 , wherein a set of parameters of the machine-learned model are trained in conjunction with an adversarial network, the adversarial network coupled to receive an output generated by applying the machine-learned model to an input text, and generate one or more likelihood predictions that indicate whether the input text is in one or more text styles.
16. The non-transitory computer-readable storage medium of claim 10 , further comprising:
training, by the online system, a plurality of machine-learned models, each machine-learned model for use with requests associated with a respective affiliation, and
responsive to receiving the request, identifying an affiliation associated with the request and selecting a machine-learned model for use based on the identified affiliation.
17. The non-transitory computer-readable storage medium of claim 16 , wherein the respective affiliation is a brand of a product advertised in the reference content item or a manufacturer of the product advertised in the reference content item.
18. The non-transitory computer-readable storage medium of claim 10 , wherein the reference content item is a sponsored content item describing a product.
19. A computer system comprising:
a computer processor; and
a non-transitory computer-readable storage medium storing computer program instructions executable to perform operations, the operations comprising:
receiving, by an online system, a request including a reference content item, wherein the reference content item includes an image and a reference text that invites a viewer of the reference content item to perform a desired response;
applying, by the online system, a machine-learned model to the reference text to generate a set of text variations corresponding to a set of text styles, wherein each text variation stylizes the textual content of the reference text in a respective text style in the set of text styles, wherein the machine-learned model includes an encoder and a set of decoders each assigned to a respective text style, and applying the machine-learned model further comprises:
applying the encoder to an encoded version of the reference text to generate a text content embedding, the text content embedding characterizing textual content of the reference text, and
for each decoder, applying the decoder to the text content embedding to generate a text variation that styles the reference text in the respective text style assigned to the decoder;
generating, by the online system, a set of content item variations, each content item variation including the image and a respective text variation; and
providing, to one or more client devices, the set of content item variations for display on one or more client devices.