Method and article of manufacture for automated generation of titles and descriptions for electronic commerce products
A method comprises collecting browsing activity data of at least one user in connection with at least one electronic commerce item, and collecting feature data of the at least one electronic commerce item. In the method, the browsing activity data and the feature data are analyzed using one or more machine learning models. At least one of a title and a description for the at least one electronic commerce item is generated based on the analysis, and are displayed on an interface for viewing by the at least one user.
1. A method, comprising:
collecting browsing activity data of at least one user in connection with at least one electronic commerce item, wherein the browsing activity data comprises purchase status data for the at least one electronic commerce item, the purchase status data comprising information corresponding to whether the at least one electronic commerce item was added to a virtual shopping cart without being purchased;
collecting feature data of the at least one electronic commerce item;
analyzing the browsing activity data and the feature data using one or more machine learning models, wherein the analyzing comprises:
inputting the browsing activity data into a first context encoder of a plurality of context encoders as a first plurality of sentences, wherein the first context encoder generates a first plurality of sentence representations based at least in part on the first plurality of sentences and extracts a first plurality of features from the first plurality of sentence representations, and wherein the first context encoder comprises a first plurality of convolutional neural networks;
computing a first plurality of convoluted features for respective ones of the first plurality of features from the first plurality of sentence representations and merging the first plurality of convoluted features, wherein an output of the first context encoder is based at least in part on the merged first plurality of convoluted features;
inputting the feature data into a second context encoder of the plurality of context encoders as a second plurality of sentences, wherein the second context encoder generates a second plurality of sentence representations based at least in part on the second plurality of sentences and extracts a second plurality of features from the second plurality of sentence representations, and wherein the second context encoder comprises a second plurality of convolutional neural networks;
computing a second plurality of convoluted features for respective ones of the second plurality of features from the second plurality of sentence representations and merging the second plurality of convoluted features, wherein an output of the second context encoder is based at least in part on the merged second plurality of convoluted features;
computing respective hidden states based at least in part on the output of the first context encoder and the output of the second context encoder;
computing respective context representations by inputting the respective hidden states and computing respective attention weights; and
merging the respective context representations to generate a combined output;
generating at least one of a title and a description for the at least one electronic commerce item based at least in part on the combined output;
training the one or more machine learning models with at least a first training dataset comprising historical user actions and one or more tags from features of a plurality of electronic commerce items;
determining a plurality of keywords used in the at least one of the generated title and the generated description;
creating a plurality of checkpoints for at least a subset of the used plurality of keywords;
re-training the one or more machine learning models with at least a second training dataset comprising the created plurality of checkpoints,
the plurality of checkpoint comprising one or more weights for the one or more machine learning models, the one or more weights being saved based at least in part on whether accuracy of the one or more machine learning models increases; and
displaying the at least one of the generated title and the generated description on an interface for viewing by the at least one user; and
wherein the steps of the method are executed by a processing device operatively coupled to a memory.
2. The method of claim 1 , wherein the browsing activity data and the feature data are collected from at least one of a cache of a user device of the at least one user and one or more servers corresponding to one or more vendors of the at least one electronic commerce item.
3. The method of claim 1 , wherein the browsing activity data further comprises at least one of one or more navigation paths of the at least one user, one or more searches performed by the at least one user for the at least one electronic commerce item, one or more keywords used by the at least one user in the one or more searches.
4. The method of claim 3 , wherein the at least one of the generated title and the generated description comprises the one or more keywords used by the at least one user in the one or more searches.
5. The method of claim 1 , wherein the feature data comprises at least one of a description and one or more attributes of the at least one electronic commerce item retrieved from one or more servers corresponding to one or more vendors of the at least one electronic commerce item.
6. The method of claim 1 , wherein the one or more machine learning models comprises a deep neural forest network.
7. The method of claim 6 , wherein the deep neural forest network is embedded with a convolutional long short-term memory (C-LSTM) network.
8. The method of claim 1 , wherein the generating of the at least one of the title and the description comprises performing at least one of text pattern modelling, text pattern matching and text generation.
9. The method of claim 1 , wherein the browsing activity data further comprises one or more searches performed by the at least one user for the at least one electronic commerce item, and wherein the analyzing comprises at least one of removing stop words from the one or more searches, tokenizing most frequently used words of the one or more searches and stemming the most frequently used words of the one or more searches.
10. The method of claim 1 , wherein the feature data comprises one or more attributes of the at least one electronic commerce item, wherein the analyzing comprises predicting keywords corresponding to the one or more attributes of the at least one electronic commerce item.
11. The method of claim 10 , wherein the generating of at least one of the title and the description comprises:
generating a sequence of the predicted keywords; and
performing a grammar correction of the generated sequence.
12. The method according to claim 1 , further comprising:
collecting additional browsing activity data of at least one other user in connection with the at least one electronic commerce item;
analyzing the additional browsing activity data and the feature data using the one or more machine learning models; and
generating at least one of an additional title and an additional description for the at least one electronic commerce item based on the analysis of the additional browsing activity data and the feature data;
wherein the at least one of the generated additional title and the generated additional description corresponding to the at least one other user are different from the at least one of the generated title and the generated description corresponding to the at least one user.
13. The method according to claim 12 , further comprising:
determining a first group of keywords used in the at least one of the generated title and the generated description corresponding to the at least one user;
determining a second group of keywords used in the at least one of the generated additional title and the generated additional description corresponding to the at least one other user;
generating a first word cloud corresponding to the first group of keywords; and
generating a second word cloud corresponding to the second group of keywords.
14. An article of manufacture comprising a non-transitory processor-readable storage medium having stored therein program code of one or more software programs, wherein the program code when executed by at least one processing device causes said at least one processing device to perform the steps of:
collecting browsing activity data of at least one user in connection with at least one electronic commerce item, wherein the browsing activity data comprises purchase status data for the at least one electronic commerce item, the purchase status data comprising information corresponding to whether the at least one electronic commerce item was added to a virtual shopping cart without being purchased;
collecting feature data of the at least one electronic commerce item;
analyzing the browsing activity data and the feature data using one or more machine learning models, wherein the analyzing comprises:
inputting the browsing activity data into a first context encoder of a plurality of context encoders as a first plurality of sentences, wherein the first context encoder generates a first plurality of sentence representations based at least in part on the first plurality of sentences and extracts a first plurality of features from the first plurality of sentence representations, and wherein the first context encoder comprises a first plurality of convolutional neural networks;
computing a first plurality of convoluted features for respective ones of the first plurality of features from the first plurality of sentence representations and merging the first plurality of convoluted features, wherein an output of the first context encoder is based at least in part on the merged first plurality of convoluted features;
inputting the feature data into a second context encoder of the plurality of context encoders as a second plurality of sentences, wherein the second context encoder generates a second plurality of sentence representations based at least in part on the second plurality of sentences and extracts a second plurality of features from the second plurality of sentence representations, and wherein the second context encoder comprises a second plurality of convolutional neural networks;
computing a second plurality of convoluted features for respective ones of the second plurality of features from the second plurality of sentence representations and merging the second plurality of convoluted features, wherein an output of the second context encoder is based at least in part on the merged second plurality of convoluted features;
computing respective hidden states based at least in part on the output of the first context encoder and the output of the second context encoder;
computing respective context representations by inputting the respective hidden states and computing respective attention weights; and
merging the respective context representations to generate a combined output;
generating at least one of a title and a description for the at least one electronic commerce item based at least in part on the combined output;
training the one or more machine learning models with at least a first training dataset comprising historical user actions and one or more tags from features of a plurality of electronic commerce items;
determining a plurality of keywords used in the at least one of the generated title and the generated description;
creating a plurality of checkpoints for at least a subset of the used plurality of keywords;
re-training the one or more machine learning models with at least a second training dataset comprising the created plurality of checkpoints,
the plurality of checkpoints comprising one or more weights for the one or more machine learning models, the one or more weights being saved based at least in part on whether accuracy of the one or more machine learning models increases; and
displaying the at least one of the generated title and the generated description on an interface for viewing by the at least one user.
15. The article of manufacture of claim 14 , wherein the program code further causes said at least one processing device to perform the steps of:
collecting additional browsing activity data of at least one other user in connection with the at least one electronic commerce item;
analyzing the additional browsing activity data and the feature data using the one or more machine learning models; and
generating at least one of an additional title and an additional description for the at least one electronic commerce item based on the analysis of the additional browsing activity data and the feature data;
wherein the at least one of the generated additional title and the generated additional description corresponding to the at least one other user are different from the at least one of the generated title and the generated description corresponding to the at least one user.
16. The article of manufacture of claim 15 , wherein the program code further causes said at least one processing device to perform the steps of:
determining a first group of keywords used in the at least one of the generated title and the generated description corresponding to the at least one user;
determining a second group of keywords used in the at least one of the generated additional title and the generated additional description corresponding to the at least one other user;
generating a first word cloud corresponding to the first group of keywords; and
generating a second word cloud corresponding to the second group of keywords.
17. The article of manufacture of claim 14 , wherein the feature data comprises one or more attributes of the at least one electronic commerce item, wherein the analyzing comprises predicting keywords corresponding to the one or more attributes of the at least one electronic commerce item.
18. The article of manufacture of claim 17 , wherein the generating of at least one of the title and the description comprises:
generating a sequence of the predicted keywords; and
performing a grammar correction of the generated sequence.
19. The article of manufacture of claim 14 , wherein the browsing activity data and the feature data are collected from at least one of a cache of a user device of the at least one user and one or more servers corresponding to one or more vendors of the at least one electronic commerce item.
20. The article of manufacture of claim 14 , wherein the generating of the at least one of the title and the description comprises performing at least one of text pattern modelling, text pattern matching and text generation.