Automatic conversation creator for news

1. A computing device comprising:

a processor;

a display device; and

a computer-readable storage medium in communication with the processor, the computer-readable storage medium having computer-executable instructions stored thereupon which, when executed by the processor, cause the computing device to:

identify lexical-level features for a plurality of pairs of comments and replies previously posted to one or more content item webpages, the lexical-level features including identifying a cosine similarity between two vectors, a first of the two vectors representing a comment of one of the plurality of pairs of comments and replies and a second of the two vectors representing a reply of the one of the plurality of pairs of comments and replies, the two vectors associated with a term frequency of a term in the one of the pair of the plurality of replies and comments;

receive information associated with a content item produced by a source system, the content item being accessible to other computing devices via a network;

in response to receiving the information, identify a start comment;

in response to identifying the start comment, create a reply to the start comment using at least a ranking model, the ranking model created at least in part on the identified lexical-level features; and

output the start comment and the reply.

2. The computing device of claim 1 , wherein the identifying the start comment comprises:

determining from a plurality of comments an amount of likes for the comments of the plurality of the comments; and

identifying the start comment as the comment of the plurality of comments having a greatest amount of likes.

3. The computing device according to claim 1 , wherein the creating the reply comprises:

identifying a plurality of pairs, the plurality of pairs comprises the start comment and one of a plurality of replies;

ranking the plurality of pairs in response to execution of the ranking model; and

identifying a reply in the pair with a highest rank as the reply to the start comment.

4. The computing device according to claim 1 , wherein the computer-readable storage medium has further computer-executable instructions stored thereon that cause the computing device to:

identify entity-level features for the plurality of pairs;

identify topic vectors for the plurality of pairs;

identify one or more quality features for the plurality of pairs;

create feature vectors for the plurality of pairs based at least in part on the lexical-level features, entity-level features, topic vectors and one or more quality features; and create the ranking model based at least in part on the feature vectors.

5. The computing device according to claim 1 , wherein the computer-readable storage medium has further computer-executable instructions stored thereon that cause the computing device to:

receive a comment from the computing device;

in response to receiving the comment from the computing device, identify a reply to the comment from the computing device based at least in part on the ranking model; and

output the reply identified by the ranking model.

6. The computing device according to claim 1 , wherein the computer-readable storage medium has further computer-executable instructions stored thereon that cause the computing device to:

apply weight values to the plurality of pairs; and

create a classifier based on the weight values.

7. The computing device according to claim 6 , wherein the computer-readable storage medium has further computer-executable instructions stored thereon that cause the computing device to create the classifier using a support vector machine.

8. The computing device according to claim 1 , wherein each of the two vectors is represented as a tf-idf vector, wherein tf is representative of a term frequency of a single word and idf is representative of an inverse document frequency of the single word.

9. The computing device according to claim 1 , wherein the lexical-level features additionally include a translation model score representative of statistical word matching in the comment and reply of the one of the plurality of pairs of comments and replies.

10. A computer-implemented method comprising:

identifying lexical-level features for a plurality of pairs of comments and replies previously posted to one or more content item webpages, the lexical-level features including a cosine similarity between two vectors, a first of the two vectors representing a comment of one of the plurality of pairs of comments and replies and a second of the two vectors representing a reply of the one of the plurality of pairs of comments and replies, the two vectors associated with a term frequency of a term in the one of the pair of the plurality of replies and comments;

receiving, at a computing device, information associated with a content item produced by a source system, the content item being accessible to other computing devices via a network;

outputting at least a portion of the content item;

in response to receiving the information, generating a start comment and identifying a reply to the start comment based at least in part on a ranking model, the ranking model created at least in part on the identified lexical-level features; and

outputting the start comment and the reply concurrently with the content item.

11. The computer-implemented method of claim 10 , wherein the identifying the start comment comprises:

determining from a plurality of comments an amount of likes for the comments of the plurality of the comments; and

identifying the start comment as the comment of the plurality of comments having a greatest amount of likes.

12. The computer-implemented method according to claim 10 , wherein the identifying the reply comprises:

identifying a plurality of pairs, the plurality of pairs comprises the start comment and one of a plurality of replies;

ranking the plurality of pairs in response to execution of the ranking model; and

identifying a reply in the pair with a highest rank as the reply to the start comment.

13. The computer-implemented method according to claim 10 , further comprising training the ranking model by:

identifying entity-level features for the plurality of pairs;

identifying topic vectors for the plurality of pairs;

identifying one or more quality features for the plurality of pairs;

creating feature vectors for the plurality of pairs based at least in part on the identified lexical-level features, entity-level features, topic vectors and one or more quality features; and

creating the ranking model based at least in part on the created feature vectors.

14. The computer-implemented method according to claim 10 , wherein creating the ranking model comprises:

applying weight values to the plurality of pairs; and

creating a classifier based on the weight values.

15. The computer-implemented method according to claim 14 , further comprising creating the classifier using a support vector machine.

16. The computer-implemented method according to claim 10 , wherein the content item comprises a news article webpage or a posting webpage.

17. The computer-implemented method according to claim 10 , further comprising:

receiving a comment from the computing device;

in response to receiving the comment from the computing device, identifying a reply to the comment from the computing device based at least in part on the ranking model; and

outputting the reply identified by the ranking model.

18. A computer-readable storage medium having computer-executable instructions stored thereupon which, when executed by a computer, cause the computer to perform the method according to claim 10 .

19. The computer-implemented method according to claim 10 , wherein each of the two vectors is represented as a tf-idf vector, wherein tf is representative of a term frequency of a single word and idf is representative of an inverse document frequency of the single word.

20. A method of creating a reply model, the method including:

identifying lexical-level features for a plurality of pairs of comments and replies previously posted to one or more content item webpages, the lexical-level features including identifying a cosine similarity between two vectors, a first of the two vectors representing a comment of one of the plurality of pairs of comments and replies and a second of the two vectors representing a reply of the one of the plurality of pairs of comments and replies, the two vectors associated with a term frequency of a term in the one of the pair of the plurality of replies and comments;

identify entity-level features for the plurality of pairs;

identify topic vectors for the plurality of pairs;

identify one or more quality features for the plurality of pairs;

create a plurality of pairs of feature vectors, each pair of feature vectors including a feature vector for each comment and a feature vector for each reply for the plurality of pairs of comments and replies based at least in part on the lexical-level features, entity-level features, topic vectors and one or more quality features;

assign a weight value to each pair of feature vectors, the weight value indicating a likelihood that the reply of the comment and reply pair is an appropriate response to the comment of the comment and reply pair;

train a reply model based on the weighted pairs of feature vectors.