Response completion in social media

Patent metadata

Field	Value
Publication number	US-9380009-B2
Application number	US-201213547930-A
Country	US
Kind code	B2
Filing date	Jul 12, 2012
Priority date	Jul 12, 2012
Publication date	Jun 28, 2016
Grant date	Jun 28, 2016

Abstract

Embodiments are directed towards providing word-by-word message completion for an incomplete response message, wherein the response message is composed in response to a received stimulus message. The message completion is based on a Response Completion Model (RCM) that may model both the language used in the incomplete response message and the contextual information in the received stimulus message. The RCM may be determined based on conversational stimulus-response data including stimulus-response message pairs. The RCM may be a mixture model and include a generic response language model based on an N-gram model, a Stimulus Model based on a Selection Model or a Topic. Model, and a mixture parameter. In some embodiments, at least one candidate next word for the incomplete response message is determined based on the RCM. The at least one candidate next word may be selected and included in the incomplete response message. A complete response message may be generated and provided to a user.

First claim

What is claimed is:

1. A method for providing response message completion, comprising:
determining a response completion model (RCM) using a linear combination of a generic response language model (LM) and a stimulus model with a mixing parameter for a mixing model, the mixing parameter being based, at least in part, on a topic probability, the topic probability to be assigned a value based, at least in part, on a likelihood that a candidate topic is associated with a received stimulus message;
receiving a stimulus message (SM);
if an incomplete response message includes at least one preceding word, determining at least one candidate next word for the incomplete response message based, at least in part, on the RCM, the SM, and the at least one preceding word and based, at least in part, on estimated word frequencies in stimulus messages from a plurality of stimulus-response message pairs;
selecting at least one word from the at least one determined candidate next word;
including the selected at least one word within the incomplete response message; and
generating a complete response message that is based, at least in part, on the incomplete response message that includes the selected at least one word.

2. The method of claim 1 , further comprising if the incomplete response message is absent one or more words, determining the at least one candidate next word for the incomplete response message based, at least in part, on at least the RCM and the SM.

3. The method of claim 1 , wherein determining the RCM further comprises:
receiving stimulus-response parameters including at least a plurality of stimulus-response message pairs;
determining a dictionary containing a plurality of words;
determining the LM based, at least in part, on the dictionary and the stimulus-response parameters;
determining the stimulus model based, at least in part, on the dictionary and the stimulus response parameters;
determining the mixing parameter for the mixing model based, at least in part, on at least the stimulus-response parameters; and
determining the RCM based, at least in part, on the LM, the stimulus model, and the mixing parameter.

4. The method of claim 1 , wherein determining the RCM is based, at least in part, on a selection model and further comprises:
determining an estimated overall frequency of one or more words in at least one stimulus message from a plurality of stimulus-response message pairs that are also included in corresponding response message from the plurality of stimulus-response message pairs, wherein the plurality of stimulus-response message pairs are included in received stimulus-response parameters; wherein
the mixing parameter for the mixing model is additionally based, at least in part, on the estimated overall frequency of one or more words in the at least one stimulus message.

5. The method of claim 1 , wherein determining the RCM is based, at least in part, on a topic model and further comprises:
determining the topic model based, at least in part on received stimulus-response parameters;
determining the mixing parameter for the mixing model based, at least in part, on at least the determined topic model; and
determining the RCM based, at least in part, on at least the determined topic model, LM, and the mixing parameter.

6. The method of claim 1 , wherein determining at least one candidate next word, further comprises:
determining at least one LM probability for a plurality of words in a dictionary based, at least in part, on at least the at least one determined preceding word within the incomplete response message;
selecting at least one word in the stimulus message;
determining at least one stimulus probability for the plurality of dictionary words based, at least in part, on the selected at least one word in the stimulus message;
ranking the plurality of dictionary words based, at least in part, at least on the at least one LM probability, the at least one determined stimulus probability, and the mixing parameter; and
determining at least one candidate word in the incomplete response message based, at least in part, on the ranked plurality of dictionary words.

7. The method of claim 1 , wherein determining at least one candidate next word further comprises:
determining at least one LM probability for a plurality of words in a dictionary based, at least in part, on the at least one determined preceding word within the incomplete response message;
determining the topic probability for a plurality of candidate topics based, at least in part, on a topic model and the SM;
selecting at least one topic based, at least in part, on the determined plurality of candidate topic probabilities;
determining at least one stimulus probability for the plurality of dictionary word based, at least in part on the selected at least one topic;
ranking the plurality of dictionary words based, at least in part on the at least one LM probability, the at least one determined stimulus probability, and the mixing parameter; and
determining the at least one candidate word for the incomplete response message based, at least in part, on the ranked plurality of dictionary words.

8. The method of claim 1 , wherein determining the mixing parameter of the mixing model further comprises:
determining a topic parameter based, at least in part, on a subset of received stimulus-response parameters; and
determining the mixing parameter of the mixing model based, at least in part, on the determined topic parameter.

9. The method of claim 1 , further comprising:
determining a most likely topic probability based, at least in part, on use of the function t* =argmax t /(topic=t|s), wherein t* corresponds to a most likely topic, wherein t corresponds to a candidate topic, and wherein s corresponds to a received stimulus message.

10. A system to provide response message completion, comprising:
at least one network device, comprising:
a memory to store parameters and instructions:
a processor to execute instructions to:
determine a response completion model (RCM) via a linear combination of a generic response language model (LM)and a stimulus model with a mixing parameter for a mixing model, the mixing parameter to be based, at least in part, on a topic probability, the topic probability to be assigned a value to be based, at least in part, on a likelihood that a candidate topic is to be associated with a received stimulus message;
receive a stimulus message (SM);
if an incomplete response message is to include at least one preceding word, determine at least one candidate next word for the incomplete response message to be based, at least in part, on the RCM, the SM, and the at least one preceding word, and to be based, at least in part, on estimated word frequencies in stimulus messages from a plurality of stimulus-response message pairs;
select at least one word from the at least one determined candidate next word;
include the selected at least one word within the incomplete response message; and
generate a complete response message that is to be based, at least in part on the incomplete response message to include the selected at least one word.

11. The system of claim 10 , wherein the processor is additionally to:
if the incomplete response message is to be absent one or more words, determine the at least one candidate next word for the incomplete response message to be based, at least in part, on at least the RCM and the SM.

12. The system of claim 10 , wherein to determine the RCM is to further comprise:
to receive stimulus-response parameters to include at least a plurality of stimulus-response message pairs;
to determine a dictionary to com

Assignees

• Ravi Sujith
• Pang Bo
• Yahoo Inc

Inventors

• Ravi Sujith
• Pang Bo

Classifications

• H04L51/02 Primary

  using automatic reactions or user delegation, e.g. automatic replies or chatbot-generated messages · CPC title

• G06F40/40

  Processing or translation of natural language (natural language analysis G06F40/20; semantic analysis G06F40/30) · CPC title

• H04L51/04

  Real-time or near real-time messaging, e.g. instant messaging [IM] · CPC title

• G06F40/274 Primary

  Converting codes to words; Guess-ahead of partial word inputs · CPC title

• G06F17/276

  Physics · mapped topic