Induction of dialog flow within a conversational natural language system
Structure of conversations between users and agents and/or systems is discovered and interactively displayed to analysts, thereby better supporting development of automated conversation handling systems for different domains. A corpus of prior dialogs of users with agents (without preexisting semantic labels indicating purposes for different parts of the dialogs) is taken as input, and embeddings are generated for textual units (e.g., rounds) of the dialogs. The embeddings are used to cluster the textual units, and the clusters and their relationships are visualized within a user interface that analysts may use to explore and fine-tune the structure of the conversations.
1. A computer-implemented method of visualizing dialog flow in an agent-user dialog system, the method comprising:
segmenting text for a plurality of dialogs between an agent and a user into a plurality of conversation units;
generating a plurality of embeddings corresponding to the plurality of conversation units by applying a pre-trained model to the conversation units of the dialogs, the pre-trained model comprising:
a first Bi-LSTM trained to encode a conversation unit into a dense vector, and
a second Bi-LSTM trained to generate an embedding for a set of dense vectors generated by the first Bi-LSTM,
wherein the first Bi-LSTM and the second Bi-LSTM are trained using examples encoded using a third Bi-LSTM, the examples comprising positive examples comprising rounds of stored dialogs and negative examples comprising randomly sampled stored dialogs;
clustering the embeddings into a plurality of embedding clusters;
forming a dialog flow graph, wherein the embedding clusters are nodes of the dialog flow graph and edges of the dialog flow graph are formed based on transitions in the dialogs corresponding to the embeddings of the conversation units; and
providing an interactive graphical user interface for visualizing the dialog flow graph.
2. The computer-implemented method of claim 1 , wherein the plurality of conversations comprised audio conversations, and wherein the text for the plurality of conversations is obtained by applying speech-to-text algorithms to the audio conversations.
3. The computer-implemented method of claim 1 , wherein the conversation units are conversation rounds, each conversation round consisting of agent text corresponding to a message of the agent to the user and user text corresponding to a response of the user to the agent.
4. The computer-implemented method of claim 1 , further comprising labeling the embedding clusters with keywords obtained from textual statistics derived from text of the conversation units corresponding to the embeddings in the clusters.
5. The computer-implemented method of claim 4 , wherein the keywords comprise both agent-representative keywords derived from agent utterances and user-representative keywords derived from user utterances.
6. The computer-implemented method of claim 1 , wherein forming the dialog flow graph comprises, for each of a plurality of the dialogs, counting transitions between clusters corresponding to the conversation units of the dialog.
7. The computer-implemented method of claim 1 , wherein the interactive graphical user interface comprises a dialog flow view portion depicting the dialog flow graph, and a cluster view portion depicting text of the conversation units corresponding to a cluster node that is currently selected within the dialog flow view.
8. The computer-implemented method of claim 7 , wherein the interactive graphical user interface additionally comprises an individual dialog view portion depicting text of a dialog that contains a conversation unit that is currently selected within the cluster view.
9. A non-transitory computer-readable storage medium storing instructions that when executed by a processor perform actions comprising:
segmenting text for a plurality of dialogs between an agent and a user into a plurality of conversation units;
generating a plurality of embeddings corresponding to the plurality of conversation units by applying a pre-trained model to the conversation units of the dialogs, the pre-trained model comprising:
a first Bi-LSTM trained to encode a conversation unit into a dense vector,
a second Bi-LSTM trained to generate an embedding for a set of dense vectors generated by the first Bi-LSTM, and
wherein the first Bi-LSTM and the second Bi-LSTM are trained using examples encoded using a third Bi-LSTM, the examples comprising positive examples comprising rounds of stored dialogs and negative examples comprising randomly sampled stored dialogs;
clustering the embeddings into a plurality of embedding clusters;
forming a dialog flow graph, wherein the embedding clusters are nodes of the dialog flow graph and edges of the dialog flow graph are formed based on transitions in the dialogs corresponding to the embeddings of the conversation units; and
providing an interactive graphical user interface for visualizing the dialog flow graph.
10. The non-transitory computer-readable storage medium of claim 9 , wherein the plurality of conversations comprised audio conversations, and wherein the text for the plurality of conversations is obtained by applying speech-to-text algorithms to the audio conversations.
11. The non-transitory computer-readable storage medium of claim 9 , wherein the conversation units are conversation rounds, each conversation round consisting of agent text corresponding to a message of the agent to the user and user text corresponding to a response of the user to the agent.
12. The non-transitory computer-readable storage medium of claim 9 , the actions further comprising labeling the embedding clusters with keywords obtained from textual statistics derived from text of the conversation units corresponding to the embeddings in the clusters.
13. The non-transitory computer-readable storage medium of claim 12 , wherein the keywords comprise both agent-representative keywords derived from agent utterances and user-representative keywords derived from user utterances.
14. The non-transitory computer-readable storage medium of claim 9 , wherein forming the dialog flow graph comprises, for each of a plurality of the dialogs, counting transitions between clusters corresponding to the conversation units of the dialog.
15. The non-transitory computer-readable storage medium of claim 9 , wherein the interactive graphical user interface comprises a dialog flow view portion depicting the dialog flow graph, and a cluster view portion depicting text of the conversation units corresponding to a cluster node that is currently selected within the dialog flow view.
16. The non-transitory computer-readable storage medium of claim 15 , wherein the interactive graphical user interface additionally comprises an individual dialog view portion depicting text of a dialog that contains a conversation unit that is currently selected within the cluster view.
17. A computer system comprising:
a computer processor; and
a non-transitory computer-readable storage medium storing instructions that when executed by the processor perform actions comprising:
segmenting text for a plurality of dialogs between an agent and a user into a plurality of conversation units;
generating a plurality of embeddings corresponding to the plurality of conversation units by applying a pre-trained model to the conversation units of the dialogs, the pre-trained model comprising:
a first Bi-LSTM trained to encode a conversation unit into a dense vector, and
a second Bi-LSTM trained to generate an embedding for a set of dense vectors generated by the first Bi-LSTM,
wherein the first Bi-LSTM and the second Bi-LSTM are trained using examples encoded using a third Bi-LSTM, the examples comprising positive examples comprising rounds of stored dialogs and negative examples comprising randomly sampled stored dialogs;
clustering the embeddings into a plurality of embedding clusters;
forming a dialog flow graph, wherein the embedding clusters are nodes of the dialog flow graph and edges of the dialog flow graph are formed based on transitions in the dialogs corresponding to the embeddings of the conversation units; and
providing an interactive graphical user interface for visualizing the dialog flow graph.
18. The computer system of claim 17 , wherein the plurality of conversations comprised audio conversations, and wherein the text for the plurality of conversations is obtained by applying speech-to-text algorithms to the audio conversations.
19. The computer system of claim 17 , wherein the conversation units are conversation rounds, each conversation round consisting of agent text corresponding to a message of the agent to the user and user text corresponding to a response of the user to the agent.
20. The computer system of claim 17 , the actions further comprising labeling the embedding clusters with keywords obtained from textual statistics derived from text of the conversation units corresponding to the embeddings in the clusters.