Dialogue act classification in group chats with DAG-LSTMs

1. A computer implemented method for classifying a dialogue act in a chat log, comprising:

mapping each word of the dialogue act to a word vector representation;

computing an utterance vector representation of the dialogue act based on the word vector representations using a bidirectional long short-term memory (LSTM) architecture;

computing an additional utterance vector representation of the dialogue act based on the utterance vector representation, the computing the additional utterance vector representation comprising:

applying a skip connection between the dialogue act of a participant and an immediately prior dialogue act of the same participant; and

computing the additional utterance vector representation based on an utterance vector representation of the skip connection to the immediately prior dialogue act of the same participant;

mapping the additional utterance vector representation, with a directed-acyclic-graph long short-term memory network (DAG-LSTM), to a classification of the dialogue act; and

outputting the classification of the dialogue act.

2. The computer implemented method of claim 1 , wherein computing an additional utterance vector representation of the dialogue act based on the utterance vector representation comprises:

computing the additional utterance vector representation based on utterance vector representations of all prior dialogue acts in the chat log.

3. The computer implemented method of claim 1 , wherein computing an additional utterance vector representation of the dialogue act based on the utterance vector representation comprises:

computing the additional utterance vector representation of the dialogue act using a modified tree long short-term memory (LSTM) based architecture.

4. The computer implemented method of claim 1 , wherein the chat log is a transcript of a conversation between a plurality of participants.

5. An apparatus comprising:

a processor; and

a memory to store computer program instructions for classifying a dialogue act in a chat log, the computer program instructions when executed on a neural network of the processor cause the processor to perform operations comprising:

mapping each word of the dialogue act to a word vector representation;

computing an utterance vector representation of the dialogue act based on the word vector representations using a bidirectional long short-term memory (LSTM) architecture;

computing an additional utterance vector representation of the dialogue act based on the utterance vector representation, the computing the additional utterance vector representation comprising:

applying a skip connection between the dialogue act of a participant and an immediately prior dialogue act of the same participant; and

computing the additional utterance vector representation based on an utterance vector representation of the skip connection to the immediately prior dialogue act of the same participant;

mapping the additional utterance vector representation, with a directed-acyclic-graph long short-term memory network (DAG-LSTM), to a classification of the dialogue act; and

outputting the classification of the dialogue act.

6. The apparatus of claim 5 , wherein computing an additional utterance vector representation of the dialogue act based on the utterance vector representation comprises:

computing the additional utterance vector representation based on utterance vector representations of all prior dialogue acts in the chat log.

7. The apparatus of claim 5 , wherein computing an additional utterance vector representation of the dialogue act based on the utterance vector representation comprises:

computing the additional utterance vector representation of the dialogue act using a modified tree long short-term memory (LSTM) based architecture.

8. The apparatus of claim 5 , wherein the chat log is a transcript of a conversation between a plurality of participants.

9. A non-transitory computer readable medium storing computer program instructions for classifying a dialogue act in a chat log, the computer program instructions when executed by a processor cause the processor to perform operations comprising:

mapping each word of the dialogue act to a word vector representation;

computing an utterance vector representation of the dialogue act based on the word vector representations using a bidirectional long short-term memory (LSTM) architecture;

computing an additional utterance vector representation of the dialogue act based on the utterance vector representation, the computing the additional utterance vector representation comprising:

applying a skip connection between the dialogue act of a participant and an immediately prior dialogue act of the same participant; and

computing the additional utterance vector representation based on an utterance vector representation of the skip connection to the immediately prior dialogue act of the same participant;

mapping the additional utterance vector representation, with a directed-acyclic-graph long short-term memory network (DAG-LSTM), to a classification of the dialogue act; and

outputting the classification of the dialogue act.

10. The non-transitory computer readable medium of claim 9 , wherein computing an additional utterance vector representation of the dialogue act based on the utterance vector representation comprises:

computing the additional utterance vector representation based on utterance vector representations of all prior dialogue acts in the chat log.

11. The non-transitory computer readable medium of claim 9 , wherein computing an additional utterance vector representation of the dialogue act based on the utterance vector representation comprises:

computing the additional utterance vector representation of the dialogue act using a modified tree long short-term memory (LSTM) based architecture.

12. The non-transitory computer readable medium of claim 9 , wherein the chat log is a transcript of a conversation between a plurality of participants.