System and method of highlighting influential samples in sequential analysis

1. A computerized method for highlighting relative importance of portions of a conversation displayed on a graphical user interface, comprising:

storing the conversation in computerized memory connected to a computer processor that is configured to display conversation text on a graphical user interface, wherein a display of the conversation illustrates conversation data according to respective conversation participants' turns in providing conversation input;

weighting respective turns of the conversation by providing the conversation input of the respective turns to a hierarchical attention network stored in the memory, wherein the hierarchical attention network uses the processor to calculate sequential long-short-term-memory cells (LSTM) in the memory;

using later LSTM cell data to update weighting values for prior LSTM cell data in a sequence of turns of the conversation input;

wherein weighting the respective turns comprises adding conversation input data from additional later turns of the conversation to new LSTM cells; and

wherein the hierarchical attention network uses the computer processor to calculate sequential long-short-term-memory cells (LSTM) in the memory when a prior weighting of turns in the conversation have had a degree of uniformity greater than a uniformity tolerance threshold.

2. A computerized method according to claim 1 , wherein weighting the respective turns comprises changing weights of the prior LSTM cell data in response to the additional later turns.

3. A computerized method according to claim 2 , wherein the processor identifies a plurality of turns in the conversation illustrating at least one change in weight distribution among the plurality of turns as an attention dependency switch.

4. A computerized method according to claim 2 , wherein the processor identifies sequential turns in the conversation illustrating at least one change in weight between two turns as a context dependency switch.

5. A computerized method according to claim 2 , wherein the processor identifies at least one turn in the conversation illustrating at least one change in weight, across the entire conversation and greater than a variation dependency variable, as a variation dependency switch.

6. A computerized method according to claim 1 , wherein weighting a group of turns in the conversation comprises forming a weight vector from occurrences of at least one attention dependency switch, at least one context dependency switch, and at least one variation dependency switch, averaging components of the vector, and representing each term in the group of terms on a graphical user interface with a pixel intensity that corresponds to the average of the components of the weight vector.

7. A computerized method according to claim 1 , wherein the processor and memory form a turn weight vector comprising weighting values for turns in the conversation and calculate a degree of uniformity (α) across members of the vector.

8. A computerized method according to claim 7 , wherein the processor and the memory use the turn weight vector to identify either uniformity or non-uniformity across the weights in the weight vector by comparing sequential weighting vectors from sequential turns to an attention dependency variable (τa).

9. A computerized method according to claim 7 , wherein the processor and the memory use the turn weight vector to identify instances across the conversation in which an addition of a turn changes the weights of previous turns by comparing the weighting vectors to a context dependency variable (τc).

10. A computerized method according to claim 7 , wherein the processor and the memory use the turn weight vector to identify individual weighting value changes across the conversation in which an addition of a turn changes the weight of a respective individual weighting value more than variation dependency variable (τv).

11. A computerized method according to claim 1 , wherein a point of escalation in the conversation is identified from the weighting.

12. A computerized system comprising the method of claim 1 implemented in system hardware comprising the processor, memory, and a graphical user interface.

13. A computerized method for highlighting relative importance of portions of a conversation displayed on a graphical user interface, comprising:

storing the conversation in computerized memory connected to a computer processor that is configured to display the conversation on a graphical user interface, wherein a display of the conversation illustrates conversation data according to respective conversation participants' turns in providing conversation input;

weighting respective turns of the conversation by providing the conversation input of the respective turns to a hierarchical attention network stored in the memory, wherein the hierarchical attention network uses the processor to calculate sequential long-short-term-memory cells (LSTM) in the memory;

using later LSTM cell data to update weighting values for prior LSTM cell data in a sequence of turns of conversation input data; and

displaying the conversation participants' turns on the graphical user interface, wherein displaying the conversation participants' turns on the graphical user interface comprises:

the processor and memory forming a turn weight vector comprising weighting values for turns in the conversation and calculating a degree of uniformity (α) across members of the vector;

the processor and the memory using the turn weight vector to identify attention dependency and either uniformity or non-uniformity across the weighting values in the weight vector by comparing sequential weighting vectors from sequential turns to an attention dependency variable (τa);

the processor and the memory using the turn weight vector to identify instances across the conversation in which an addition of a turn identifies context dependency and changes in the weighting values of previous turns by comparing the weighting vectors to a context dependency variable (τc);

the processor and the memory using the turn weight vector to identify variation dependency and individual weighting value changes across the conversation, in which an addition of a turn changes the weight of a respective individual weighting value more than variation dependency variable (τv); and

selecting a visible indication for displaying the respective turns according to combinations of attention dependency, context dependency, and variation dependency across the turn weight vector for the conversation.