Mass-based clustering of utterances in financial contexts
A set of embeddings is generated from the data of a set of utterances originating in a financial data processing environment. An embedding corresponds to an utterance and includes a multidimensional vector whose dimensionality is reduced forming a compressed vector. A mass value is assigned to the compressed vector (mass-bearing data point (MD)). A set of MDs corresponds to the set of utterances. For a neighborhood of the MD, a torque value is iteratively adjusted using mass values and pairwise distances between pairs of members of the neighborhood of MD. After reaching an exit condition, a cluster is output with a better coherency correspondence with a singular actionable label as compared to another coherency correspondence of another cluster formed without the mass assignment or torque based adjustment. An operation is triggered from the actionable label in the financial data processing environment.
1 . A computer-implemented method, comprising:
generating, by reducing a dimensionality of a multidimensional vector of an embedding corresponding to an utterance, a compressed vector;
forming, by assigning a mass value to the compressed vector to form a mass-bearing data point (MD) in a set of MDs;
iteratively adjusting a torque value using mass values and pairwise distances between pairs of members of a first subset of MDs from MDs from the set of MDs, the first subset of MDs forming a neighborhood of the MD;
increasing, responsive to the iterative adjusting, a coherency in the first subset of MDs with a singular actionable label; and
triggering from the singular actionable label, an operation in a data processing environment.
2 . The computer-implemented method of claim 1 , further comprising:
receiving the utterance from a device configured to perform a function relative to the data processing environment.
3 . The computer-implemented method of claim 1 , wherein the utterance comprises natural language (NL) speech, and wherein the NL speech comprises a phrase spoken with an associated intent of causing the operation of a function relative to the data processing environment.
4 . The computer-implemented method of claim 3 , wherein the phrase comprises a contraction of an expression, wherein the expression maps to a plurality of functions.
5 . The computer-implemented method of claim 4 , wherein different placements of a word in the NL speech are indicative of different functions in the plurality of functions.
6 . The computer-implemented method of claim 1 , wherein the utterance comprises video data of a nonverbal gesture, and wherein the nonverbal gesture has an associated intent of causing the operation of a function relative to the data processing environment.
7 . The computer-implemented method of claim 6 , wherein the nonverbal gesture corresponds to an expression, wherein the expression maps to a plurality of functions.
8 . The computer-implemented method of claim 7 , wherein different patterns in the nonverbal gesture are indicative of different functions in the plurality of functions.
9 . The computer-implemented method of claim 1 , further comprising:
preprocessing the utterance, wherein the preprocessing comprises speech-to-text conversion of an NL speech in the utterance.
10 . The computer-implemented method of claim 1 , further comprising:
preprocessing the utterance, wherein the preprocessing comprises video-to-text conversion of a gesture in the utterance.
11 . The computer-implemented method of claim 1 , wherein the utterance is a member of a set of utterances, wherein the set of utterances comprises a subset of past utterances and a subset of present utterances, and wherein both the subsets are from a first user.
12 . The computer-implemented method of claim 11 , wherein the set of utterances comprises:
a subset of present utterances of the first user;
a subset of present utterances of a second user;
a subset of past utterances of the first user; and
a subset of present past utterances of the second user.
13 . The computer-implemented method of claim 1 , wherein the utterance is a member of a set of utterances comprising a subset of past transactions and a subset of present utterances, wherein both the subsets are from a common user.
14 . The computer-implemented method of claim 1 , wherein the reducing comprises executing a feature thresholding operation on the embedding.
15 . The computer-implemented method of claim 1 , further comprising:
computing, as a part of assigning the mass value to the compressed vector, the mass value as a proportion of a sum of absolute values of a set of features in the compressed vector.
16 . The computer-implemented method of claim 1 , further comprising:
computing, as a part of assigning the mass value to the compressed vector, the mass value as a function of: (1) a variance in a set of features in the compressed vector, (2) a local density of the neighborhood of the MD, or a combination of (1) and (2).
17 . The computer-implemented method of claim 1 , further comprising:
changing, as a part of the iteratively adjusting, a value in the compressed vector of the MD such that a vector distance between the MD and the MD1 is reduced.
18 . A computer program product comprising one or more computer readable storage media, and program instructions collectively stored on the one or more computer readable storage media, the program instructions executable by a set of one or more processors to cause the set of one or more processors to perform operations comprising:
generating, by reducing a dimensionality of a multidimensional vector of an embedding corresponding to an utterance, a compressed vector;
forming, by assigning a mass value to the compressed vector to form a mass-bearing data point (MD) in a set of MDs;
iteratively adjusting a torque value using mass values and pairwise distances between pairs of members of a first subset of MDs from the set of MDs, the first subset of MDs forming a neighborhood of the MD;
increasing, responsive to the iterative adjusting, a coherency in the first subset of MDs with a singular actionable label; and
triggering from the singular actionable label, an operation in a data processing environment.
19 . The computer program product of claim 18 , wherein the stored program instructions are stored in a computer readable storage device in a data processing system, and wherein the stored program instructions are transferred over a network from a remote data processing system.
20 . A computer system comprising a set of one or more processors and one or more computer readable storage media, and program instructions collectively stored on the one or more computer readable storage media, the program instructions executable by the set of one or more processors to cause the processor to perform operations comprising:
generating, by reducing a dimensionality of a multidimensional vector of an embedding corresponding to an utterance, a compressed vector;
forming, by assigning a mass value to the compressed vector to form a mass-bearing data point (MD) in a set of MDs;
iteratively adjusting a torque value using mass values and pairwise distances between pairs of members of a first subset of MDs from the set of MDs, the first subset of MDs forming a neighborhood of the MD;
increasing, responsive to the iterative adjusting, a coherency in the first subset of MDs with a singular actionable label; and
triggering from the singular actionable label, an operation in a data processing environment.