NATURAL LANGUAGE GENERATION USING PINNED TEXT AND MULTIPLE DISCRIMINATORS
A personality model is created for a population and used as an input to a text generation system. Alternative texts are created based upon the emotional effect of the generated text. Certain words or phrases are “pinned” in the output, reducing the variability of the generated text so as to preserve required information content, and a number of tests provide input to a discriminator network so that proposed outputs both match an outside objective regarding the information content, emotional affect, and grammatical acceptability. A feedback loop provides new “ground truth” data points for refining the personality model and associated generated text.
1 . A system for generating natural language, the system comprising:
a prospect modeling component, operable to correlate known information about a target prospect personality with a quantitative personality model, the quantitative personality model being expressed as a vector indicating the relative expression of a plurality of relatively mutually orthogonal personality traits;
a neural sequence-to-sequence encoder-decoder, wherein the encoder is operable to deconstruct a source text and represent it as a sequence of weights on a pre-built conditional text model; and
wherein the decoder is operable to create a generated text with approximately equal semantic content but differing syntax and word choice; and
wherein the syntax and word choice of the generated text varies as a function of the expression of the quantitative personality model.
2 . The system of claim 1 , further comprising an evaluator coupled to the encoder and decoder, wherein the evaluator is operable to compare the generated text with a set of measurements made against the source text to create a text similarity evaluation; and
wherein the text similarity evaluation is provided to one or both of the encoder and decoder; and
wherein the internal weights associated with one or both of the encoder and decoder are updated to reinforce high text similarity and to discourage low text similarity.
3 . The system of claim 1 , further comprising a discriminator coupled to the encoder and decoder, wherein the discriminator provides a distinguishability score reflecting a weighted probability that the generated text is human-generated; and
wherein the distinguishability score is provided to one or both of the encoder and decoder; and
wherein the internal weights associated with one or both of the encoder and decoder are updated to reinforce low distinguishability and to discourage high distinguishability.
4 . The system of claim 1 , further comprising an evaluator coupled to the encoder and decoder, wherein the evaluator provides a personality score reflecting the association of the language use in the generated text with the personality model input to the decoder; and
wherein the personality score is provided to the decoder; and
wherein the internal weights associated with the decoder are updated to reinforce high personality association and to discourage lower personality association.
5 . The system of claim 1 , wherein the generated text is provided to a representative of the modeled prospect class; and wherein the response of the representative is used to update the prospect model.
6 . The system of claim 5 , wherein the updating of the prospect model is relative to the measured receptiveness of the representative to the text.
7 . The system of claim 5 , wherein the updating of the prospect model is relative to the imputed measurement of personality traits.
8 . A method for generating natural language, comprising:
providing a quantitative personality model, using a prospect modeling component, based on correlating known information about a target prospect personality, the quantitative personality model being expressed as a vector indicating a relative expression of a plurality of relatively mutually orthogonal personality traits;
deconstructing a source text and representing it as a sequence of weights on a prebuilt conditional text model, using an encoder of a neural sequence-to-sequence encoder-decoder; and
creating a generated text with approximately equal semantic content but differing syntax and word choice, using a decoder of the neural sequence-to-sequence encoder-decoder,
wherein the syntax word choice of the generated text varies as a function of the expression of the quantitative personality model.
9 . The method of claim 13 , further comprising:
comparing, using an evaluator coupled to the encoder and decoder, the generated text with a set of measurements made against the source text to create a text similarity evaluation,
wherein the text similarity evaluation is provided to one or both of the encoder and decoder, and
wherein internal weights associated with one or both of the encoder and decoder are updated to reinforce high text similarity and to discourage low text similarity.
10 . The method of claim 13 , further comprising:
creating, using a discriminator coupled to the encoder and decoder, a distinguishability score reflecting a weighted probability that the generated text is human-generated,
wherein the distinguishability score is provided to one or both of the encoder and decoder, and
wherein internal weights associated with one or both of the encoder and decoder are updated to reinforce low distinguishability and to discourage high distinguishability.
11 . The method of claim 13 , further comprising:
creating, using an evaluator coupled to the encoder and decoder, a personality score reflecting an association of the language use in the generated text with the qualitative personality model,
wherein the personality score is provided to the decoder, and
wherein internal weights associated with the decoder are updated to reinforce high personality association and to discourage lower personality association.
12 . The method of claim 13 , wherein the generated text is provided to a representative of the modeled prospect class, and wherein a response of the representative is used to update the quantitative personality model.
13 . The method of claim 17 , wherein updating of the quantitative personality model is relative to a measured receptiveness of the representative to the text.
14 . The method of claim 17 , wherein updating of the quantitative personality model is relative to an imputed measurement of personality traits.
15 . A non-transitory computer readable medium storing instructions which, when executed by one or more processors, cause the one or more processors to:
provide a quantitative personality model, using a prospect modeling component, based on correlating known information about a target prospect personality, the quantitative personality model being expressed as a vector indicating a relative expression of a plurality of relatively mutually orthogonal personality traits;
deconstruct a source text and represent it as a sequence of weights on a pre-built conditional text model, using an encoder of a neural sequence-to-sequence encoder-decoder; and
create a generated text with approximately equal semantic content but differing syntax and word choice, using a decoder of the neural sequence-to-sequence encoder-decoder,
wherein the syntax word choice of the generated text varies as a function of the expression of the quantitative personality model.
16 . The non-transitory computer readable medium of claim 20 further comprising instructions which, when executed by the one or more processors, cause the one or more processors to:
compare, using an evaluator coupled to the encoder and decoder, the generated text with a set of measurements made against the source text to create a text similarity evaluation,
wherein the text similarity evaluation is provided to one or both of the encoder and decoder, and
wherein internal weights associated with one or both of the encoder and decoder are updated to reinforce high text similarity and to discourage low text similarity.
17 . The non-transitory computer readable medium of claim 20 further comprising instructions which, when executed by the one or more processors, cause the one or more processors to:
create, using a discriminator coupled to the encoder and decoder, a distinguishability score reflecting a weighted probability that the generated text is human-generated,
wherein the distinguishability score is provided to one or both of the encoder and decoder, and
wherein internal weights associated with one or both of the encoder and decoder are updated to reinforce low distinguishability and to discourage high distinguishability.
18 . The non-transitory computer readable medium of claim 20 further comprising instructions which, when executed by the one or more processors, cause the one or more processors to:
create, using an evaluator coupled to the encoder and decoder, a personality score reflecting an association of the language use in the generated text with the qualitative personality model,
wherein the personality score is provided to the decoder, and
wherein internal weights associated with the decoder are updated to reinforce high personality association and to discourage lower personality association.
19 . The non-transitory computer readable medium of claim 20 , wherein
the instructions comprise instructions which, when executed by the one or more processors, cause the one or more processors to provide the generated text to a representative of the modeled prospect class and to update the quantitative personality model based on a response of the representative.
20 . The non-transitory computer readable medium of claim 24 , wherein the instructions comprise instructions which, when executed by the one or more processors, cause the one or more processors to update the quantitative personality model relative to a measured receptiveness of the representative to the text.
21 . (canceled)