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 method for maintaining a specific semantic content in a generated natural language text, the method comprising:
identifying words or phrases essential to the specific semantic content of a source text;
performing an evaluation of the source text,
wherein the evaluation includes a first summarization procedure, followed by a first measurement of a relative order and precedence of essential words and phrases in a source summarization result;
encoding the source text in a neural sequence-to-sequence encoder-decoder to generate a code;
sampling from a distribution implied by the code to create a candidate generated text;
performing an evaluation of the generated text,
wherein the evaluation of the generated text includes a second summarization procedure, followed by a second measurement of the relative order and precedence of the essential words and phrases in a generated summarization result;
computing a first difference between the first measurement and the second measurement;
in response to the first difference being greater than an essential similarity threshold, providing a negative feedback response to the encoder and decoder; and
in response to the first difference being less than the essential similarity threshold, providing a positive feedback response to the encoder and decoder; and
updating internal weights associated with the encoder and decoder in response to the feedback.
2. The method of claim 1 , wherein in response to the first difference being greater that the essential similarity threshold, the encode, decode, and evaluation steps are repeated until the first difference is less than the essential similarity threshold.
3. The method of claim 1 , wherein the performing of the evaluation further comprises computing a source statistical measurement associated with a length and textual form of the source text, and wherein the evaluation of the generated text includes computing a generated statistical measurement comparable to the source statistical measurement and computing at least one statistical similarity measurement based on a comparison of the source statistical measurement and the generated statistical measurement.
4. The method of claim 1 , wherein the performing of the evaluation further comprises computing a source readability measurement associated with a length and textual form of the source text, and wherein the evaluation of the generated text includes computing a generated readability measurement comparable to the source readability measurement and computing at least one readability similarity measurement based on a comparison of the source readability measurement and the generated readability measurement.
5. The method of claim 4 , wherein the readability similarity measurement is compared relative to an expected reading capacity of a prospect.