System and method for automatic, unsupervised paraphrase generation using a novel framework that learns syntactic construct while retaining semantic meaning

• US 9,984,063 B2
• Filed: 09/15/2016
• Issued: 05/29/2018
• Est. Priority Date: 09/15/2016
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Bennett, Laura J., Krishnamurthy, Lakshminarayanan, Parameswaran, Niyati, Sudarsan, Sridhar
• Primary Examiner(s): Lerner, Martin
• Application Number:
  US15/266,754
• Publication Number:
  US 20180075015A1
• Time in Patent Office: 621 Days
• Field of Search: 704 1, 704 9, 704257, 704 4
• CPC Class Codes: G06F 40/211 Syntactic parsing, e.g. based on context-free grammar [CFG] or unification grammars
• CPC Class Codes: G06F 40/289 Phrasal analysis, e.g. finite state techniques or chunking
• CPC Class Codes: G06F 40/30 Semantic analysis
• CPC Class Codes: G06F 40/56 Natural language generation

Abstract

A system includes a question answering system executed by a computer, a processor, and a memory coupled to the processor. The memory is encoded with instructions that when executed cause the processor to provide training for training the question answering system. The training system is configured to receive a first phrase and a second phrase, the first and second phrases being paraphrases of each other, convert the first phrase into a first logical form and the second phrase into a second logical form, generate a phrasal edit that includes a difference between the first logical form and the second logical form, convert the phrasal edit into a disjunctive logical form in two directions, and generate a first plurality of paraphrases of the first and second phrases based on the disjunctive logical form.

Claims

1. A method, comprising:

receiving, by a paraphrase generation system, a first phrase and a second phrase, the first and second phrases being paraphrases of each other;

converting, by the paraphrase generation system, the first phrase into a first logical form and the second phrase into a second logical form;

generating, by the paraphrase generation system, a plurality of phrasal edits that includes differences between the first logical form and the second logical form;

converting, by the paraphrase generation system, the plurality of phrasal edits into a plurality of disjunctive logical forms in two directions;

generating, by the paraphrase generation system based on the plurality of disjunctive logical forms in two directions, a first plurality of paraphrases of the first and second phrases;

determining a first score for a first paraphrase of the first plurality of paraphrases;

determining a second score for a second paraphrase of the first plurality of paraphrases, wherein the first score is higher than the second score based in part on a first syntactic variation between the first paraphrase and the first phrase and the second phrase being greater than a second syntactic variation between the second paraphrase and the first phrase and the second phrase; and

pruning, by the paraphrase generation system, the first plurality of paraphrases to yield a second plurality of paraphrases including grammatical alternatives to the first and second phrases, wherein the first and second paraphrases are pruned based on the first score and the second score such that the first paraphrase is included in the second plurality of paraphrases and the second paraphrase is not included in the second plurality of paraphrases based on the first score being higher than the second score.

2. The method of claim 1, wherein the converting the first phrase into the first logical form includes dependency parsing the first phrase to identify parts of speech of the first phrase thereby generating a first parsed representation of the first phrase and the converting the second phrase into the second logical form includes dependency parsing the second phrase to identify parts of speech of the second phrase thereby generating a second parsed representation of the second phrase.

3. The method of claim 2, wherein the generating the plurality of phrasal edits includes comparing the first parsed representation of the first phrase with the second parsed representation of the second phrase to determine a word from the first phrase that corresponds with a word from the second phrase.

4. The method of claim 1, wherein the plurality of disjunctive logical forms includes a plurality of disjunctions that represent alternative choices of words at a level of semantic dependencies.

5. The method of claim 4, wherein the converting the plurality of phrasal edits into the plurality of disjunctive logical forms includes generating a bi-directional frequency distribution model for words in the first and second phrases, wherein the bi-directional frequency distribution model includes a probability that, given a particular first word of the words in the first and second phrases, a particular second word follows the particular first word and a particular third word precedes the particular first word.

6. A system, comprising:

a question answering system executed by a computer;

a processor; and

a memory coupled to the processor, the memory encoded with instructions that when executed cause the processor to provide training for the question answering system at least in part by causing the processor to;

receive a first phrase and a second phrase, the first and second phrases being paraphrases of each other;

convert the first phrase into a first logical form and the second phrase into a second logical form;

generate a phrasal edit that includes a difference between the first logical form and the second logical form;

convert the phrasal edit into a disjunctive logical form in two directions;

generate a first plurality of paraphrases of the first and second phrases based on the disjunctive logical form;

determine a first score for a first paraphrase of the first plurality of paraphrases;

determine a second score for a second paraphrase of the first plurality of paraphrases, wherein the first score is higher than the second score based in part on a first syntactic variation between the first paraphrase and the first phrase and the second phrase being greater than a second syntactic variation between the second paraphrase and the first phrase and the second phrase; and

prune the first plurality of paraphrases to yield a second plurality of paraphrases including grammatical alternatives to the first and second phrases, wherein the first and second paraphrases are pruned based on the first score and the second score such that the first paraphrase is included in the second plurality of paraphrases and the second paraphrase is not included in the second plurality of paraphrases based on the first score being higher than the second score.

7. The system of claim 6, wherein the processor is further configured to provide the second plurality of paraphrases and intent of the first and second phrases to the question answering system as enriched ground truth.

8. The system of claim 6, wherein the disjunctive logical form includes a disjunction that represents an alternative word choice at a level of semantic dependencies.

9. The system of claim 8, wherein the processor is configured to convert the phrasal edit into the disjunctive logical form by generating a first trigram model.

10. The system of claim 9, wherein the first trigram model is a bi-directional frequency distribution model for words in the first and second phrases, and wherein the bi-directional frequency distribution model includes a probability that, given a particular first word of the words in the first and second phrases, a particular second word follows the particular first word and a particular third word precedes the particular first word.

11. The system of claim 6, wherein the processor is configured to:

convert the first phrase into the first logical form by dependency parsing the first phrase to identify parts of speech of the first phrase thereby generating a first parsed representation of the first phrase; and

convert the second phrase into the second logical form includes dependency parsing the second phrase to identify parts of speech of the second phrase thereby generating a second parsed representation of the second phrase.

12. A computer program product for generating paraphrases, the computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computer to cause the computer to:

receive a first phrase and a second phrase, the first and second phrases being paraphrases of each other;

convert the first phrase into a first logical form and the second phrase into a second logical form;

generate a plurality of phrasal edits that include differences between the first logical form and the second logical form;

convert the plurality of phrasal edits into a plurality of disjunctive logical forms in two directions;

generate a first plurality of paraphrases of the first and second phrases based on the plurality of disjunctive logical forms;

determine a first score for a first paraphrase of the first plurality of paraphrases;

determine a second score for a second paraphrase of the first plurality of paraphrases, wherein the first score is higher than the second score based in part on a first syntactic variation between the first paraphrase and the first phrase and the second phrase being greater than a second syntactic variation between the second paraphrase and the first phrase and the second phrase; and

prune the first plurality of paraphrases to yield a second plurality of paraphrases containing grammatical alternatives to the first and second phrases, wherein the first and second paraphrases are pruned based on the first score and the second score such that the first paraphrase is included in the second plurality of paraphrases and the second paraphrase is not included in the second plurality of paraphrases based on the first score being higher than the second score.

13. The computer program product of claim 12, wherein the program instructions are further executable by the computer to cause the computer to convert the first phrase into the first logical form by dependency parsing the first phrase to identify parts of speech of the first phrase thereby generating a first parsed representation of the first phrase and the second phrase into the second logical form by dependency parsing the second phrase to identify parts of speech of the second phrase thereby generating a second parsed representation of the second phrase.

14. The computer program product of claim 13, wherein the program instructions are further executable by the computer to cause the computer to generate the plurality of phrasal edits by comparing the first parsed representation of the first phrase with the second parsed representation of the second phrase to determine a word from the first phrase that corresponds with a word from the second phrase.

15. The computer program product of claim 12, wherein the plurality of disjunctive logical forms includes a plurality of disjunctions that represent alternative choices of words at a level of semantic dependencies.

16. The computer program product of claim 15, wherein the program instructions are further executable by the computer to cause the computer to convert the plurality of phrasal edits into the plurality of disjunctive logical forms by generating a first trigram model.

17. The computer program product of claim 16, wherein the first trigram model is a bi-directional frequency distribution model for words in the first and second phrases and wherein the bi-directional frequency distribution model includes a probability that, given a particular first word of the words in the first and second phrases, a particular second word follows the particular first word and a particular third word precedes the particular first word.