Learning neuro-symbolic multi-hop reasoning rules over text

Patent metadata

Field	Value
Publication number	US-11645526-B2
Application number	US-202016911645-A
Country	US
Kind code	B2
Filing date	Jun 25, 2020
Priority date	Jun 25, 2020
Publication date	May 9, 2023
Grant date	May 9, 2023

Abstract

A method and a system for learning and applying neuro-symbolic multi-hop rules are provided. The method includes inputting training texts into a neural network as well as pre-defined entities. The training texts and the entities relate to a specific domain. The method also includes generating an entity graph made up of nodes and edges. The nodes represent the pre-defined entities, and the edges represent passages in the training texts with co-occurrence of the entities connected together by the edges. The method further includes determining a relation based on the passages for each of the pre-defined entities connected together by the edges, calculating a probability relating to the relation, generating a potential reasoning path between a head entity and a target entity. The method also includes learning a neuro-symbolic rule by converting the edges along the potential reasoning path into symbolic rules and combining those rules into the neuro-symbolic rule.

First claim

What is claimed is:

1. A computer-implemented method for generating neuro-symbolic rules, the computer-implemented method comprising:
inputting training texts and pre-defined entities into a neural model, wherein the training texts and the pre-defined entities relate to a domain;
generating an entity graph including nodes and edges, wherein the nodes represent the pre-defined entities and the edges represent passages in the training texts with a co-occurrence of the pre-defined entities connected together by the edges;
determining a relation based on the passages for each of the pre-defined entities connected together by the edges;
calculating a probability relating to the relation for each of the pre-defined entities;
generating a potential reasoning path between a head entity and a target entity;
scoring the potential reasoning path based on a beam search of the potential reasoning path using the-probabilities of the edges;
learning a neuro-symbolic rule by converting the edges along the potential reasoning path into symbolic rules and combining the symbolic rules into the neuro-symbolic rule; and
applying the neuro-symbolic rule to perform multi-hop reasoning, thereby increasing a possibility that an entity answer selected using the multi-hop reasoning is correct.

2. The computer-implemented method of claim 1 , further comprising:
inputting answers into the neural model; and
scoring the potential reasoning path based on the answers.

3. The computer-implemented method of claim 1 , further comprising:
providing answer candidates to the neural model; and
learning neuro-symbolic rules for the answer candidates.

4. The computer-implemented method of claim 1 , further comprising:
inputting pre-defined symbolic rules into the neural model; and
combining the pre-defined symbolic rules with the neuro-symbolic rule.

5. The computer-implemented method of claim 1 , wherein the neural model is a long short-term memory (LSTM) recurrent neural network (RNN).

6. The computer-implemented method of claim 1 , wherein the neural model includes entity-aware encoding.

7. A computer-implemented method for applying neuro-symbolic rules, the computer-implemented method comprising:
inputting texts comprising passages pertaining to a domain into a neural model;
inputting a query relating to the domain, wherein the query includes a head entity and a target relationship into the neural model;
extracting entities from the texts, wherein the entities relate to the domain;
generating an entity graph with nodes connected by edges, wherein the nodes represent the entities extracted from the text and the edges represent passages within the texts with a co-occurrence of the entities connected together by the edges;
determining a textual relation for each of the edges in the entity graph including a probability relating to the textual relation;
extracting potential reasoning paths for candidate answers in the entity graph by applying the neuro-symbolic rules learned by the neural model, wherein applying the neuro-symbolic rules increases a possibility that the entity answer is correct;
scoring the potential reasoning paths based on a structured prediction; and
providing an entity answer based on scoring the potential reasoning paths.

8. The computer-implemented method of claim 7 , further comprising:
weighing the symbolic relation for each of the edges within the potential reasoning paths.

9. The computer-implemented method of claim 7 , further comprising:
converting a reasoning path with a highest score into a neuro-symbolic rule; and
combining the neuro-symbolic rule with the neuro-symbolic rules already learned by the neural model.

10. The computer-implemented method of claim 7 , wherein the entities are pre-defined entities relating to the texts.

11. The computer-implemented method of claim 7 , wherein scoring the potential reasoning paths is performed by a beam search through the entity graph to the candidate answer.

12. The computer-implemented method of claim 7 , further comprising:
optimizing the potential reasoning paths based on the structured prediction, wherein a potential reasoning path that has a neuro-symbolic rule indicating a likely correct answer is assigned a higher score as compared to scores of other potential reasoning paths, and
a potential reasoning path that has a neuro-symbolic rule indicating a likely incorrect answer is assigned a lower score as compared to scores of other potential reasoning paths.

13. The computer-implemented method of claim 7 , wherein the query includes multiple queries inputted into the neural model.

14. A neuro-symbolic rules system for learning neuro-symbolic rules using multi-hop reasoning, the neuro-symbolic rules system comprising:
one or more computer-readable storage media storing program instructions and one or more processors which, in response to executing the program instructions, are configured to:
extract entities from texts for a domain, wherein the entities are nouns relating to the domain;
generate an entity graph including nodes and edges, wherein the nodes correspond to the entities and the edges correspond to passages within the texts with a co-occurrence of the entities connected together by the edges;
determine, using a neural model, reasoning paths between a head entity and an answer entity that result in a target relationship, wherein the neural model scores the reasoning paths based on probabilities calculated by the neural model for each edge along the reasoning path;
learn, by the neural model, a neuro-symbolic rule; and
apply the neuro-symbolic rule to the reasoning paths, for increasing a possibility increases a likelihood of selecting a correct entity.

15. The neuro-symbolic rules system of claim 14 , wherein the neural model scores the reasoning paths based on inputted ground truth answers during a training process.

16. The neuro-symbolic rules system of claim 14 , wherein the neural model is a long short-term memory recurrent neural network.

17. The neuro-symbolic rules system of claim 14 , wherein the neural model scores the reasoning paths based on a beam search and using structured prediction.

18. The neuro-symbolic rules system of claim 14 , wherein the neuro-symbolic rule is learned by converting the edges along the reasoning paths into symbolic rules and combining those rules into the neuro-symbolic rule.

19. The neuro-symbolic rules system of claim 18 , wherein the neuro-symbolic rule is added to other neuro-symbolic rules learned by the neural model.

20. The neuro-symbolic rules system of claim 18 , wherein the neural model is configured to determine weights for the symbolic rules.

Assignees

• IBM

Inventors

• Yu Mo
• Zhang Li
• Klinger Tamir
• Guo Xiaoxiao

Classifications

• G06N3/09

  Supervised learning · CPC title

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N3/042

  Knowledge-based neural networks; Logical representations of neural networks · CPC title

• G06N5/025

  Extracting rules from data · CPC title

• G06N3/08 Primary

  Learning methods · CPC title