Learning word representations via commonsense reasoning

1. A computer-implemented method for training a recursive reasoning unit (RRU) to learn word representations, comprising:

receiving a knowledge graph for a set of words and a co-occurrence matrix for a different set of words in a training phase, the knowledge graph (i) mapping each of the words in the set of words to a node having a node label and (ii) indicating a relation between adjacent nodes by an edge having an edge label, the co-occurrence matrix indicating word co-occurrence frequency of the different set of words;

using a weighting function in the se matrix to discount an influence of frequent words occurring greater than a first threshold amount;

discovering, by the RRU, reasoning paths from the knowledge graph for word pairs by mapping word pairs from the set of words into a source word and a destination word and finding the reasoning paths therebetween;

predicting word co-occurrence frequency using the reasoning paths; and

updating, responsive to the word co-occurrence frequency, model parameters of the RRU until a difference between a predicted word occurrence and a true word occurrence are less than a second threshold amount to provide a trained RRU.

2. The computer-implemented method of claim 1 , wherein said predicting step applies a sigmoid neural network activation function to the reasoning paths and learnable parameters of the RRU.

3. The computer-implemented method of claim 1 , further comprising:

receiving the knowledge graph, a learned RRU, and input test words to be represented by the word representations in a test phase; and

selectively generating, by the RRU using the knowledge graph, vector word representations for at least one of individual words and word pairs, from among the input test words.

4. The computer-implemented method of claim 3 , further comprising performing a question-answering session using the vector word representations as an input query.

5. The computer-implemented method of claim 3 , further comprising performing a sentiment analysis using the vector word representations as an input query.

6. The computer-implemented method of claim 1 , further comprising obtaining vector word representations by concatenating all word embedding from each hop along a given reasoning path.

7. The computer-implemented method of claim 1 , further comprising, for intermediate words at each hop along a given one of the reasoning paths, calculating transition probabilities to a next hop word along the given one of the reasoning paths.

8. A computer program product for training a recursive reasoning unit (RRU) to learn word representations, the computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computer to cause the computer to perform a method comprising:

receiving, by the RRU implemented by a hardware processor of the computer, a knowledge graph for a set of words and a co-occurrence matrix for a different set of words, the knowledge graph (i) mapping each of the words in the set of words to a node having a node label and (ii) indicating a relation between adjacent nodes by an edge having an edge label, the co-occurrence matrix indicating word co-occurrence frequency of the different set of words;

using a weighting function in the co-occurrence matrix to discount an influence of frequent words occurring greater than a first threshold amount:

discovering, by the RRU, reasoning paths from the knowledge graph for word pairs by mapping word pairs from the set of words into a source word and a destination word and finding the reasoning paths therebetween;

predicting, by the RRU, word co-occurrence frequency using the reasoning paths; and

updating, by the RRU responsive to the word co-occurrence frequency, model parameters of the RRU until a difference between a predicted word occurrence and a true word occurrence are less than a second threshold amount to provide a trained RRU.

9. The computer program product of claim 8 , wherein said predicting step applies a sigmoid neural network activation function to the reasoning paths and learnable parameters of the RRU.

10. The computer program product of claim 8 , further comprising:

inputting the knowledge graph, a learned recursive reasoning unit (RRU), and input test words to be represented by the word representations; and

selectively generating, by the RRU using the knowledge graph, vector word representations for at least one of individual words and word pairs, from among the input test words.

11. The computer program product of claim 10 , further comprising performing a question-answering session using the vector word representations as an input query.

12. The computer program product of claim 10 , further comprising performing a sentiment analysis using the vector word representations as an input query.

13. The computer program product of claim 8 , further comprising obtaining vector word representations by concatenating all word embedding from each hop along a given reasoning path.

14. The computer program product of claim 8 , further comprising, for intermediate words at each hop along a given one of the reasoning paths, calculating transition probabilities to a next hop word along the given one of the reasoning paths.

15. A computer processing system for training a recursive reasoning unit (RRU) to learn word representations, comprising:

a memory device for storing program code; and

a hardware processor operatively coupled to the memory device to form the RRU for running the program code to:

receive a knowledge graph for a set of words and a co-occurrence matrix for a different set of words in a training phase, the knowledge graph (i) mapping each of the words in the set of words to a node having a node label and (ii) indicating a relation between adjacent nodes by an edge having an edge label, the co-occurrence matrix indicating word co-occurrence frequency of the different set of words;

use a weighting function in the co-occurrence matrix to discount an influence of frequent words occurring greater than a first threshold amount;

discover reasoning paths from the knowledge graph for word pairs by mapping word pairs from the set of words into a source word and a destination word and finding the reasoning paths therebetween;

predict word co-occurrence frequency using the reasoning paths; and

update, responsive to the word co-occurrence frequency, model parameters of the RRU until a difference between a predicted word occurrence and a true word occurrence are less than a second threshold amount to provide a trained RRU.

16. The computer processing system of claim 15 , wherein the word co-occurrence frequency is predicted by applying a sigmoid neural network activation function to the reasoning paths and learnable parameters of the RRU.

17. The computer processing system of claim 15 , wherein the hardware processor is further configured to run the program code as the RRU to:

receive the knowledge graph, a learned RRU, and input test words to be represented by the word representations in a test phase; and

selectively generate, using the knowledge graph, vector word representations for at least one of individual words and word pairs, from among the input test words.

18. The computer processing system of claim 17 , wherein the hardware processor is further configured to run the program code as the RRU to perform a question-answering session using the vector word representations as an input query.