Methods and systems for graphically organizing the meanings of words
Described are methods and systems for graphically organizing words and their meanings. The words, in an input language such as English, are added as nodes to a graph. Synonyms of each input word taken from a second language, such as Sanskrit, are likewise added as nodes to the graph and connected to the corresponding input-word nodes via synonym edges. Root elements of the synonyms, semantic meanings, are also added to the graph as meaning nodes. The resulting graph depicts linguistic and semantic interrelations between input words, their second-language synonyms, and the meanings derived therefrom.
1 . A method of graphically organizing meanings of a word training a neural network for natural-language processing using words expressed in an input language, the method comprising:
extracting Dhatu tensors from the words by:
for each of the words:
adding the word to a graph as a word node;
adding Sanskrit synonyms of the word as Sanskrit nodes of the graph;
representing synonym edges of the graph between the word node and the Sanskrit nodes;
adding Dhatus of the Sanskrit synonyms as Dhatu nodes of the graph;
representing Dhatu edges of the graph between the Sanskrit nodes and the Dhatu nodes;
adding meanings of the Dhatus as meaning nodes of the graph; and
representing meaning edges of the graph between the Dhatu nodes and the meaning nodes;
mapping the meaning nodes to the Dhatu tensors, the Dhatu tensors serving as embeddings for the neural network; and
training the neural network on the embeddings.
2 . The method of claim 1 , wherein adding one of the Dhatus as one of the Dhatu nodes of the graph comprises populating a table with all possible combinations of character spans for the one of the Dhatus and comparing the character spans against a Dhatu list.
3 . The method of claim 2 , wherein some of the meaning nodes are more closely connected than others of the meaning nodes, and wherein the Dhatus tensors of the more closely connected meaning nodes are more similar than the Dhatus tensors of less closely connected meaning nodes.
4 . The method of claim 1 , wherein the word is in English.
5 . The method of claim 2 , wherein the comparing provides matches between the character spans and elements of the Dhatu list, the method further comprising scoring the matches and selecting among the elements of the Dhatu list based on the scoring.
6 . The method of claim 5 , further comprising using scores from the scoring to select Dhatus for substrings of the Sanskrit synonym such that an aggregate of the scores is maximized.
7 . A system for extracting embeddings for natural-language processing from a word expressed in an input language, the system comprising:
one or more processors configured to:
add the word to a graph as a word node;
add Sanskrit synonyms of the word as Sanskrit nodes of the graph;
represent synonym edges of the graph between the word node and the Sanskrit nodes;
add Dhatus of the Sanskrit synonyms as Dhatu nodes of the graph;
represent Dhatu edges of the graph between the Sanskrit nodes and the Dhatu nodes;
add meanings of the Dhatus as meaning nodes of the graph;
represent meaning edges of the graph between the Dhatu nodes and the meaning nodes;
map the meaning nodes to Dhatu tensors, the Dhatu tensors serving as embeddings for a neural network; and
train the neural network on the embeddings.
8 . The system of claim 7 , wherein the one or more processors are further configured to add one of the Dhatus as one of the Dhatu nodes of the graph by populating a table with all possible combinations of character spans for the one of the Dhatus and comparing the character spans against a Dhatu list.
9 . The system of claim 7 , wherein some of the meaning nodes are more closely connected than others of the meaning nodes, and wherein the Dhatu tensors of the more closely connected meaning nodes are more similar than the Dhatu tensors of less closely connected meaning nodes.
10 . The system of claim 7 , wherein the word is in English.
11 . The system of claim 7 , the one or more processors further configured to combine the Dhatu tensors with second embeddings to enrich semantic representation for natural-language processing.