Employing abstract meaning representation to lay the last mile towards reading comprehension

1. A method of verifying an alignment of text, the method comprising:

creating a first semantic tree from a question and second semantic tree from a candidate answer, wherein each semantic tree comprises nodes and edges, wherein the nodes represent entities, and wherein each edge represents a relationship between two of the entities;

identifying, between the first semantic tree and the second semantic tree, a plurality of common subtrees, wherein a common subtree comprises (i) nodes, each node representing a common entity that is common between the first semantic tree and the second semantic tree and (ii) edges between the nodes that represent a semantic relationship between the entities;

calculating a semantic alignment score from a sum of sizes of each of the plurality of common subtrees, wherein a size of a common subtree equals a number of nodes in the common subtree;

forming a first syntactic tree for the question and a second syntactic tree for the candidate answer, wherein each syntactic tree comprises syntactic nodes that represent a word and an associated part of speech;

identifying, between the first syntactic tree and the second syntactic tree, a number of common syntactic nodes;

calculating a syntactic alignment score based on the number of common syntactic nodes; and

responsive to determining that a sum of the semantic alignment score and the syntactic alignment score is greater than a threshold, outputting the candidate answer to a device.

2. The method of claim 1 , wherein identifying the number of common syntactic nodes comprises:

applying a trained classification model to the first syntactic tree and the second syntactic tree; and

obtaining the number of common syntactic nodes from the classification model.

3. The method of claim 1 , further comprising:

creating a third semantic tree from an additional candidate answer;

identifying an additional plurality of common subtrees between the question and the additional candidate answer;

calculating an additional semantic alignment score that is based on a sum of sizes of each of the additional plurality of common subtrees;

forming a third syntactic tree for the candidate answer;

identifying, between the first syntactic tree and the third syntactic tree, an additional number of common syntactic nodes;

calculating an additional syntactic alignment score based on the additional number of common syntactic nodes; and

responsive to determining that an additional sum of the additional semantic alignment score and the additional syntactic alignment score is greater than a threshold, outputting the additional candidate answer to the device.

4. The method of claim 1 , further comprising:

providing the question into a classification model, wherein the classification model is trained to select a body of text corresponding to the question; and

obtaining, from the classification model, the candidate answer.

5. The method of claim 4 , further comprising training the classification model by iteratively:

providing, to the classification model, a training question and text comprising context corresponding to the training question;

receiving, from the classification model, an answer;

comparing entities in answer to entities in a reference answer that corresponds to the training question; and

adjusting, based on the comparing, a parameter of the classification model to minimize a loss function.

6. The method of claim 1 , further comprising:

translating text into a set of vectors, wherein each vector represents words and sentences from the text;

providing the set of vectors to a classification model, wherein the classification model is trained to generate answers from text; and

obtaining, from the classification model, the candidate answer.

7. The method of claim 6 , wherein the translating comprises providing the question and the candidate answer to a word2vec model.

8. A system comprising:

a non-transitory computer-readable medium storing computer-executable program instructions; and

a processing device communicatively coupled to the non-transitory computer-readable medium for executing the computer-executable program instructions, wherein executing the computer-executable program instructions configures the processing device to perform operations comprising:

creating a first semantic tree from a question and second semantic tree from a candidate answer, wherein each semantic tree comprises nodes and edges, wherein the nodes represent entities, and wherein each edge represents a relationship between two of the entities;

identifying, between the first semantic tree and the second semantic tree, a plurality of common subtrees, wherein a common subtree comprises (i) nodes, each node representing a common entity that is common between the first semantic tree and the second semantic tree and (ii) edges between the nodes that represent a semantic relationship between the entities;

calculating a semantic alignment score from a sum of sizes of each of the plurality of common subtrees, wherein a size of a common subtree equals a number of nodes in the common subtree;

forming a first syntactic tree for the question and a second syntactic tree for the candidate answer, wherein each syntactic tree comprises syntactic nodes that represent a word and an associated part of speech;

identifying, between the first syntactic tree and the second syntactic tree, a number of common syntactic nodes;

calculating a syntactic alignment score based on the number of common syntactic nodes; and

responsive to determining that a sum of the semantic alignment score and the syntactic alignment score is greater than a threshold, outputting the candidate answer to a device.

9. The system of claim 8 , wherein identifying the number of common syntactic nodes comprises:

applying a trained classification model to the first syntactic tree and the second syntactic tree; and

obtaining the number of common syntactic nodes from the classification model.

10. The system of claim 8 , the operations further comprising:

creating a third semantic tree from an additional candidate answer;

identifying an additional plurality of common subtrees between the question and the additional candidate answer;

calculating an additional semantic alignment score that is based on a sum of sizes of each of the additional plurality of common subtrees;

forming a third syntactic tree for the candidate answer;

identifying, between the first syntactic tree and the third syntactic tree, an additional number of common syntactic nodes;

calculating an additional syntactic alignment score based on the additional number of common syntactic nodes; and

responsive to determining that an additional sum of the additional semantic alignment score and the additional syntactic alignment score is greater than a threshold, outputting the additional candidate answer to the device.

11. The system of claim 8 , the operations further comprising:

providing the question into a classification model, wherein the classification model is trained to select a body of text corresponding to the question; and

obtaining, from the classification model, the candidate answer.

12. The system of claim 11 , the operations further comprising training the classification model by iteratively:

providing, to the classification model, a training question and text comprising context corresponding to the training question;

receiving, from the classification model, an answer;

comparing entities in answer to entities in a reference answer that corresponds to the training question; and

adjusting, based on the comparing, a parameter of the classification model to minimize a loss function.

13. The system of claim 8 , the operations further comprising:

translating text into a set of vectors, wherein each vector represents words and sentences from the text;

providing the set of vectors to a classification model, wherein the classification model is trained to generate answers from text; and

obtaining, from the classification model, the candidate answer.

14. A non-transitory computer-readable storage medium storing computer-executable program instructions, wherein when executed by a processing device, the program instructions cause the processing device to perform operations comprising:

creating a first semantic tree from a question and second semantic tree from a candidate answer, wherein each semantic tree comprises nodes and edges, wherein the nodes represent entities, and wherein each edge represents a relationship between two of the entities;

identifying, between the first semantic tree and the second semantic tree, a plurality of common subtrees, wherein a common subtree comprises (i) nodes, each node representing a common entity that is common between the first semantic tree and the second semantic tree and (ii) edges between the nodes that represent a semantic relationship between the entities;

calculating a semantic alignment score from a sum of sizes of each of the plurality of common subtrees, wherein a size of a common subtree equals a number of nodes in the common subtree;

forming a first syntactic tree for the question and a second syntactic tree for the candidate answer, wherein each syntactic tree comprises syntactic nodes that represent a word and an associated part of speech;

identifying, between the first syntactic tree and the second syntactic tree, a number of common syntactic nodes;

calculating a syntactic alignment score based on the number of common syntactic nodes; and

responsive to determining that a sum of the semantic alignment score and the syntactic alignment score is greater than a threshold, outputting the candidate answer to a device.

15. The non-transitory computer-readable storage medium of claim 14 , wherein identifying the number of common syntactic nodes comprises:

applying a trained classification model to the first syntactic tree and the second syntactic tree; and

obtaining the number of common syntactic nodes from the classification model.

16. The non-transitory computer-readable storage medium of claim 14 , wherein the operations further comprise:

creating a third semantic tree from an additional candidate answer;

identifying an additional plurality of common subtrees between the question and the additional candidate answer;

calculating an additional semantic alignment score that is based on a sum of sizes of each of the additional plurality of common subtrees;

forming a third syntactic tree for the candidate answer;

identifying, between the first syntactic tree and the third syntactic tree, an additional number of common syntactic nodes;

calculating an additional syntactic alignment score based on the additional number of common syntactic nodes; and

responsive to determining that an additional sum of the additional semantic alignment score and the additional syntactic alignment score is greater than a threshold, outputting the additional candidate answer to the device.

17. The non-transitory computer-readable storage medium of claim 14 , wherein the operations further comprise:

providing the question into a classification model, wherein the classification model is trained to select a body of text corresponding to the question; and

obtaining, from the classification model, the candidate answer.

18. The non-transitory computer-readable storage medium of claim 17 , wherein the operations further comprise training the classification model by iteratively:

providing, to the classification model, a training question and text comprising context corresponding to the training question;

receiving, from the classification model, an answer;

comparing entities in answer to entities in a reference answer that corresponds to the training question; and

adjusting, based on the comparing, a parameter of the classification model to minimize a loss function.

19. The non-transitory computer-readable storage medium of claim 14 , wherein the operations further comprise:

translating text into a set of vectors, wherein each vector represents words and sentences from the text;

providing the set of vectors to a classification model, wherein the classification model is trained to generate answers from text; and

obtaining, from the classification model, the candidate answer.

20. The non-transitory computer-readable storage medium of claim 19 , wherein the translating comprises providing the question and the candidate answer to a word2vec model.