Performing fine-grained question type classification

Patent metadata

Field	Value
Publication number	US-11520762-B2
Application number	US-201916713776-A
Country	US
Kind code	B2
Filing date	Dec 13, 2019
Priority date	Dec 13, 2019
Publication date	Dec 6, 2022
Grant date	Dec 6, 2022

Abstract

A computer-implemented method according to one embodiment includes converting an input question into a vector form using trained word embeddings; constructing a type similarity matrix using a predetermined ontology; and determining a score for all possible types for the input question, based on the input question in vector form and the type similarity matrix.

First claim

What is claimed is:

1. A computer-implemented method, comprising:
converting, by a neural network, an input question into a vector form using trained word embeddings;
constructing, by the neural network, a type similarity matrix using a predetermined ontology;
performing, by the neural network, matrix multiplication between the input question in vector form and the type similarity matrix to determine a plurality of possible types for the input question; and
scoring, by the neural network, each of the plurality of possible types for the input question to create a ranked list of possible types for the input question,
wherein the predetermined ontology includes a plurality of labeled nodes representing entities and their associated types,
wherein the labeled nodes within the predetermined ontology are restructured and reformatted into a series of vectors in Euclidean vector space to construct the type similarity matrix,
wherein each vector within the series of vectors conveys a predetermined type within the predetermined ontology,
wherein a location of each vector within the Euclidean vector space is used to determine a similarity score between the associated types of the labeled nodes within the predetermined ontology,
wherein the neural network performs matrix multiplication between the input question in vector form and the type similarity matrix to determine a plurality of possible types for the input question, and
wherein the plurality of possible types for the input question is refined utilizing a list of potential types for training data derived from the predetermined ontology.

2. The computer-implemented method of claim 1 , wherein the input question is converted into a vector by a bidirectional long short-term memory (LSTM) and a combination layer.

3. The computer-implemented method of claim 1 , wherein:
the input question includes a text string, and
the input question is converted into the vector form by a recurrent neural network (RNN).

4. The computer-implemented method of claim 1 , wherein each of a plurality of vectors are created within a Euclidean vector space to represent an associated type within the predetermined ontology, where a proximity of the vectors within the Euclidean vector space is associated with a similarity of the vectors.

5. The computer-implemented method of claim 1 , wherein the type similarity matrix is created utilizing a low rank decomposition method such as SVD on top of a co-occurrence matrix.

6. The computer-implemented method of claim 1 , wherein:
the neural network is trained utilizing a plurality of question/answer pairs, wherein each of the question/answer pairs includes a question string and an associated answer string, and
for each of the plurality of question/answer pairs, a type of the associated answer string is determined from the predetermined ontology and is assigned to the question string as its label to create a list of question/type pairs.

7. The computer-implemented method of claim 1 , wherein:
the input question includes a text string that is converted into the vector form by a recurrent neural network (RNN).

8. The computer-implemented method of claim 1 , further comprising:
encoding a difference between a vector of possible types for the input question as predicted by a system and a list of potential types as provided by training data by utilizing a loss function such as a weighted negative log-likelihood (NLL) operation; and
minimizing the difference during a training phase of a neural network which modifies parameters of the neural network.

9. The computer-implemented method of claim 8 , further comprising back-propagating a loss value for the input question through the neural network.

10. A computer program product comprising one or more computer readable storage media, and program instructions collectively stored on the one or more computer readable storage media, the program instructions comprising instructions configured to cause one or more processors to perform a method comprising:
converting, by a neural network utilizing the one or more processors, an input question into a vector form using trained word embeddings;
constructing, by the neural network utilizing the one or more processors, a type similarity matrix using a predetermined ontology;
performing, by the neural network utilizing the one or more processors, matrix multiplication between the input question in vector form and the type similarity matrix to determine a plurality of possible types for the input question; and
scoring, by the neural network utilizing the one or more processors, each of the plurality of possible types for the input question to create a ranked list of possible types for the input question,
wherein the predetermined ontology includes a plurality of labeled nodes representing entities and their associated types,
wherein the labeled nodes within the predetermined ontology are restructured and reformatted into a series of vectors in Euclidean vector space to construct the type similarity matrix,
wherein each vector within the series of vectors conveys a predetermined type within the predetermined ontology,
wherein a location of each vector within the Euclidean vector space is used to determine a similarity score between the associated types of the labeled nodes within the predetermined ontology,
wherein the neural network performs matrix multiplication between the input question in vector form and the type similarity matrix to determine a plurality of possible types for the input question, and
wherein the plurality of possible types for the input question is refined utilizing a list of potential types for training data derived from the predetermined ontology.

11. The computer program product of claim 10 , wherein the input question is converted into a vector by a bidirectional long short-term memory (LSTM) and a combination layer.

12. The computer program product of claim 10 , wherein:
the input question includes a text string, and
the input question is converted into the vector form by a recurrent neural network (RNN).

13. The computer program product of claim 10 , wherein each of a plurality of vectors are created within a Euclidean vector space to represent an associated type within the predetermined ontology, where a proximity of the vectors within the Euclidean vector space is associated with a similarity of the vectors.

14. The computer program product of claim 10 , wherein the type similarity matrix is created utilizing a low rank decomposition method such as SVD on top of a co-occurrence matrix.

15. The computer program product of claim 12 , wherein:
the neural network is trained utilizing a plurality of question/answer pairs, wherein each of the question/answer pairs includes a question string and an associated answer string, and
for each of the plurality of question/answer pairs, a type of the associated answer string is determined from the predetermined ontology and is assigned to the question string as its label to create a list of question/type pairs.

16. A system, comprising:
a processor; and
logic integrated with the processor, executable by the processor, or integrated with and executable by the processor, the logic being configured to:
convert, by a neural network, an input question into a vector form using trained word embeddings;
construct, by the neural network, a type similarity matrix using a predetermined ontology;
perform, by the neural network, matrix multiplication between the input question in vector form and the type similarity matrix to determine a plurality of possible types for

Assignees

• IBM

Inventors

• Dash Sarthak
• Rossiello Gaetano
• Gliozzo Alfio Massimiliano
• Farrell Robert G
• Makni Bassem
• Sil Avirup
• Castelli Vittorio
• Florian Radu

Classifications

• G06F16/35 Primary

  Clustering; Classification · CPC title

• G06F16/2237 Primary

  Vectors, bitmaps or matrices · CPC title

• G06F18/24133

  Distances to prototypes · CPC title

• G06F18/24

  Classification techniques · CPC title

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title