Semi-supervised learning of word embeddings

• US 9,659,560 B2
• Filed: 09/30/2015
• Issued: 05/23/2017
• Est. Priority Date: 05/08/2015
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Cao, Liangliang, Fan, James J., Wang, Chang, Xiang, Bing, Zhou, Bowen
• Primary Examiner(s): SINGH, SATWANT K
• Application Number:
  US14/870,204
• Publication Number:
  US 20160328388A1
• Time in Patent Office: 601 Days
• Field of Search: None
• CPC Class Codes: G06F 16/3347 using vector based model
• CPC Class Codes: G06F 40/169 Annotation, e.g. comment data or footnotes
• CPC Class Codes: G06F 40/40 Processing or translation of natural language natural language analysis G06F40/20; semantic analysis G06F40/30
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G06N 3/084 Backpropagation, e.g. using gradient descent
• CPC Class Codes: G06N 3/088 Non-supervised learning, e.g. competitive learning
• CPC Class Codes: G06N 5/022 Knowledge engineering; Knowledge acquisition
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/16 using artificial neural networks
• CPC Class Codes: G10L 15/18 using natural language modelling

Abstract

Software that trains an artificial neural network for generating vector representations for natural language text, by performing the following steps: (i) receiving, by one or more processors, a set of natural language text; (ii) generating, by one or more processors, a set of first metadata for the set of natural language text, where the first metadata is generated using supervised learning method(s); (iii) generating, by one or more processors, a set of second metadata for the set of natural language text, where the second metadata is generated using unsupervised learning method(s); and (iv) training, by one or more processors, an artificial neural network adapted to generate vector representations for natural language text, where the training is based, at least in part, on the received natural language text, the generated set of first metadata, and the generated set of second metadata.

Claims

1. A method comprising:

receiving, by one or more processors, a set of natural language text;

generating, by one or more processors, a set of first metadata for the set of natural language text, where the first metadata is generated using supervised learning method(s);

generating, by one or more processors, a set of second metadata for the set of natural language text, where the second metadata is generated using unsupervised learning method(s);

training, by one or more processors, an artificial neural network adapted to generate vector representations for natural language text, where the training is based, at least in part, on the received natural language text, the generated set of first metadata, and the generated set of second metadata;

generating, by one or more processors, a set of at least two vector representations for the set of natural language text using the trained artificial neural network, where each vector representation of the set of at least two vector representations pertains to a respective subset of natural language text from the set of natural language text;

generating, by one or more processors, a vector representation pertaining to the set of natural language text by adding each of the vector representations in the generated set of at least two vector representations; and

storing, by one or more processors, the generated vector representation pertaining to the set of natural language text for use by a natural language processing system.

2. The method of claim 1, further comprising:

determining, by one or more processors, an amount of similarity between at least two subsets of natural language text from the set of natural text by comparing their respectively generated vector representations.

3. The method of claim 2, wherein each of the at least two subsets of natural language text is a word.

4. The method of claim 1, further comprising:

generating, by one or more processors, a set of first metadata for the generated set of at least two vector representations, where the first metadata for the generated set of at least two vector representations is generated using supervised learning method(s);

generating, by one or more processors, a set of second metadata for the set of at least two vector representations, where the second metadata for the generated set of at least two vector representations is generated using unsupervised learning method(s); and

training, by one or more processors, the artificial neural network based, at least in part, on the generated set of at least two vector representations, the generated set of first metadata for the set of at least two vector representations, and the generated set of second metadata for the set of at least two vector representations.

5. The method of claim 1, further comprising:

generating, by one or more processors, a set of initial vector representations for the set of natural language text;

generating, by one or more processors, a set of first metadata vector representations for the generated set of first metadata; and

generating, by one or more processors, a set of second metadata vector representations for the generated set of second metadata;

wherein the training of the artificial neural network is further based, at least in part, on the generated set of initial vector representations, the generated set of first metadata vector representations, and the generated set of second metadata vector representations.

6. The method of claim 1, wherein the supervised learning methods utilize at least one of a natural language processing annotator or an ontology.

7. The method of claim 1, wherein the unsupervised learning methods are based on at least one of reconstruction error or language modeling.