Domain-specific stopword removal from unstructured computer text using a neural network

Patent metadata

Field	Value
Publication number	US-2018225568-A1
Application number	US-201715426958-A
Country	US
Kind code	A1
Filing date	Feb 7, 2017
Priority date	Feb 7, 2017
Publication date	Aug 9, 2018
Grant date	—

Abstract

Methods and apparatuses are described for analyzing unstructured computer text for domain-specific stopword identification and removal. A computer data store stores unstructured text. A server computing device splits the unstructured text into phrases and generates tokens from the phrases. The server computing device generates a set of bootstrap keywords using the tokens. An artificial intelligence neural network executing on the server computing device generates a stopword training model. The server computing device generates a first set of candidate stopwords using the bootstrap keywords and the stopword training model. The server computing device generates regular expressions using the bootstrap keywords, and generates a second set of candidate stopwords using the regular expressions. The server computing device stores the candidate stopwords in the data store, and removes stopwords from the unstructured text using the data store.

First claim

What is claimed is:

1 . A system used in a computing environment in which unstructured computer text is analyzed for domain-specific stopword identification and removal, the system comprising:
a computer data store including domain-specific unstructured text,
the unstructured text being input via a web page, input directly into the computer data store via a first computer file, or any combination thereof, and
a server computing device in communication with the computer data store and programmed to:
split the unstructured text into one or more phrases;
generate tokens from the phrases, wherein each token comprises a word;
generate a set of bootstrap keywords by:
a) selecting a subset of the tokens that (i) appear frequently in the unstructured text or (ii) are determined to be uninformative, and
b) inserting each token in the subset of the tokens into the computer data store as a bootstrap keyword;

generate, using an artificial intelligence neural network executing on the server computing device, a stopword training model by creating a word vector for each token and inserting the word vector in a high-dimensional space that comprises the training model, wherein a position of each word vector in the high-dimensional space is based upon a semantic relationship between the corresponding word and surrounding words in the unstructured text,
generate a first set of candidate stopwords by executing the stopword training model against the set of bootstrap keywords to determine one or more words in the unstructured text that are similar to the bootstrap keywords based upon the word vector for the bootstrap keywords;
generate regular expressions based upon the set of bootstrap keywords by:
a) identifying a sequence of characters in one or more of the bootstrap keywords,
b) identifying a syntactical pattern in the sequence of characters, and
c) generating a regular expression using the syntactical pattern;

generate a second set of candidate stopwords using the regular expressions and
insert the first set of candidate stopwords and the second set of candidate stopwords into the computer data store as a set of final stopwords; and

remove stopwords from the unstructured text that match the set of final stopwords.

2 . The system of claim 1 , wherein the artificial intelligence neural network positions each word vector in the high-dimensional space to be in proximity to similar word vectors.

3 . The system of claim 1 , wherein the server computing device splits the unstructured text into one or more phrases by locating a terminator in the unstructured text and separating the unstructured text on either side of the terminator into a phrase.

4 . The system of claim 1 , wherein the server computing device is further programmed to determine one or more themes in the unstructured text after the stopwords are removed.

5 . The system of claim 1 , wherein the high-dimensional space comprises hundreds of dimensions.

6 . The system of claim 1 , wherein the second set of candidate stopwords comprises a plurality of stopwords with a syntax that matches at least one regular expression.

7 . The system of claim 6 , wherein the second set of candidate stopwords includes one or more stopwords that are not present in the unstructured text.

8 . A computerized method in which unstructured computer text is analyzed for domain-specific stopword identification and removal, the method comprising:
storing, in a computer data store, domain-specific unstructured text, the unstructured text being input via a web page, input directly into the computer data store via a first computer file, or any combination thereof; splitting, by a server computing device in communication with the computer data store, the unstructured text into one or more phrases; generating, by the server computing device, tokens from the phrases, wherein each token comprises a word; generating, by the server computing device, a set of bootstrap keywords by:
a) selecting a subset of the tokens that (i) appear frequently in the unstructured text or (ii) are determined to be uninformative, and
b) inserting each token in the subset of the tokens into the computer data store as a bootstrap keyword;
generating, by an artificial intelligence neural network executing on the server computing device, a stopword training model by creating a word vector for each token and inserting the word vector in a high-dimensional space that comprises the training model, wherein a position of each word vector in the high-dimensional space is based upon a semantic relationship between the corresponding word and surrounding words in the unstructured text; generating, by the server computing device, a first set of candidate stopwords by executing the stopword training model against the set of bootstrap keywords to determine one or more words in the unstructured text that are similar to the bootstrap keywords based upon the word vector for the bootstrap keywords; generating, by the server computing device regular expressions based upon the set of bootstrap keywords by:
a) identifying a sequence of characters in one or more of the bootstrap keywords,
b) identifying a syntactical pattern in the sequence of characters, and
c) generating a regular expression using the syntactical pattern;
generating, by the server computing device, a second set of candidate stopwords using the regular expressions and insert the first set of candidate stopwords and the second set of candidate stopwords into the computer data store as a set of final stopwords; and removing, by the server computing device, stopwords from the unstructured text that match the set of final stopwords.

9 . The method of claim 8 , wherein the artificial intelligence neural network positions each word vector in the high-dimensional space to be in proximity to similar word vectors.

10 . The method of claim 8 , wherein the server computing device splits the unstructured text into one or more phrases by locating a terminator in the unstructured text and separating the unstructured text on either side of the terminator into a phrase.

11 . The method of claim 8 , wherein the server computing device is further programmed to determine one or more themes in the unstructured text after the stopwords are removed.

12 . The method of claim 8 , wherein the high-dimensional space comprises hundreds of dimensions.

13 . The method of claim 8 , wherein the second set of candidate stopwords comprises a plurality of stopwords with a syntax that matches at least one regular expression.

14 . The method of claim 13 , wherein the second set of candidate stopwords includes one or more stopwords that are not present in the unstructured text.

15 . A computer readable storage medium comprising programmatic instructions for operation of a computing environment in which unstructured computer text is analyzed for domain-specific stopword identification and removal, the instructions operable to cause
a computer data store to store domain-specific unstructured text, the unstructured text being input via a web page, input directly into the computer data store via a first computer file, or any combination thereof; and a server computing device in communication with the computer data store, and including an executable artificial intelligence neural network executing on the server computing device, to:
split the unstructured text into one or more phrases;
generate tokens from the phrases, wherein each token comprises a word;
generate a set of bootstrap keywords by
a) selecting a subset of the tokens that (i) appear frequently

Assignees

• Fmr Llc

Inventors

• Chandramouli Aravind

Classifications

• G06F40/30

  Semantic analysis · CPC title

• G06F16/313

  Selection or weighting of terms for indexing · CPC title

• G06F16/3344 Primary

  using natural language analysis · CPC title

• G06F40/216

  using statistical methods · CPC title

• G06F40/284

  Lexical analysis, e.g. tokenisation or collocates · CPC title