U.S. patent application number 14/845508 was filed with the patent office on 2016-03-03 for natural language processing system and method.
This patent application is currently assigned to Fido Labs Inc.. The applicant listed for this patent is Fido Labs Inc.. Invention is credited to Tomasz Krupa, Gniewosz Leliwa, Michal Stanczyk, Piotr Wiacek, Michal Wroczynski.
Application Number | 20160062982 14/845508 |
Document ID | / |
Family ID | 50628267 |
Filed Date | 2016-03-03 |
United States Patent
Application |
20160062982 |
Kind Code |
A1 |
Wroczynski; Michal ; et
al. |
March 3, 2016 |
NATURAL LANGUAGE PROCESSING SYSTEM AND METHOD
Abstract
A natural language processing system is disclosed herein.
Embodiments of the NLP system perform hand-written rule-based
operations that do not rely on a trained corpus. Rules can be added
or modified at any time to improve accuracy of the system, and to
allow the same system to operate on unstructured plain text from
many disparate contexts (e.g. articles as well as twitter contexts
as well as medical articles) without harming accuracy for any one
context. Embodiments also include a language decoder (LD) that
generates information which is stored in a three-level framework
(word, clause, phrase). The LD output is easily leveraged b various
software applications to analyze large quantities of text from any
source in a more sophisticated and flexible manner than previously
possible. A query language (LDQL) for information extraction from
NLP parsers' output is disclosed, with emphasis on on its
embodiment implemented for LD. It is also presented, how to use
LDQL for knowledge extraction on the example of application named
Knowledge Browser.
Inventors: |
Wroczynski; Michal; (Gdynia,
PL) ; Krupa; Tomasz; (Sopot, PL) ; Leliwa;
Gniewosz; (Gdansk, PL) ; Wiacek; Piotr;
(Gdansk, PL) ; Stanczyk; Michal; (Gdansk,
PL) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Fido Labs Inc. |
Palo Alto |
CA |
US |
|
|
Assignee: |
Fido Labs Inc.
Palo Alto
CA
|
Family ID: |
50628267 |
Appl. No.: |
14/845508 |
Filed: |
September 4, 2015 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
14071631 |
Nov 4, 2013 |
9152623 |
|
|
14845508 |
|
|
|
|
61721792 |
Nov 2, 2012 |
|
|
|
Current U.S.
Class: |
704/9 |
Current CPC
Class: |
G06F 40/211 20200101;
G06F 40/30 20200101; G06F 40/284 20200101; G06F 40/205 20200101;
G06F 40/55 20200101; G06F 40/56 20200101 |
International
Class: |
G06F 17/27 20060101
G06F017/27; G06F 17/28 20060101 G06F017/28 |
Claims
1. (canceled)
2. A system for natural language processing (NLP) comprising: at
least one processor configured to execute a process comprising
extracting and structuring information from output of an NLP
parser, the process further comprising, receiving input comprising
output from the NLP parser; receiving a description of conditions
characterizing fragments of the NLP parser output, wherein the
fragments correspond to information to be extracted; and receiving
a description of a form in which the extracted information is to be
structured, wherein the description of conditions and the
description of a form are provided as one or more queries, wherein
each query comprises, a target section describing the form in which
the extracted information is to be structured; and a conditions
section describing conditions characterizing the fragments of the
NLP parser output to be extracted.
3. The system of claim 2, wherein: the conditions section comprises
one or more atomic formulas connected with zero or more logical
connectives, and wherein a set of possible atomic formulas
comprises: relation of identity; relation of membership; relation
of inclusion; relations capturing properties of character strings;
relations capturing the structure of NLP parser output; and the set
of possible logical connectives comprises, negation; conjunction;
and alternative.
4. The system of claim 3, wherein: the description of conditions
and the description of a form are provided in a form of one or more
queries, and zero or more user-defined predicate definitions; a
conditions section of each query, and each user-defined predicate
definition comprises a first order logic formula with variables
ranging over fragments of the NLP parser output: and the set of
possible atomic formulas comprises, first order predicate of
identity; first order predicate of membership; first order
predicate of inclusion; first order predicates capturing properties
of character strings; first order predicates capturing the
structure of NLP parser output; and each of the user-defined
predicates; and each group of two or more queries may comprise a
same user-defined predicate.
5. The system of claim 2, further comprising a compiler, wherein
the process further comprises converting description of conditions
and the description of a form into a program, wherein input of the
program comprises the NLP parser output.
6. The system of claim 5, wherein: the program comprises a sequence
of one or more operations to be performed on the NLP parser output
in order to extract and structure the information; and the compiler
comprises, a parser for converting the description of conditions
and the description of a form, into an abstract syntax tree; a
proper compiler, for converting the abstract syntax tree into the
sequence of the operations to be performed on the NLP parser
output.
7. The system of claim 6, wherein the compiler further comprises an
optimizer configured to modify the abstract syntax tree and the
sequence of the operations.
8. The system of claim 6, wherein the proper compiler produces a
standalone executable.
9. The system of claim 6, wherein the system further comprises a
virtual machine, and wherein input of the virtual machine comprises
the NLP parser output.
10. The system of claim 9, wherein input of the virtual machine
further comprises the sequence of the operations to be performed on
the NLP parser output.
11. The system of claim 10, wherein the proper compiler produces
the bytecode for the virtual machine.
12. A computer-implemented method for natural language queries of
textual data, the method comprising: at least one processor
executing a process comprising extracting and structuring
information from output of an NLP parser, the process further
comprising, receiving input comprising output from the NLP parser;
receiving a description of conditions characterizing fragments of
the NLP parser output, wherein the fragments correspond to
information to be extracted; and receiving a description of a form
in which the extracted information is to be structured, wherein the
description of conditions and the description of a form are
provided as one or more queries, wherein each query comprises, a
target section describing the form in which the extracted
information is to be structured; and a conditions section
describing conditions characterizing the fragments of the NLP
parser output to be extracted.
13. The method of claim 12, wherein: the conditions section
comprises one or more atomic formulas connected with zero or more
logical connectives, and wherein a set of possible atomic formulas
comprises: relation of identity; relation of membership; relation
of inclusion; relations capturing properties of character strings;
relations capturing the structure of NLP parser output; and wherein
the set of possible logical connectives comprises, negation;
conjunction; and alternative.
14. The method of claim 13, wherein: the description of conditions
and the description of a form are provided in a form of one or more
queries, and zero or more user-defined predicate definitions; a
conditions section of each query, and each user-defined predicate
definition comprises a first order logic formula with variables
ranging over fragments of the NLP parser output: and the set of
possible atomic formulas comprises, first order predicate of
identity; first order predicate of membership; first order
predicate of inclusion; first order predicates capturing properties
of character strings; first order predicates capturing the
structure of NLP parser output; and each of the user-defined
predicates; and each group of two or more queries may comprise a
same user-defined predicate.
15. The method of claim 12, further comprising a compiler, wherein
the process further comprises converting description of conditions
and the description of a form into a program, wherein input of the
program comprises the NLP parser output.
16. The method of claim 15, wherein: the program comprises a
sequence of one or more operations to be performed on the NLP
parser output in order to extract and structure the information;
and the compiler comprises, a parser for converting the description
of conditions and the description of a form into an abstract syntax
tree; a proper compiler, for converting the abstract syntax tree
into the sequence of the operations to be performed on the NLP
parser output.
17. The method of claim 16, wherein the compiler further comprises
an optimizer configured to modify the abstract syntax tree and the
sequence of the operations.
18. The method of claim 16, wherein the proper compiler produces a
standalone executable.
19. The method of claim 16, wherein the system further comprises a
virtual machine, and wherein input of the virtual machine comprises
the NLP parser output.
20. The method of claim 19, wherein input of the virtual machine
further comprises the sequence of the operations to be performed on
the NLP parser output.
21. The method of claim 20, wherein the proper compiler produces
the bytecode for the virtual machine.
Description
RELATED APPLICATIONS
[0001] This application is a continuation of U.S. patent
application Ser. No. 14/071,631, filed Nov. 4, 2013, which claims
the benefit of U.S. Provisional Patent Application No. 61/721,792,
filed Nov. 2, 2012. Both of these application are incorporated by
reference herein in their entirety.
FIELD OF THE INVENTION
[0002] Inventions disclosed and claimed herein are in the field of
natural language processing (NLP).
BACKGROUND
[0003] Natural language processing (NLP) systems are
computer-implemented methods for taking natural language input (for
example, computer-readable text), and operating on the input so as
to generate output that is useful for computers to derive meaning.
Examples of NLP systems applications include spell checkers/grammar
checkers, machine translation systems, and speech-to-text systems.
Increasingly, there is interest in developing methods for machines
to more intelligently interpret human language input data (such as
text) for the purpose of directing, the computer as if it were
another person who could understand speech. One application for
such methods is search engines that receive a typed query from a
person and perform web searches to attempt to generate a set of
meaningful answers to the query. An important subclass of NLP
systems is NLP parsers, especially grammatical parsers such as
Part-Of-Speech tagger, constituency parsers, dependency parsers,
and shallow semantic parsers such as SRL (Semantic Role Labeling).
Their role is to preprocess text and add additional information to
words to prepare it for further usage. Current NLP systems are
mostly built on top of NLP parsers and they features and accuracy
strongly relies on the information produced by these parsers.
Quality of the information delivered by these parsers is strongly
correlated with the efficiency of NLP systems.
[0004] FIG. 1 is a block diagram of prior art NLP parsers 100. The
input is a text, which is processed by NLP parser (102) consisting
of machine learning techniques (104) trained on manually annotated
corpus (105). The parser (102) produces the output (103) which is
then used by other systems/applications (106).
[0005] All current parsers are dependent on corpora and therefore
on the context in which they were written. Typically corpora are in
a context of correctly written, grammatically correct sentences and
common syntactic structures which are manually annotated by humans.
The system is then trained using this corpus.
[0006] This is one reason that traditional NLP parsers are most
accurate on the same type of content they were trained on (the same
corpus). That is why always-changing language, such as user
generated content (e.g. reviews, tips, comments, tweets, social
media content) presents a challenge for NLP parsers built with
machine learning techniques. Such content often includes
grammatically incorrect sentences and non-standard usage of
language, as well as emoticons, acronyms, strings of non-letter
characters and so on. This content is constantly changing and
expanding with different words or syntactic structures. All of this
content has meaningful information and it is easy to understand by
humans but it is still difficult for NLP applications to extract
meaning out of it.
[0007] One way in which current NLP parsers can be updated or
improved (for better accuracy or extracting additional information)
is to modify the existing corpus, or create a new corpus or
re-annotate existing one and retrain the system with it to
understand new content. However, this is a tedious, costly and
time-consuming process. For example all current NLP parsers are
using corpus as a training data annotated by linguists with
predefined tags (e.g. Penn Treebank)--especially use
machine-learning algorithms.
[0008] If there was a need to distinct pronominal or adjectival
aspect of "that" (giving them different POS tags in different
context), one would need to manually re-annotate by linguists all
the sentences in the whole corpus that contain word "that"
regarding the context of each usage of "that" and retrain the
parser.
[0009] Building a particular application on top of and NLP parser
requires building a module to transform the NLP parser output into
usable data. The application using the parser's output could be
coded in a programming language, use a rule based systems or be
trained with machine learning techniques--or created with
combination of any of the above methods.
[0010] Using NLP parsers can be challenging due to the need to
understand the structure of the output and parameters (requires
expert knowledge). One of the challenges in NLP parsers is to
provide constant consistent structure of information. Also, the
output of the NLP parsers rely on the quality of the input text
data.
[0011] For example,
[0012] Let's consider these sentences [0013] 1. John likes math.
[0014] 2. John likes to learn. [0015] 3. John likes learning math
in the evening.
[0016] By using grammar parsers in each case you will get different
notations for the object what John likes.
[0017] In constituency parsers the number of levels (depth) in
parse tree depends of the length and the grammatical structure of
the processed sentences. In a given example the first sentence has
3 levels, the second sentence has 5 levels and the third example
has 6 levels in a tree representation.
[0018] In state-of-the-art dependency parsers the structure of the
output and number of levels in the dependency tree representation
also vary. Adding even one word in the sentence can alter the
grammatical information of all the other words.
[0019] The given example about John would produce different
structure for each sentence. The first sentence require extracting
dependents of "dobj" relation connected to the word "likes", in
second all dependents of "xcomp" relation connected to the word
"likes" and in third example there is a need of analyzing all
governors connected to dependents of "xcomp" related to the word
"likes".
[0020] All of the above is the reason why it is difficult for
people and especially non-linguists (developers, analysts) to use
the parser output and write rules to adjust it to their current
needs. E.g. to write information extraction engine to extract
information about product features from reviews you could use
constituency or dependency parser but you need to write complex
algorithms to search through the parse tree. To move to another
domain (e.g. extracting information from twitter) the algorithms
must be redesigned, and part of the code rewritten.
[0021] To deal with these problems NLP systems use machine learning
techniques. This approach has some limitations in terms of accuracy
and amount of extracted information.
[0022] There are query languages to process structured data (e.g.
SQL for relational databases, Cypher for graph databases,
SPARQL-RDF tagged texts (resource description framework)) but there
are no languages designed directly to query structure of the
natural language (output of the NLP parser).
[0023] It would be desirable to have an efficient framework of
storing information decoded from text. It should provide invariant
and consistent way of storing information which would be
insensitive to different types of input. Having such a framework,
would be possible for non-experts to write efficient rules on top
of NLP parser's output.
[0024] It would be desirable to have a parser for natural language
processing that is built fully algorithmically so it allows for
constantly improvement in accuracy, and the addition of new
features, without building or re-annotating any corpus. It would
desirable to have an NLP system that is more capable than current
NLP parsers of dealing with non-typical grammatical input, deals
well with constantly-changing language on the web, and produces
accurate output which can be stored into an efficient framework of
information.
[0025] It would also be desirable to have a query language that can
be used on the logical layer across different input contexts
allowing humans to write efficient rules for extracting information
and is capable of effectively leveraging many NLP systems.
BRIEF DESCRIPTION OF THE DRAWINGS
[0026] FIG. 1 is a block diagram of as prior art natural language
processing system.
[0027] FIG. 2 is a block diagram of a natural language processing
environment according to an embodiment.
[0028] FIG. 3 is at block diagram of a natural language processing
system according to an embodiment.
[0029] FIG. 4 is at block diagram of a natural language processing
system according to an embodiment.
[0030] FIG. 5 is a block diagram of word tagger according to an
embodiment.
[0031] FIG. 6 is as flow diagram of as word tagger updating or
exception handling method according to an embodiment.
[0032] FIG. 7 is a block diagram of clause divider elements and
processes according to an embodiment.
[0033] FIG. 8 is a flow diagram of a clause divider updating or
exception handling method according to an embodiment.
[0034] FIGS. 9-14 are diagrams of a clause divider process
according to an embodiment.
[0035] FIG. 16 is a block diagram of accomodator according to an
embodiment.
[0036] FIG. 17 is a flow diagram of the "it" test according to an
embodiment
[0037] FIGS. 18-23 are diagrams illustrating the operation of the
accomodator according to an embodiment.
[0038] FIG. 24 is a flow diagram of Language Decoder updating or
exception handling method according to an embodiment.
[0039] FIG. 25 is a diagram of a three-level output structure for
LD 203 according to an embodiment
[0040] FIG. 26 is a diagram illustrating the operation of a
language decoder (LD) according to an embodiment.
[0041] FIG. 27 is a diagram illustrating an overview of an
information extraction methodology according to an embodiment.
[0042] FIGS. 28-34 illustrate a user interface of a knowledge
browser application according to an embodiment.
DETAILED DESCRIPTION
[0043] Embodiments of inventions disclosed herein include
improvements on current NLP systems and methods that are especially
relevant to processing input that consists of plain text from
different types of context. As such, the embodiments disclosed
herein provide a highly accessible platform for natural language
processing. The output of the disclosed NLP system is easily
leveraged by various software applications to analyze large
quantities of text from any source in a more sophisticated and
flexible manner than previously possible.
[0044] The Language Decoder is a novel fully algorithmic NLP parser
that is decoding information out of text and store it into
three-level framework which is capable of handling various type of
texts from different domain like reviews, news, formal documents,
tweets, comments etc.
[0045] The algorithmic nature of the system allows to achieve high
accuracy on user generated content.
[0046] Embodiments of the NLP system can thus work properly on
different kinds of domains at the same time.
[0047] This three-level hierarchical framework of processed text is
leveraged by embodiments of a language decoder query language LDQL
as further described herein. The LDQL is particularly easy to use
for developers, without requiring specific linguistics training.
However, other rule-based query languages could conceivably be
developed for extraction (by query) of information from text
processed by the LD.
[0048] FIG. 2 is a block diagram of as natural language processing
environment 200 according to an embodiment. A natural language
processing (NLP) system 200 accepts text as input. Text can include
electronic data from many sources, such as the Internet, physical
media (e.g. hard disc), a network connected data base, etc. The NLP
system 200 includes multiple databases 202A and multiple processors
202B. Databases 202A and processors 202B can be located anywhere
that is accessible to a connected network 208, which is typically
the Internet. Databases 202A and processors 202B can also he
distributed geographically in the known manner. Data sources 210
include: 1) any source of electronic data that could serve as a
source of text input to NLP system 202, and 2) any source of
electronic data that could be searched using the output of NLP
system, 202 as described further below.
[0049] Other systems and applications 106 are systems, including
commercial systems and associated software applications that have
the capability to access and use the output of the NLP system 202
through one or more application programming interface (APIs) as
further described below. For example, other systems/applications
106 can include an online application offering its users a search
engine for answering specific queries. End users 212 includes
individuals who might use applications 106 through one or more of
end user devices 121A. User devices 212A include without
limitations personal computers, smart phones, tablet computers, and
so on. In some embodiments, end users 121 access NLP system 202
directly through one or more APIs presented by NLP system 202.
[0050] FIG. 3 is a block diagram illustrating NLP 202 in more
detail of according to an embodiment. Input text is received by
Language Decoder (LD) 203. LD 203 parses and tags (among other
things) the input text as further described below. The output of
the LD 203 is referred to here as LD output API 205, and can be
operated on by a variety of downstream software applications or
systems as described in this document. The format of LD output 205
is described in further detail below, but it is easily accessible
to various machine learning, system 104, which can process it to
provide input to yet other systems/applications 106. Other
systems/applications 106 can also directly use LD output 205. Once
the format of LD output 205 is understood, anyone can access
desired information from the original input text.
[0051] The LD output 205 can also be operated on by embodiments of
an LD query language (LDQL) 204. LDQL 204 is described according to
various embodiments below as an improved query language designed to
take advantage of LD output 205. However LDQL 204 can also operate
on the output of any prior NLP system. Also provided in various
embodiments is a higher level of APIs (as compared to LD output
205) for providing machine learning systems 104 and other
systems/applications 106 more intuitive access to LD output 205.
Other systems/applications 106 can include semantic databases,
Freebase.TM., Wordnet.TM., etc.
[0052] In general, LD APIs 206 are relatively easy to use to for
other systems seeking to manipulate LD output 205. However, machine
learning systems 104 and other systems/applications 106 can also
directly write queries using LDQL 204.
[0053] FIG. 4 is block diagram showing more detail of an NLP system
202 according to an embodiment. The LD 203 includes a word tagger
203A which is further described below. The word tagger 203 includes
a tokenizer 203A1, and a sentence divider 203A2. The word tagger
203A communicates its output to a predicate finder 203B. The LD 203
also includes a clause divider 203C and an accommodator 203D.
[0054] FIG. 26 is a diagram summarizing the operation of the LD 203
is an example according to an embodiment. In this example, the text
input is the sentence "John is a nice guy and Susanne likes him."
Tokenizer and sentence divider (203A1 and 203A2) divide the input
text into separate words grouped into sentences (a single sentence
is shown here as an example). A corrector module 2604 cooperates
with the tokenizer 203A1, the sentence divider 203A2 and the word
tagger 203A and applies rules to improve the text input, for
example by fixing typos in the input. A named entity recognition
module 2602 locates and classifies elements such as names of
persons, names of organizations, names of locations, etc.
[0055] The word tagger 203A labels each word with a tag from the
proprietary tagset previously described. The predicate finder 203B
builds predicates from verb components. Then the clause divider
203C joins words into scraps, groups them into separate clauses,
and determines relationships and types of those clauses. Next the
accommodator 203D converts scraps within clauses into phrases and
determines relationships and types of those phrases. At the end of
the process, the accommodator determines types and relationships of
each word within each phrase. The output 205 of the accommodator as
shown is in a human readable form in contrast to tree-style output
of various prior systems. The original sequence of the input words
is preserved and the relationships, types, tags, etc. are simple to
view.
[0056] Word Tagger Methodology
[0057] Contrary to all prior-art word taggers, present word tagger
does not use any of machine learning techniques and contains only
human-written disambiguation rules.
[0058] Word taggers rules are described by a language resembling
well-known regular expressions, which is comprehensible and easy to
change for human operators. Hence, the process of changes is
relatively fast, especially in comparison to the prior-art machine
learning based algorithms.
[0059] A set of rules is divided into two groups (referred to
herein as "exact group" and "inexact group") along with different
usages within the tagging algorithm.
[0060] Contrary to machine learning approaches, present word tagger
uses its own dictionary to recognize words and assign all possible
tags to them. It is proven that word taggers using machine learning
techniques achieves higher accuracy for tagging known tokens
(words) than unknown. The term "known token" means that token
appeared in training corpus at least once. Differential in accuracy
for known and unknown tokens is about 7-8% for good parsers (see
for example
http://nlp.stanford.edu/pubs/CICLing2011-manning-tagging.pdf).
Embodiments of the presented word tagger allow new words to be
added directly to the referenced dictionary with all possible tags.
For example, "google" can be added with tags related to noun as
well as verb. The present word tagger contains rules responsible
for resolving that kind of ambiguity, so it can automatically
handle that problem.
[0061] Word Tagger Output
[0062] Present word tagger provides unique tagset.
[0063] Most important differences between present tagset and
prior-art tagsets: [0064] The present tagset does not exclude
particles as a separate tag. The definition of particle is very
inconsistent. Therefore, a distinction for adverbs and prepositions
is preferred. [0065] The present tagset provides distinction for
prepositions and subordinating conjunctions. [0066] For
determiners, the present tagset provides distinction for adjectival
and pronominal function. [0067] For pronouns, the present tagset
provides distinction for relative and interrogative function.
[0068] The process of the Word Tagger
[0069] The input to the word tagger 203 consists of plain text. The
text can come from any source such as article, document, user
generated content and any other source. Referring to FIG. 5, a word
tagging process according to an embodiment is illustrated. In an
embodiment, the input to the word tagger element 103A is text that
is tokenized into sentences and words by preceding elements
tokenizer 203A1 and sentence divider 203A2. A dictionary 502 is
accessed to find words with possible tags. All possible tags are
assigned to each word (504). The words assigned with all possible
tags are input to a dehomonymizer 506 which resolves ambiguities
(i.e. situations in which for one word there is more than one
possible matching tag). The dehomonymizer 506 operates using exact
rules 508 and inexact rules 510 as further described below. After
ambiguities (if any) are resolved, tagged text 512 is output by the
word tagger 203. The word tagger output consists of text tokenized
into sentences and words, each of which has assigned exactly one
tag.
[0070] Exact rules 508 and inexact rules 510 are two groups of
rides. Each rule examines the context of a word with an ambiguous
tag (that is, more than one possible tag). Such context is not
limited or restricted; it can address any element or the text, in
both directions at the same time. This is a range aspect of the
claimed invention describing the facility to access or examine any
element of the text before or after an ambiguous element, without
limit.
[0071] Another aspect of the word tagger 203A is the manner of
expression employed. In order to examine the context of a word,
rules permit the creation of patterns and sub-patterns which
address words and/or their tags and which can be subject to
following operations; [0072] conditionality (for single element or
subpattern) (corresponding to "?" regular expression operator),
[0073] alternatives (for single element or subpattern)
(corresponding to "|" regular expression operator), [0074]
repetition (for single element or subpattern) (corresponding to "+"
and "*" regular expression operator), [0075] negation for single
element or subpattern) (corresponding to "!" regular expression
operator).
[0076] This form of rules expands their expressiveness in
comparison to prior static look-ups and ranges ("next word is . . .
", there is a . . . word in range of 3 words before" etc.) This in
effect allows better description of situations in which a given
rule should apply.
[0077] In an embodiment, the rules are applied separately for each
sentence according to following algorithm:
[0078] 1. Stop algorithm at any point when there is no ambiguity
left in current sentence.
[0079] 2. For all ambiguous words try to apply all exact rules in
proper order.
[0080] 3. Apply inexact rules until any is met. Go to step 2.
[0081] Methodology of Improving the Word Tagger and Resolving
Exceptions
[0082] FIG. 6 is a flow diagram illustrating a method 600 of
improving the word tagger or modifying the word tagger as
exceptions occur according to an embodiment. Exceptions include
failure of the word tagger to tag an element of input text.
[0083] At 602 an exception indication is received within the system
202. It is determined which word caused the exception (604). The
indicated word is searched for in the dictionary 502 (606). If the
word is not in the dictionary, it is added to the dictionary at
608, then check if there is an existing set of tags for the word
612--if so check if the problem is solved 618, if not--create a new
set of rules for the new set of tags 614 and then check if the
problem is solved 618.
[0084] If, after checking 606, the word is in the dictionary, it is
determined which rule is responsible for the exception, and the
responsible rule is edited or a new rule is added 610. In an
embodiment, the editing is performed by a person. In another
embodiment, the editing is performed by a machine according to a
separate rule-making algorithm.
[0085] If the exception is resolved (at 618), the process ends. If
the exception is not resolved, the process returns to 604 to
examine which word caused the exception.
[0086] The above methodology of resolving exceptions to the system
may become an automated proces.
[0087] Clause Divider Methodology
[0088] Referring to FIG. 4, the LD 203 includes a predicate finder
203B and a clause divider 203C. The output of the word tagger 203A
is received by the predicate finder 203B. Predicate is understood
here as a function over arguments, i.e. it consists of a main verb
grouped with other verb components like modals, auxiliaries,
negations and particles, contrary to traditional approach where
predicate also contains other sentence elements like objects,
predicative expressions and adjuncts. Predicate finders 203B can
include any of the known predicate finder methods. The output of
the predicate finder 203B is received by the clause divider 203C
which performs a clause divider methodology according to
embodiments.
[0089] Embodiments of the clause divider 203C comprises an
algorithm emulating human reasoning behind dividing text into
clauses. To this end, a sequential approach is employed rather than
applying patterns to an entire sentence. In effect, collections of
one or more words (scraps) are sequentially considered (which is
similar to what a human actually does while hearing or reading a
sentence) and an attempt is made to assign each scrap to build a
clause "on the fly" constantly trying and rejecting different
alternatives. This approach simplifies aspects of language
decoding, such as handling of nested relative clauses and
facilitates the creation of simpler and more accurate rules than
those based on patterns.
[0090] Output of the Clause Divider
[0091] According to embodiments, the clause divider provides
division for clauses. Clauses can possess at the most one predicate
(either finite or non-finite like infinitives, participles and
gerunds) and all of its arguments. LD provides connections between
clauses based on criteria concerning their function towards their
superiors. Clauses can be connected directly or through the nodes.
Nodes are sentence elements provided to connecting clauses, e.g.
coordinating or subordinating conjunctions. LD provides unique
classifications of clauses which correspond to the LD system 202
classification of phrases. Main clauses are distinguished. Other
clauses can function as subjects, objects, complements and
attributes, and therefore those can he labeled with proper function
name (e.g. attribute clause).
[0092] In an embodiment, the clause divider includes at least the
following characteristics:
[0093] With reference to one of the current, typical dependency
parsers:
[0094] 1. Embodiments provide implicit division into clauses. In
referenced parser, derivation from the relations structure is
required.
[0095] 2. Coordinating nodes (elements connecting two coordinated
clauses (e.g. "and", "but", "or") are distinguished by embodiments.
Referenced parser does not provide a distinction between a relation
connecting two coordinated clauses and a relation connecting two
coordinated phrases or words. In addition, elements connecting a
subordinated clause to its superior clause (e.g. "which", "after",
"when") are distinguished by embodiments. In typical dependency
parser the main connection holds between two verbs representing the
clauses; hence the equivalent of a subordinate node has to be
connected with a proper verb.
[0096] 3. In the clause divider 203C, classification is based on
different criteria (than in the typical dependency parser) and in
consequence not every type from one classification is equivalent to
a subset of the types of the other classification (although in a
number of situations it is).
[0097] 4. Types of clauses in the LD 202 are equivalent to types of
phrases, making the LD system 202 more coherent in comparison to
the typical parser. For example, some of referenced parser's
relations corresponding to clause types are common with those for
phrases and/or words.
[0098] With reference to one of the current, typical constituency
parsers:
[0099] 1. Typical constituency parser provides classification that
is based on criteria concerning grammatical construction. This is
in sharp contrast to the clause divider 203C, whose criteria is
based on a clause's function towards its superior. In effect every
type from one classification can match almost any type from the
other, depending on the particular situation.
[0100] 2. The LD 202 treats nodes connecting clauses (e.g. "and",
"but", "which", "after") as separate elements on a clause level,
whereas the referenced parser includes them into the following
clause.
[0101] The Process of the Clause Divider
[0102] FIG. 7 is a diagram showing elements of a clause divider
203C and various tasks performed by the clause divider 203C.
Generally, the clause divider input is text divided into sentences
and tokenized into words, along with their tags as assigned by word
tagger 203A. Grammatical information for each sentence is provided
by predicate tinder 203B. The tagged word input is received by a
scrapper element or module 702. The scrapper 702 groups the tagged
words into scraps, which are groups of one or more words that are
strictly connected, but do not necessarily from a phrase.
[0103] More precisely, a scrap is a maximal set of words that will
certainly form a phrase either by itself or along with some other
scrap(s). That is, if two elements are able to form two separate
phrases, but this is not evident at the point of creating scraps,
they become separate scraps. There is an exception to the
foregoing. Specifically, prepositions are joined in scraps with
following noun phrases in principle, despite the fact that they
will later be divided into two separate phrases.
[0104] At 804, the clause divider 203C makes a preliminary estimate
of which scraps are the strongest candidates to introduce a new
clause (that is, which scraps are responsible for dividing the
sentence into clauses. The output of the preliminary estimate
process 704 is received by a main divider engine 706, which
includes a scrap dispatcher and a correctness evaluator. The scrap
dispatcher assigns scraps to appropriate clauses. In an embodiment,
this assignment is based on constant interaction between the scrap
dispatcher and the correctness evaluator. The correctness evaluator
evaluates the decision made by the scrap dispatcher by determining
whether the decision generated as correct clause or not. The output
of the main divider engine 706 is received by a nodes extraction
process 708. This process extracts nodes as separate elements on a
clause level. Relationships between clauses are established by as
clause connection detection process 710. Clause types are detected
(e.g. subject, object, complement) by a clause type classification
process 712. The output 714 of the clause divider is text divided
into sentences and clauses, which contain words (along with their
tags from the word tagger) grouped into scraps. The grouping serves
only as an aid in prior and subsequent processes and is not itself
a crucial element of the output. The words can also be considered
as assigned directly to clauses. Moreover each clause has its
proper grammatical information assigned, along with type and
connection to its superior clause.
[0105] FIGS. 9-14 are diagrams illustrating a process of the clause
divider 203C module according to an embodiment. Referring to FIG.
9, the input to the clause divider module 203C is the sentence "The
man I saw last night became my friend." with all words tokenized
and tagged by the Word Tagger 203A module and with predicates
grouped and distinguished by the Predicate Finder 203B module.
Scrapper 702 groups words into scraps. For example, determiners are
joined with nouns, hence "the man" becomes one scrap (analogous
with "my friend"). Referring to FIG. 10, the preliminary estimation
of dividing scraps 704 begins from the last word of a sentence.
Predicate "became" is found and added to the predicate stack.
Analogous with next predicate "saw". The introduction of dependent
clause is detected and therefore predicate "saw" is removed from
the top of predicate stack. The first word is reached and the
preliminary estimation ends. Referring to FIG. 11 the main divider
engine 706 sequentially tries to assign scraps to clauses. First
scrap "the man" is assigned to the first clause of temporary stack
by default. Second scrap "I" is assigned to the second clause of
temporary stack, because the introduction of dependent clause was
detected between the first two scraps. Scrap "saw" is assigned to
the second clause of temporary stack. Referring to FIG. 12 the main
divider engine 706 assigns the scrap "last night" to the second
clause of temporary stack. Once the scrap "became" is reached, the
clause from the top of the temporary stack is considered as closed,
because "became" is a predicate scrap and the clause already
possess a predicate scrap "saw". Hence the clause is moved to the
final stack and "became" is added to the top clause of the
temporary stack. The scrap "my friend" is added to the top clause
of the temporary stack. Because all of the scraps were processed,
all clauses from temporary stack are moved to the final stack and
the process of assigning scraps ends. Referring to the FIG. 13
nodes are extracted. In the considered example there are no scraps
that function as nodes, hence the process leaves the clause
unaltered. Afterwards the connections of clauses are detected.
Because the clause "I saw last night" is interjected right after NP
scrap, we consider it dependent clause and hence determine
connection from it to "The man became my friend" clause. Moreover
in this situation connection between clause "I saw last night" and
phrase "The man" is established. The clause connection detection
ends. Referring to FIG. 14 the clause "The man became my friend"
has no connection to superior clause hence it obtains "main" type.
The clause "I saw last night" is connected to superior clause and
to superior phrase as well, thus it obtains "complement" type. The
process of the Clause Divider 203C ends and returns output composed
of sentence "The man I saw last night became my friend" which
contains two clauses: "The man became my friend" and "I saw last
night" along with their types, connections and respective
scraps.
[0106] Methodology of Improving the Clause Divider 203C and
Resolving Exceptions
[0107] By analyzing eases of wrong divisions, connections and types
if clauses or adjusting the clause divider 203C into different type
of content one can decide whether to: [0108] add or modify new
divider candidates. [0109] add or modify rules in choosing possible
alternative divisions. [0110] add or modify rules in correctness
evaluation.
[0111] This provides flexibility and continuous improvement for
different types of input context.
[0112] FIG. 8 is a now diagram illustrating method 800 of improving
the clause divider or modifying the clause divider as exceptions
occur according to an embodiment. Exceptions include failure of the
clause divider to divide clauses or determine the proper type or
connection of the clauses.
[0113] At 802 an exception indication is received within the system
202. It is determined if the clause is correctly divided (804). If
not, it is determined if clause dividing scraps were chosen
correctly (806). If not, a new candidate is added or the rules
responsible for choice are changed (808).
[0114] After the changes (808) or in case if it was determined,
that the dividing scraps were chosen correctly (806), it is
determined if scraps were correctly distributed to clauses (810).
If not, scrap dispatcher or correctness evaluator rules are
adjusted (814).
[0115] After the changes in rules of scrap dispatcher or
correctness evaluator (814), or in case if it was earlier
determined that either scraps were correctly distributed to clauses
(810) or clause is correctly divided (808), it is determined if
connections between clauses are correct (812). If not, the rules
governing connection determination are adjusted (816).
[0116] After adjusting the determination rules (816) or if the
connections between clauses were determined as correct (812) it is
determined if the clause classifiaction is correct (818). If yes,
the process ends. Otherwise the clause classification rules are
adjusted (820).
[0117] The above methodology of resolving exceptions to the system
may become an automated proces.
[0118] Accomodator Methodology
[0119] Referring briefly to FIG. 4, the output of clause divider
203C is received by an accomodator module 203D. FIG. 16 is a block
diagram of an accomodator module 203D architecture according to an
embodiment. In an embodiment, the accomodator 203D uses a set of
types of phrases (corresponding to simple grammatical types such as
predicate, subject, object, complement etc.). The accomodator
employs a sequential algorithm, which in one embodiment is based on
hand-written rules, but embodiments are not so limited. Maintaining
control over rules and having the ability to modify and add rules
to the system at will enables the LD 203 to work properly on
different kinds of domains (or language contexts) at the same time.
This tight control over rules and the adaptability of the rules
also provides the ability to handle new domains while avoiding any
distortion in the original or prior domain(s). A separate semantic
database is referenced to resolve ambiguities with connections
(e.g., a semantic n-gram database 1603). In addition, the
accomodator 203D module uses a method for determining connections
between phrases (the "it" test described below with reference to
FIG. 17).
[0120] The accomodator 203D receives words grouped into scraps and
further into clauses (along with connections and types). In other
embodiments the accomodator is based on words grouped directly into
clauses. A phrase creator module 1602 detects boundaries of every
phrase. Then a phrase connection detection module 1604 detects
connections between phrases relying on a semantic n-gram database
1603. A phrase type classification module 1606 denotes types for
each phrase in ever clause in a sentence. Then a word component
classification module 1608 assigns components to each word within
all of the phrases. Finally, as word connection detection module
1610 detects connections between each word within a phrase.
[0121] The accomodator 203D output 1612 consists of words grouped
into phrases, which are further grouped into clauses. The phrases
and words have proper connections and types assigned. In an
embodiment the accomodator 203D output is the final output of the
LD system 203, or LD output API 205 as shown in FIG. 4. As further
described in other parts of this document, the output is useable by
many other systems via APIs or other programming tools or
interfaces. In an embodiment, types of the phrases are determined
after the connections; they are evaluated on the basis of the
contained words (along with their tags), types of the scraps from
which a phrase comes, and types of scraps from which connected
phrases come. Although it has proved efficient and accurate to
first connect the phrases and classify them afterward, it is just
as viable to do it the other way around.
[0122] An advantage of the accomodator over prior modules that
perform analogous tasks is that the accomodator 203D method
determines connections between phrases and their grammatical
functions in way that is similar to the way in which is human
processes language. The accomodator 203D sequentially considers
separate scraps, and chooses, for a set of syntactic possibilities,
the one that semantically makes most sense. In an embodiment this
is achieved in part though using a database 1603 containing n-grams
(representing these syntactic possibilities) along with the
quantity of their occurrence in a large corpus. Ambiguities in
interpretation are always reduced to a set of syntactic
possibilities consisting of a few elements and are then solved on a
semantic basis. The result is simple and intuitive for a human to
understand. Thus, a human can readily see and understand the
decisions the LD 203 makes and, when appropriate, correct its
mistakes by modifying rules so that the system is continually
improved.
[0123] The accomodator employs an unambiguous criterion for
determining connections between phrases, which causes its
performance to be reliable and its output to be easily understood
by humans. This allows the output of the LD 203 to be more readily
incorporated into other existing language applications and their
respective rules. As an example, to determine the connection
between certain prepositional phrases, the accomodator employs the
so-called "it" test, which is illustrated in FIG. 17. An "it"
algorithm according to an embodiment is based on an implementation
of the "it" test aided by the semantic n-gram database 1603.
Referring to the example of FIG. 17, consider the sentence
(consisting of one clause) "I recommend coffee with sweetener for
your father." To determine if a given clause element is superior,
it is replaced with "it". If the original meaning of the sentence
is not preserved, the replaced element is indeed the superior
element. Otherwise, the replaced element is not the superior
element. As shown, for this clause, it is determined that "coffee"
is a superior phrase to prepositional phrase "with sweetener".
[0124] FIGS. 18-23 are diagrams illustrating a process of the
accomodator 203D module according to an embodiment. Referring to
FIG. 18, the input to the accomodator module 203D is the sentence
"Construction worker bought coffee with milk and cigarettes."
Phrases are created by minor manipulation of scraps, such as
merging, dividing and changing the type of scraps. Then phrases are
connected. In this instance, "construction worker" is first
connected to a predicate because of its position. The "construction
worker" phrase is added to a stack of phrases. In an embodiment,
only NP and P phrases can be added to the stack. Referring to FIG.
19, and continuing phrase connection, the predicate "bought" is by
default connected to the root. The "construction worker" phrase is
removed from the stack because of the appearance of the predicate.
Then "coffee" is connected to the predicate because of noun-verbal
proximity, and this phrase is put on the stack. In the following
operation, two situations are considered. Either "with" connects
the phrase on the right with one from the top of the stack, or
"with" connects the phrase on the right with the predicate. The
first solution is better semantically, therefore this one is
selected. Next, "milk" is connected to the "with" because it stands
immediately after a preposition, and "with milk" is added to the
stack. Referring to FIG. 20 and continuing with phrase detection,
two further situations are considered. "And" links the phrase on
the right with one of the phrases from the stack: either "coffee
and cigarettes" or "with milk and cigarettes". These two pairs
force the following semantic relationships: "bought cigarettes" and
"coffee with cigarettes", respectively. The first makes more sense
semantically, therefore it is chosen, and it takes over the
connection to "bought" from "coffee" and links it to itself.
"Cigarettes" is then connected to the "and" phrase because it
stands immediately after a CN (connector) phrase.
[0125] Referring to FIG. 21, the phrases are then classified. The
"Construction worker" phrase gets SUBJECT for its type because of
its position and connection to the predicate. The predicate phrase
type was assigned earlier by the clause divider module 203C. The
"coffee" phrase gets OBJECT for its type because of its position
and relationship to the predicate. The "with" scrap type is changed
to a PREPOSITION phrase type. Referring to FIG. 22, the
classification of phrases continues. The "milk" phrase gets
ATTRIBUTE for its type because of as relationship to a preposition
phrase. "And" gets CONNECTOR for its type because it links other
phrases. "Cigarettes" gets OBJECT for its type because of its
relationship to a particular connector.
[0126] The next operation is word categorization. The word
"construction" gets a specifier component because it is a noun
specifying a phrase core. Referring to FIG. 23, word categorization
continues. The rest of the words in the sentence are classified as
core components because they are single within corresponding
phrases and are not modifying any other phrases. Next, words are
connected within phrases. All specifying words are connected to the
core of the respective phrase. Other words in the sentence are not
considered here because they are single within their corresponding
phrases.
[0127] Methodology of Improving Accomodator and Resolving
Exceptions
[0128] FIG. 15 is a flow diagram illustrating method 1500 of
improving the accomodator or modifying the accomodator as
exceptions occur according to an embodiment. Exceptions include
wrong detection of phrase boundaries or failure of determining type
of phrase or its connection.
[0129] At 1502 an exception indication is received within the
system 202. It is determined if the phrase is built correctly
(1504). If not, it is determined if the problem results from phrase
creator (1506). If so, the rules of creating phrases are adjusted
(1508). If it was determined that the problem of wrongly built
phrase does not result from phrase creator (1506), determine if the
problem results from scraps (1510). If so, repair scrapper engine
(1514). If the scrapper engine was repaired (1514), the rules of
creating phrases were adjusted (1508) or if it was determined that
the problem does not result from scraps (1510) or that the phrase
was built correctly (1504), it is determined if the connections of
the phrases are correct (1512). If no, the connection rules are
changed or the semantic n-gram database is edited (1518). After
these changes (1518) or if the connection between phrases is
determined as correct (1512), it is determined if the type of the
phrase is correct (1520). If so, the process ends. Otherwise the
rules for determining type of phrase are adjusted (1522).
[0130] The above methodology of resolving exceptions to the system
may become an automated proces.
[0131] Methodology for Improving Language Decoder and Resolving
Exceptions
[0132] FIG. 24 is a flow diagram illustrating method 2400 of
improving the Language Decoder or modifying the Language Decoder as
exceptions occur according to an embodiment. Exceptions include
failure of the Language Decoder on any level: tokenization,
sentence division, word tagging, predicate finder, clause divider
or accomodator.
[0133] At 2402 an exception indication is received within the
system XX. It is determined if the input text is correctly
tokenized (2404). If not, adding or changing existing rules in
tokenizer is performed (2408).
[0134] After the changes (2408) or in case if it is determined,
that text is correctly divided into sentences (2406). If not,
adding or changing existing rules in sentence divider is performed
(2412).
[0135] After the changes (2412) or in case if it is determined,
that word tagger worked correctly (2410). If not, word tagger
exception resolving process is started (600).
[0136] After the changes (600) or in case if it is determined, that
predicate finder worked correctly (2414). If not, adding or
changing existing rules in predicate finder is performed
(2420).
[0137] After the changes (2420) or in case if it is determined,
that clause divider worked correctly (2418). If not, clause divider
exception resolving process is started (800).
[0138] After the changes (800) or in case if it is determined, that
accommodator worked correctly (2432). If not, accommodator
exception resolving process is started (1500).
[0139] The above methodology of resolving exceptions to the system
may become an automated process.
[0140] LD Output as a Three-Level Framework of Information
Interface
[0141] The output of the LD module 203 is a three-level framework
of storing information (words, phrases and clauses). The
three-level structure enables efficient organization of
information. Embodiments of a three-level LD output framework
capture a maximum amount of the information coded into text while
maintaining the simplest possible structure for easy access. The
first step was to design efficient framework for storing
information and then using only algorithms written by humans (no
machine learning techniques and corpus) to decode the information
from any given text. The output of the LD thus conveys the logical
structure of information stored in text, along with the grammatical
structure.
[0142] The three-level LD output structure is compositional and
predictable and consist of a relatively small number of elements on
each level. Having a minimal set of components and relations speeds
up the learning process, simplifies writing extraction rules, and
can reduce chances of miscategorization. The output of the LD is
effectively an easy-to-access interface standardizing the
information coded into natural language.
[0143] Description of the Three-Level Framework
[0144] A three-level structure consisting of word-level,
phrase-level and clause-level stores the information coded in text.
The LD 203 is a three-level structure, invariant of the text
processed through the system. The text could be short, simple
sentences written in proper grammar, short non-proper grammar
twitter content or long non-proper grammar reviews. This contrasts
with prior systems, in which the parameters are responsible for
attributes and the structure at the same time, which can produce
varying results depending on the context of some fragment of
text.
[0145] In an example from BACKGROUND Section: [0146] 1. John Likes
Math.
[0147] 2. John Likes to Learn.
[0148] 3. John Likes Learning Math in the Evening.
[0149] When using LD parser in each case one gets consistent
notation for the object which John likes. In the first example
"math" is an object, in the second example "to learn" is object
clause (its role is the same, but on different level) and in the
third example "learning math in the evening" is also an object
clause. This approach allows to separate grammatical layer from
logical layer, so that a single rule can cover many different
syntactic structures of the sentence.
[0150] As a result information extraction rules written on top of
the LD are efficient. Fewer rules need to be written to capture
more information, and the information is less ambiguous.
[0151] FIG. 25 is as diagram of three-level output structure for LD
203 according to an embodiment. In an embodiment a word is a
minimal unit of information. Combinations of words (possibly also
one-element) form a reference to some entity, concept, action etc.
The word's meaning can only be understood in context of the phrase
to which it belongs and this is described in the LD by a word
component. The connection further indicates which other word inside
the phrase is directly concerned by the role of the former. In an
embodiment, a phrase is a group of words which acts as a single
unit referring to some entity, action etc. A combination of phrases
in as clause can produce a minimal statement that can be part of as
more elaborate message. The internal meaning of a clause neglecting
its context, which is managed by clause level) is determined by
contained phrases and their relations. The role of given phrase in
the clause is represented in the LD by the phrase type and the
connection indicates to which other phrase this type relates.
[0152] In an embodiment, a clause is a group of phrases that form a
single minimal statement. Its internal meaning can he analyzed
separately (and this is done by the LD on a phrase level), but only
its context in combination with other clauses can lead to
understanding the text and its meaning. This is caused by the fact
that the original, literal meaning of the clause is often
significantly altered by its relation towards some other clauses.
The LD provides such information about relations between clauses as
well as their functions in the whole utterance.
[0153] The three-level framework consists of elements on each level
and allows for the addition of layers storing new kinds of
information. This 3-level structure allows to integrate the decoded
information with semantic databases (e.g. freebase.TM.,
wordnet.TM.), ontologies taxonomies in order to add additional
semantic information into existing components on the phrase and/or
word level.
[0154] The clause level has additional layer of information about
the tense (e.g. present simple, past simple) and construction (e.g.
positive, negative, question). In embodiment, different layers of
information are kept separate for several reasons. Layers of
abstraction can be easily formed by combining only the relevant
types of information and ignoring the other (e.g. if one needs only
information about phrases division and not their types). It is
possible to add other, new layers (e.g. coreference) in addition to
the existing ones without distorting the already present
information thanks to the separation of layers.
[0155] The three level framework can be seen and treated as an
interface for information extraction from text. The information can
be coded into different layers of the text. For example, in "He
owns that red ear" the information about the car is described as
"red" on the word level. In "The boy threw that ball" the
information about which object was thrown is coded on a phrase
level. In "It started to rain after we got home" the circumstances
are stored on the clause level. It is also possible to include
higher levels on which the information can be stored, e.g. the
causation can be seen as pattern of clauses connected with certain
nodes ("[If] you eat too much chocolate [then] your stomach will
hurt"), therefore the information is coded into higher level than
clause. Likewise it is possible to include yet higher levels,
building patterns on top of other patterns. The three-level
framework of the LD output reflects that natural division of
information into separate levels and provide fundaments for
creating higher level patterns to capture information store in
multiple sentences.
[0156] Language Decoder Query Language
[0157] 1. General Description and Purpose
[0158] LDQL 204 is a declarative, domain-specific querying language
for text-structuring NLP systems outputs' information extraction.
It is based on first-order predicate calculus and enables users to
predefine their own formulas (including recursive definitions).
[0159] LDQL 204 queries can express a wide range of actions--from
simple listing all subjects to actions as complicated as e.g.
finding opinions about people.
[0160] LDQL 204 queries have SQL-like syntax, which is relatively
easy to write and read for human operator, and does not require
linguistic knowledge from them. The queries can be also created
automatically, which is described below.
[0161] Our implementation of LDQL 204 is optimized for Language
Decoder, however LDQL 204 can also be implemented for other
text-structuring NLP systems, e.g. Stanford Parser or Illinois
SRL.
[0162] It is possible (e.g. with the use of LD 203 and LDQL 204) to
formulate queries in natural language, and then translate them to
[high-level] LDQL 204 queries.
[0163] 2. LDQL in Comparison to Other Query Languages
[0164] There are some prior-art query languages which were used for
NLP (e.g. Prolog, SPARQL, IQL). As compared with them, LDQL 204 has
some unique features: [0165] it was designed for querying NLP
parsers' output: it's types and relations reflect 3-level
information structure, regardless of particular parser's output
structure, [0166] it is based on full first-order predicate
calculus, [0167] users can define their own formulas (including
recursive ones) within the language, and use them in queries, or
other definitions.
[0168] LDQL 204 queries have SQL-like syntax: a SELECT section
containing goals of extraction, an optional FROM section for fixing
the search range (search by clauses, sentences, or whole text at
once), and an optional WHERE section, containing restrictions on
extracted objects in the form of (syntactically sugared)
first-order formulas, where the variable's range over 3-level
structure of words, phrases and clauses (regardless of particular
parser's output structure).
[0169] E.g. to extract pairs of subject-predicate phrases from
given text we could use the following LDQL 204 query:
TABLE-US-00001 SELECT P:sub AS Actor, P:prd AS Action FROM clauses
WHERE sub.phrase_type=`subject` AND prd.phrase_type=`predicate` AND
sub->prd
[0170] LDQL's orientation towards formula composition encourages
users to building new formulas out of previously defined ones,
which we consider crucial in dealing with natural language's
complexity; for example instead of the following pair of
queries:
TABLE-US-00002 SELECT P:x,P:y WHERE x.phrase_type=`subject` AND
y.phrase_type=`object` AND x.phrase_core_base=`crocodile` AND
EXISTS P:z (z.phrase_type=`predicate` AND x->z AND y->z);
SELECT P:q,P:w WHERE q.phrase_type=`subject` AND
w.phrase_type=`object` AND w.phrase_core_base=`lamb` AND EXISTS P:z
(z.phrase_type=`predicate` AND q->z AND w->z)
[0171] We would rather abstract the common pattern for finding
"subject.fwdarw.predicate.rarw.object" connection:
TABLE-US-00003 DEFINE subobj_connection(P:s,P:o) AS
s.phrase_type=`subject` AND o.phrase_type=`object` AND EXISTS P:p
(p.phrase_type=`predicate` AND s->p AND o->p) SELECT P:x,P:y
WHERE x.phrase_core_base=`crocodile` AND subobj_connection(x,y);
SELECT P:x,P:y WHERE y.phrase_core_base=`lamb` AND
subobj_connection(xy)
[0172] which makes the queries more "granular", and thus easy to
maintain (in analogy with subroutines in imperative programming
languages).
[0173] As far as we're concerned, this is the first use of full
first-order query language for unstructured text information
extraction. The presence of EXISTS quantifier (in addition to the
Boolean connectives AND, OR and NOT present in e.g. SQL, Prolog
or--in some flavor--SPARQL) with the ability to write recursive
formula definitions makes LDQL 204 expressively stronger formalism
than e.g., pure SQL, Prolog or even SPARQL (which seems to have
some quantification constructs, but no recursive formula
definitions). An example benefit of having such an expressive power
(also present in e.g. Godel, or--indirectly--Cypher, or any
imperative language like Java) is the ability to describe closures
of relations:
[0174] suppose we would like to know whether two phrases are linked
by a sequence of one or more ".fwdarw." connections. In LDQL 204 we
could simply write recursive formula:
TABLE-US-00004 DEFINE linked(P:x,P:y) AS x!=y AND (x->y OR
EXISTS P:z (z!=x AND z!=y AND x->z AND NOT z->x AND
linked(z,y)))
[0175] Which literally reads "x and y linked if x.fwdarw.y or
x.fwdarw.z for some z, such that z is linked with y"--notice the
circularity (recursion). Because of LDQL's expressive strength,
some measures need to he taken to avoid infinite evaluations (the
"x.fwdarw.z AND NOT z.fwdarw.x" and "z!=x AND z!=y"
restrictions).
[0176] 3. Merging LDQL with External Systems
[0177] LDQL 204 could be implemented to operate on the output of
any NLP text-structuring system (e.g. Stanford Parser), provided
that the implementation will contain accessors to attributes that
given system offers.
[0178] LDQL 204 can be connected to external semantic bases (e.g.
Freebase, Wordnet), lexical data bases (e.g. Wordnet),
domain-specific data bases, ontologies, word banks, taxonomies,
etc., in order to support it with more semantic content.
[0179] LDQL 204 rules can be hand-written, generated automatically
or semi-automatically by either connecting LDQL 204 to sources of
semantic content (as above), or by machine learning means (in case
of LD's output it is possible to extract common structures out of
annotated text corpora by means of unification heuristics or e.g.
Koza's genetic programming), or both, or generated
semi-automatically by merging any of the methods mentioned.
[0180] 4. LDQL Implementation
[0181] The LDQL 204 implementation for LD 203 consists of three
modules: parser, optimizer and compiler, as shown on FIG. 35.
[0182] The parser module 3508 processes text of LDQL script 3502
(i.e. list of definitions 3504 and queries 3506) in order to
produce its abstract representation.
[0183] The optimizer module 3510 takes the parser's output and
performs additional annotations to guide compiler.
[0184] The compiler module 3512 takes annotated abstract
representation and generates as php/c/some-imperative-language
output program 3514 which, given valid NLP parser output 103
returns the query's results 3516 for that output.
[0185] FIG. 27 is a diagram illustrating a process of information
extraction from plain text according to an embodiment. This is an
example of what is achieved by a software application (generally an
example of a software application 106) written to apply LDQL 204
rules for a specific purpose of extracting information from
multiple plain text online reviews. The input text "The chicken in
this restaurant is very delicious and the music is awful! I want to
say the sea view is awesome!" is operated on by 203, which outputs
(2702) the sentences in the form of naked API 205, which as
previously described maintains the original word sequence.
Previously chosen LDQL rules are applied (2704) to the output of
the LD process, and the result is a two-column table (subject and
complement) 2706 that lists the sea view is awesome, the chicken is
very delicious, and the music is awful. This tabulated output is
easily accessed for further operations.
[0186] With using output of Language Decoder it is possible to
build language understanding applications or language understanding
engines. It can be done by applying on top of LD one or more of the
following: LDQL rules, raw code (e.g. PHP, Java, C#, .NET, C++),
machine learning algorithms (supervised or unsupervised).
[0187] Usage of Language Decoder
[0188] Language Decoder is the component technology for building
application in many areas including (but not limited to): [0189]
Information Extraction [0190] Sentiment Analysis and Extraction
[0191] Event and Relationship Extraction [0192] Opinion Mining
[0193] Text Mining [0194] Document indexing [0195] Text
Summarization/curation of information [0196] Speech processing
[0197] Question&Answering [0198] Text proofing [0199]
Translations [0200] Natural Language Search/Structured Search
[0201] Query expansion [0202] Automated scoring (essay scoring)
[0203] Language Decoder may be used to process any type of texts
including (but not limited to): [0204] User generated content,
[0205] Social media content, microblogging (e.g. twitter) [0206]
Reviews (e.g. review sites content, websites, wikipedia, etc.)
[0207] Formal language documents [0208] Articles and new [0209]
Biomedical free text [0210] Research papers
[0211] System built on top of language decoder can represent many
domains including (but not limited to): [0212] Big data [0213]
Customer feedback analytics [0214] Social listening/sentiment
analytics [0215] Email analysis [0216] Text Analytics [0217] Search
[0218] Voice Search [0219] Advertising [0220] Predictive Analytics
[0221] Google Glasses and other reality devices [0222] Voice
interface
[0223] Information Extraction Example: Knowledge Browser
[0224] Bubble visualization is a concept of showing multi-level
structure of knowledge representation (LDQL extracted) in a form of
clickable bubbles.
[0225] This representation can vary in size, color and position of
bubbles representing frequency and additional properties (for
example, color palette can represent polarity of detected object
sentiments).
[0226] This multilevel approach allows to browse (e.g. zoom in or
zoom out) through summarized knowledge encoded in analysed text,
which can help to understand wider context along with more exact
properties of extracted data.
[0227] This concept does not restrict data to be in specific format
or order, pairs (object--opinion), triplets, quadruples
(suggestion, aim, suggestion modifier, aim modifier) and many more
are suitable here. Also additional categorization can be applied on
top of extracted data to improve knowledge integrity and
understanding. It can be achieved in multiple ways, starting from
applying external lexicons or semantic databases (e.g. Freebase,
FrameNet), through human made categorization to even logical
correlations.
[0228] Bubble visualization concept is used in the information
extraction example application described below.
[0229] FIGS. 28-34 illustrate a user interface of a knowledge
browser application. In this example, the knowledge browser is for
San Francisco restaurants, but any other subject matter is
possible. The knowledge browser is an embodiment of an application
106 that can help people to make quickly learn, what they want to
know about San Francisco restaurants. By summarizing some reviews
about eateries to a very user-friendly form of so-called bubbles,
users can browse through multiple food-oriented categories and find
exactly what people said about particular aspects of their
experiences in restaurants. The knowledge browser application
illustrates some of the possibilities for applying LDQL 204 rules
to user-written reviews processed by the LD 204. The rules used in
this example are built to extract pairs of entities (object and
related opinion) from user-generated reviews.
[0230] With reference to FIG. 28, after choosing a particular
restaurant (by search input or by clicking on one from a list of
popular restaurants), the application displays a set of bubbles
grouped into five clusters. The central bubble of each cluster
represents a different meta-category. Each meta-category has
different sets of category bubbles attached to it. This set depends
on what people wrote in reviews about chosen restaurant.
[0231] With reference to FIG. 29, after clicking on category
bubble, the application displays a set of extracted objects
attached to the chosen category (FOOD in this example). Bubbles can
differ in size and color to visually convey information. In this
example, the size of a bubble represents object frequency in
processed user reviews. More frequent objects appear on bigger
bubbles. The present example shows that people wrote more about
meal and dish than course or amount. The color of a bubble can vary
from green through orange to red to represent the sum of positive
and negative opinions about the particular object. In the present
example object portion has relatively more negative opinions than
object brunch so it is closer to being red in color.
[0232] With reference to FIG. 30, after clicking on an object
bubble, the application shows the set of opinions about the chosen
object (portion in this example). Bubbles can differ in size and
color here as well. Here, the size of a bubble or radius represents
opinion about the particular object frequency in processed user
reviews. In this example small portion appeared more frequently in
reviews than good portion. The color of the opinion bubble is
constrained to three values: red, green and grey. Red opinions are
meant to be negative, green positive and grey are treated as
neutral. In this example opinions about size of a portion are
rather negative (small, tiny, extremely small) but opinions about
general quality of portion are certainly positive (good, perfect,
healthy).
[0233] Referring to FIG. 31, double clicking on an opinion bubble
shows a general view of every object, which was connected to the
chosen opinion. This view gives a better insight into what
generally, despite categorization, was related to the chosen
opinion (good in this example). For example, the user can look up
what was good in the selected restaurant. Bubbles showed in this
screen have the same properties (color and size) as in previous
steps.
[0234] Referring to FIG. 32, a similar view can be obtained by
clicking on a central category bubble on the screen of FIG. 28.
This displays a graphic of a set of opinions about an entire
category (FOOD in this example), which means that all of those
opinion bubbles are related to some objects categorized as
FOOD.
[0235] Referring to FIGS. 33 and 34, the application provides an
option for comparing two restaurants. By placing two instances of
bubbles next to each other, the user can browse through categories
and objects in one instance (representing one restaurant) and
simultaneously choices are transferred to the second instance
showing differences.
[0236] Bubble size and bubble color are just examples of visual
properties that can be used to convey the desired information. Any
other visual characteristic that can be varied in a similar manner
would be just as appropriate.
[0237] The various functions or processes disclosed herein may be
described as data and/or instructions embodied in various
computer-readable media, in terms of their behavioral, register
transfer, logic component, transistor, layout geometries, and/or
other characteristics. Computer-readable media in which such
formatted data and/or instructions may be embodied include, but are
not limited to, non-volatile storage media in various forms (e.g.,
optical, magnetic or semiconductor storage media) and carrier waves
that may be used to transfer such formatted data and/or
instructions through wireless, optical, or wired signaling media or
any combination thereof. Examples of transfers of such formatted
data and/or instructions by carrier waves include, but are not
limited to, transfers (uploads, downloads, e-mail, etc.) over the
internet and/or other computer networks via one or more data
transfer protocols (e.g., HTTP, FTP, SMTP, etc.). When received
within a computer system via one or more computer-readable media,
such data and/or instruction-based expressions of components and/or
processes under the system described may be processed by a
processing entity (e.g., one or more processors within the computer
system in conjunction with execution of one or more other computer
programs.
[0238] Aspects of the systems and methods described herein may be
implemented as functionality programmed into any of a variety of
circuitry, including programmable logic devices (PLDs), such a
field programmable gate arrays (FPGAs), programmable array logic
(PAL) devices, electrically programmable logic and memory devices
and standard cell-based devices, as well as application specific
integrated circuits (ASICs). Some other possibilities for
implementing aspects of the system include: microcontrollers with
memory (such as electronically erasable programmable read only
memory (EEPROM)), embedded microprocessors, firmware, software,
etc. Furthermore, aspects of the system may be embodied in
microprocessors having software-based circuit emulation, discrete
logic (sequential and combinatorial), custom devices, fuzzy
(neural) logic, quantum devices, and hybrids of any of the above
device types. Of course the underlying device technologies may be
provided in a variety of component types, e.g., metal-oxide
semiconductor field-effect transistor (MOSFET) technologies like
complementary metal-oxide semiconductor (CMOS), bipolar
technologies like emitter-coupled logic (ECL), polymer technologies
(e.g., silicon-conjugated polymer and metal-conjugated
polymer-metal structures), mixed analog and digital, etc.
[0239] Unless the context clearly requires otherwise, throughout
the description and the claims, the words "comprise," "comprising,"
and the like are to be construed in an inclusive sense as opposed
to an exclusive or exhaustive sense: that is to say, in a sense of
"including, but not limited to." Words using the singular or plural
number also include the plural or singular number respectively.
Additionally, the words "herein," "hereunder," "above," "below,"
and words of similar import refer to this application as a whole
and not to any particular portions of this application. When the
word "or" is used in reference to a list of two or more items, that
word covers all of the following interpretations of the word: any
of the items in the list, all of the items in the list and any
combination of the items in the list.
[0240] The above description of illustrated embodiments of the
systems and methods is not intended to be exhaustive or to limit
the systems and methods to the precise forms disclosed. While
specific embodiments of, and examples for, the systems components
and methods are described herein for illustrative purposes, various
equivalent modifications are possible within the scope of the
systems, components and methods, as those skilled in the relevant
art will recognize. The teachings of the systems and methods
provided herein can be applied to other processing systems and
methods, not only for the systems and methods described above.
[0241] The elements and acts of the various embodiments described
above can be combined to provide further embodiments. These and
other changes can be made to the systems and methods in light of
the above detailed description.
[0242] In general, in the following claims, the terms used should
not be construed to limit the systems and methods to the specific
embodiments disclosed in the specification and the claims, but
should be construed to include all processing systems that operate
under the claims. Accordingly, the systems and methods are not
limited by the disclosure, but instead the scope of the systems and
methods is to be determined entirely by the claims.
[0243] While certain aspects of the systems and methods are
presented below in certain claim forms, the inventors contemplate
the various aspects of the systems and methods in any number of
claim forms. For example, while only one aspect of the systems and
methods may be recited as embodied in machine-readable medium,
other aspects may likewise be embodied in machine-readable medium.
Accordingly, the inventors reserve the right to add additional
claims after filing the application to pursue such additional claim
forms for other aspects of the systems and methods.
* * * * *
References