U.S. patent application number 13/228617 was filed with the patent office on 2011-12-29 for systems and methods for filtering dictated and non-dictated sections of documents.
This patent application is currently assigned to Dictaphone Corporation. Invention is credited to Alwin B. Carus, Larissa Lapshina, Bernardo Rechea.
Application Number | 20110320189 13/228617 |
Document ID | / |
Family ID | 38445102 |
Filed Date | 2011-12-29 |
![](/patent/app/20110320189/US20110320189A1-20111229-D00000.png)
![](/patent/app/20110320189/US20110320189A1-20111229-D00001.png)
![](/patent/app/20110320189/US20110320189A1-20111229-D00002.png)
![](/patent/app/20110320189/US20110320189A1-20111229-D00003.png)
![](/patent/app/20110320189/US20110320189A1-20111229-D00004.png)
![](/patent/app/20110320189/US20110320189A1-20111229-D00005.png)
United States Patent
Application |
20110320189 |
Kind Code |
A1 |
Carus; Alwin B. ; et
al. |
December 29, 2011 |
SYSTEMS AND METHODS FOR FILTERING DICTATED AND NON-DICTATED
SECTIONS OF DOCUMENTS
Abstract
A system and method for filtering documents to determine section
boundaries between dictated and non-dictated text. The system and
method identifies portions of a text report that correspond to an
original dictation and, correspondingly, those portions that are
not part of the original dictation. The system and method include
comparing tokenized and normalized forms of the original dictation
and the final report, determining mismatches between the two forms,
and applying machine-learning techniques to identify document
headers, footers, page turns, macros, and lists automatically and
accurately.
Inventors: |
Carus; Alwin B.; (Waban,
MA) ; Lapshina; Larissa; (Shirley, MA) ;
Rechea; Bernardo; (Belmont, MA) |
Assignee: |
Dictaphone Corporation
Stratford
CT
|
Family ID: |
38445102 |
Appl. No.: |
13/228617 |
Filed: |
September 9, 2011 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
11362646 |
Feb 27, 2006 |
8036889 |
|
|
13228617 |
|
|
|
|
Current U.S.
Class: |
704/9 |
Current CPC
Class: |
G06F 40/20 20200101;
G10L 15/22 20130101; G10L 15/18 20130101; G06F 40/103 20200101 |
Class at
Publication: |
704/9 |
International
Class: |
G06F 17/27 20060101
G06F017/27 |
Claims
1-13. (canceled)
14. A method for filtering dictated and non-dictated sections of
documents, the method comprising: gathering a first set of
documents having dictated and non-dictated section boundaries;
featurizing text in at least one document from the first set of
documents; differentiating dictated and non-dictated sections of
text in the at least one document from the first set of documents;
categorizing text of a second set of documents to identify dictated
and non-dictated sections of text within at least one document from
the second set of documents; and outputting dictated sections of
the at least one document from the second set of documents to an
automatic speech recognition process.
15. The method according to claim 14, wherein the gathering further
comprises pre-processing text in the first set of documents.
16. The method according to claim 15, further comprising
interpreting featurization rules.
17. The method according to claim 16, further comprising generating
featurized data.
18. The method according to claim 16, wherein the differentiating
further comprises creating classification models in order to
distinguish between dictated and non-dictated sections of text in
the at least one document from the first set of documents.
19. The method according to claim 18, wherein the categorizing is
performed based on the classification models.
20. The method according to claim 19, wherein the first set of
documents does not equal the second set of documents.
21-33. (canceled)
34. A system for filtering dictated and non-dictated sections of
documents, the system comprising: a central processing unit; and a
computer code operatively associated with the central processing
unit, the computer code including instructions to cause the central
processing unit to: gather a first set of documents having dictated
and non-dictated section boundaries; featurize text in at least one
document from the first set of documents; differentiate dictated
and non-dictated sections of text in the at least one document from
the first set of documents; categorize text of a second set of
documents to identify dictated and non-dictated sections of text
within at least one document from the second set of documents; and
output dictated sections of the at least one document from the
second set of documents to an automatic speech recognition
process.
35. The system according to claim 34, wherein the instructions
further cause the central processing unit to pre-process text in
the first set of documents.
36. The system according to claim 35, wherein the instructions
further cause the central processing unit to interpret
featurization rules.
37. The system according to claim 36, wherein the instructions
further cause the central processing unit to generate featurized
data.
38. The system according to claim 36, wherein the instructions
further cause the central processing unit to create classification
models in order to distinguish between dictated and non-dictated
sections of text in the at least one document from the first set of
documents.
39. The system according to claim 38, wherein the instructions
cause the central processing unit to categorize the text of the
second set of documents based on the classification models.
40. The system according to claim 39, wherein the first set of
documents does not equal the second set of documents.
Description
BACKGROUND OF THE INVENTION
[0001] The present invention relates generally to a system and
method for filtering sections of documents for automatic speech
recognition processes.
[0002] In the process of converting dictated speech files to
finished reports, transcriptionists typically transform the
original dictation substantially to generate the final report.
Transcriptionists may revise the original text stylistically to
conform to personal and site standards, to re-order the dictated
material, construct lists, invoke text macros to generate tables
and other large text sections, and insert document headings.
Transcriptionists may also act on specific instructions to correct
speech disfluencies such as repeated or fragmentary words.
[0003] Additionally, many electronic documentation systems
automatically insert document formatting and metadata such as
headers, footers, page turns, macros, and default demographic
information. Such automatic formatting varies widely from one
automatic documentation system to another. Moreover, when using
such documentation systems, transcriptionists may actually remove
redundant information from the text file, which may have been added
by an electronic documentation system.
[0004] It has been found that automatically formatted sections are
easily and frequently overwritten. Document sections may be
improperly filled in or modified, where dictated and transcribed
segments may be inserted into macros and tags may be deleted.
Consequently, final transcribed text reports often differ
substantially from the original recorded dictation.
[0005] In many automatic speech recognitions systems however, it is
desirable to maintain a close alignment between the dictated audio
saved in wave files or other formats and the corresponding
transcribed text reports. Such alignment of the wave files and
corresponding text is critical for many tasks associated with
automatic speech recognition systems. In particular, language model
identification (LMID), language model adaptation (LMA), acoustic
model adaptation (AMA), automatic error correction, speaker
evaluation, report evaluation, and post-processing techniques have
been found to benefit from improvements in the alignment of the
dictated wave files and the corresponding transcribed text. These
processes rely on matches between the originally dictated acoustics
and text as what was intended, known as "truth." In most automated
environments, "truth" is unavailable, so finished reports are used
instead to produce "pseudo-truth."
[0006] Such mismatches or misalignments between the original
recorded dictation and the final text report have been found to
degrade automatic speech recognition processes as described above.
It has been found for example that non-filtered non-dictated
sections in finished documents negatively and significantly effect
LMID, most of the LM and AM augmentation processes, and speaker
classification. Bad filtering or a lack of filtering is also a
serious problem for automatic rewrite techniques. In this
connection, it has been found that LMID is often highly inaccurate
when headers and footers are not removed from the documents. For
example, radiology reports have been identified as belonging to
general medicine or mental health domain. In the same example, with
headers and footers filtered out these reports have been recognized
properly as radiology domain reports. Overall, it has been found
that in some cases there is about 5% absolute accuracy degradation
due to the erroneous behavior of LMID.
[0007] Unfortunately, the current state of technology does not
provide a suitable solution for theses issues. For example, current
solutions for these issues include manual rewrites of the documents
using post-processing. This, of course, increases time of
processing for the finished reports as well as the costs associated
therewith.
[0008] Another sensitive process is LMA, especially for narrow
domains like radiology. LMA is a process that includes adjusting
word N-gram counts of the existing LM. The goal of LMA is to make
the existing language model reflect better the specific speaking
style of the particular user or group of users. Traditionally LMA
is performed on text of finished reports which is considered to be
the best available approximation of the way users dictate. It has
been found that leaving the most likely non-dictated sections of
reports in the text submitted for LMA leads to the opposite effect,
when the LM counts are skewed and end up being further apart from
the targeted individual or group specific dictation style.
[0009] As a result, finished reports usually differ substantially
from the original dictation and techniques to bring them into
closer alignment with the original dictation are needed.
[0010] Therefore, there exists a need for an automatic document
section filtering technique. It is desirable that such a technique
is both accurate and automatic. It is also desirable to have such a
technique that does not require intervention by transcriptionists
or other staff since this is not only time-consuming and expensive,
but frequently performed inaccurately and inconsistently.
[0011] There also exists a need for a simple and reliable system
and method of automatic document section filtering to identify the
most likely non-dictated sections of the medical reports in order
to filter them out.
[0012] There further exists a need to determine reliable heuristics
in order to identify non-dictated sections based on alignment of
finished reports against recognition output.
[0013] There also exists a need for a system and method of
automatic document section filtering to filter sufficient amount of
data to train classifiers independent of recognition output capable
of identifying non-dictated sections such headers, footers, page
turns, and macros based on solely text.
[0014] There also exists a need for a system and method of
automatic document section filtering that uses trained models to
classify document sections for documents that are available only in
the text form where recognition output is unavailable.
SUMMARY OF THE INVENTION
[0015] The present invention includes a system and method for
automatic document section filtering. The present invention
includes identifying portions of a text report that correspond to
an original dictation and, correspondingly, those portions that are
not part of the original dictation. The invention includes
comparing tokenized and normalized forms of the original dictation
and the final report, determining mismatches between the two forms,
and applying machine-learning techniques to identify document
headers, footers, page turns, macros, and lists automatically and
accurately.
[0016] In a first aspect, the present invention includes a method
for filtering documents including gathering speech recognition
output and a first set of corresponding documents, conforming at
least one associated document from the first set of corresponding
documents to a selected speech recognition format, comparing the
speech recognition output and at least one associated document,
determining long homogeneous sequences of misaligned tokens from
the speech recognition output and at least one associated document,
detecting boundaries between dictated and non-dictated sections in
at least one associated document and annotating at least one
associated document with the boundaries.
[0017] In some embodiments the conforming step may include
pre-processing at least one of the associated documents and the
comparing step may include performing label smoothing on the
recognition output and at least one associated document. The label
smoothing may be performed using a sliding average with a fixed
window size. In some embodiments, in particular for medical
documents, the window size was determined to be about 3.
[0018] In some embodiments determining long homogeneous sequences
of misaligned tokens may include the step of detecting formatting
anchors and identifying end points of the detected long homogeneous
sequences of misaligned tokens.
[0019] Some embodiments include outputting the dictated sections to
at least one automatic speech recognition process such as language
model identification, language model adaptation, acoustic model
adaptation, automatic error correction, and speaker evaluation.
[0020] Some embodiments include creating classification models in
order to distinguish between dictated and non-dictated sections of
text in at least one associated document. Based on the
classification models, some embodiments include categorizing text
of a second set of documents to identify dictated and non-dictated
sections of text within at least one of the second set of
documents. Additional embodiments may include outputting dictated
sections of at least one document from the second set of documents
to an automatic speech recognition process. The first set of
documents may not equal the second set of documents.
[0021] In a second aspect, the present invention includes a method
for filtering documents, including gathering a set of documents
having dictated and non-dictated section boundaries, featurizing
text in at least one document from the set of documents,
differentiating dictated and non-dictated sections of text in at
least one document, categorizing text of a second set of documents
to identify dictated and non-dictated sections of text within at
least one of the second set of documents and outputting dictated
sections of at least one document from the second set of documents
to an automatic speech recognition process.
[0022] The gathering step may include pre-processing text in the
documents. Some embodiments may include interpreting featurization
rules and generating featurized data.
[0023] In some embodiments the differentiating step includes
creating classification models in order to distinguish between
dictated and non-dictated sections of text in at least one document
where the categorizing may be performed based on the classification
models.
[0024] In a third aspect, the present invention includes a system
for filtering sections of electronic documents to determine
dictated and non-dictated text in the documents having a central
processing unit with computer code operatively associated with the
central processing unit. In some embodiments the computer code may
include a first set of instructions configured to gather speech
recognition output and a first set of documents corresponding to
the speech recognition output, a second set of instructions
configured to conform at least one associated document from the
first set of corresponding documents to a selected speech
recognition format, a third set of instructions configured to
compare the speech recognition output and at least one associated
document, a fourth set of instructions configured to determine long
homogeneous sequences of misaligned tokens from the speech
recognition output and at least one associated document, a fifth
set of instructions configured to detect boundaries between
dictated and non-dictated sections in at least one associated
document and a sixth set of instructions configured to annotate at
least one associated document with the boundaries.
[0025] In some embodiments the second set of instructions may
include pre-processing of at least one of the associated documents.
The third set of instructions may include performing label
smoothing on the recognition output and at least one associated
document where the label smoothing is performed using a sliding
average. In some embodiments the label smoothing is performed using
a window size of about 3.
[0026] In some embodiments the fourth set of instructions includes
detecting formatting anchors and the fifth set of instructions may
include identifying end points of the detected long homogeneous
sequences of misaligned tokens.
[0027] In some embodiments the computer code includes a seventh set
of instructions configured to output the dictated sections to at
least one automatic speech recognition process. At least one
automatic speech recognition process may include language model
identification, language model adaptation, acoustic model
adaptation, smart rewrite and speaker evaluation.
[0028] Some embodiments include an eighth set of instructions
configured to create classification models in order to distinguish
between dictated and non-dictated sections of text in at least one
associated document. There may also be a ninth set of instructions
configured to, based on the classification models, categorize text
of a second set of documents to identify dictated and non-dictated
sections of text within at least one of the second set of
documents.
[0029] Some embodiments of the present invention may include the
computer code with a tenth set of instructions configured to output
dictated sections of at least one document from the second set of
documents to an automatic speech recognition process.
BRIEF DESCRIPTION OF THE DRAWINGS
[0030] While the specification concludes with claims particularly
pointing out and distinctly claiming the present invention, it is
believed the same will be better understood from the following
description taken in conjunction with the accompanying drawings,
which illustrate, in a non-limiting fashion, the best mode
presently contemplated for carrying out the present invention, and
in which like reference numerals designate like parts throughout
the Figures, wherein:
[0031] FIG. 1 is a diagram of a document filtering system according
to one embodiment of the invention;
[0032] FIG. 2 is a diagram of alignment-based filtering system
according one embodiment of the invention;
[0033] FIG. 3 is a diagram of population-based filtering system
according to one embodiment of the invention;
[0034] FIG. 4 is a diagram of a training module process system
according to one embodiment of the invention; and
[0035] FIG. 5 is a diagram of a data evaluation module according to
one embodiment of the invention.
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0036] The present disclosure will now be described more fully with
reference to the Figures in which at least one embodiment of the
present invention is shown. The subject matter of this disclosure
may, however, be embodied in many different forms and should not be
construed as being limited to the embodiments set forth herein.
[0037] In the various embodiments described herein, each such
embodiment may be implemented through the use of a central
processing unit having a computer code mechanism structured and
arranged to carry out the various steps and functions corresponding
to the elements in the various diagrams.
[0038] In one embodiment the invention includes identifying and
selectively processing non-dictated sections of text reports. Such
selective processing may include either removing or ignoring
certain sections of the reports. The final text of reports may
contain substantial document formatting and metadata generated in
the post-dictation, transcription phase of document preparation.
Identifying those portions of the final report that match the
original dictation can substantially improve the accuracy of
language model identification, language model adaptation, acoustic
model adaptation, automatic error correction model learning,
speaker evaluation, report evaluation, and post-processing.
[0039] The invention also includes comparing tokenized and
normalized versions of the original dictation and its corresponding
finished report using a variant of the Levenshtein edit distance
algorithm. Mismatches between the two forms are identified and once
a sufficiently large set of dictation-report pairs and their
associated mismatches have been collected (on the order of the same
number as required for language model adaptation), these data are
used to teach classifiers to identify headers, footers, macros, and
lists. Still other embodiments may employ the IGTree and IB
algorithms of the TiMBL machine-learning classifier system to
generate accurate and compact classifiers.
[0040] The resulting information about the boundaries of headers,
footers, page turns, macros, and lists may be then used to
construct appropriate input data for different processes. For
example, headers, footers, and page turns are removed for all
processes. For LMID nothing further may be removed. For LMA, AMA,
automatic error correction, and speaker and report evaluation,
macros may be removed. It has been found that post-processing
benefits from improvements in identifying lists as described
herein.
[0041] In one embodiment of the present invention, an
alignment-based approach is implemented to correct misalignments,
which may be based on comparison of tokenized finished reports and
recognition output. This approach allows identification of headers,
footers, and text body of the report with a great deal of accuracy
where headers, footers, and text body detected with alignment-based
method can be used for training header, footer, and text body
models that provide reliable classification of the the sections
based on text only. Proper implementation of this approach has
demonstrated section classification accuracy to about 97% or even
higher for site-specific and speaker-specific models. Accuracy of
site-specific models proves to be comparable or sometimes better
than speaker-specific models for small amounts of training
data.
[0042] In another embodiment of the present invention, a best
performing classification method is implemented using line-by-line
classification with two initial tokens of each line as features
used with the context window size 2-2. Such a context window may
include features of two previous lines, two following lines and a
target line combined.
[0043] The present invention may be implemented with relatively few
site reports including text and wave file pairs, while maintaining
sufficient standards to train reliable filters. Such filters may
then be implemented to process text without recognition output.
Speaker specific filters with about 93% accuracy may be trained for
about between 10-20 speaker wave/text pairs.
[0044] In another embodiment of the present invention, there
includes a method to determine reliable heuristics to identify
non-dictated sections based on alignment of finished reports
against recognition output. In some embodiments there is an
assumption that non-dictated sections of the report will be labeled
by the aligner as misalignments (insertions or deletions). Another
assumption may include that normally sections have certain
formatting and lexical features that can be utilized by the present
invention. Such formatting and lexical features may include
sections separated by paragraph or a new line, headers that are
normally expected at the beginning of the report, footers at the
end, and page turns that may have a key word such as "page".
[0045] Using alignment and heuristics, one method may include
filtering a sufficient amount of data to train classifiers
independent of recognition output capable of identifying
non-dictated sections such as headers, footers, page turns, and
macros based on text only. For documents that are available only in
the text form, where recognition output is unavailable, the
invention may use trained models to classify document sections.
[0046] No form of annotation, manual or automatic, is 100%
accurate. As such, and since the terms "header" and "footer" are
meant in a broad sense as sections at the beginning and the end of
the document that are normally not dictated or unlikely to be
dictated, it can be difficult to make a decision about segment
boundaries based on some strict textual or formatting rules.
Moreover, speakers may not always be consistent. Some speakers tend
to never dictate meta-data that usually belongs to headers and
footers. Other speakers may dictate some of the meta-data including
signatures, distribution ("cc") lists, topic of the report,
demographic information, and medical record of the patient. As a
result, the medical report data may demonstrate significant inter-
and intra-speaker variability.
[0047] In addition, the present invention may be applied across
clinic sites with a broad range of medical domains and document
worktypes.
[0048] In some embodiments implementing an alignment-based
filtering method for the detection of headers and footers may
include simple text preprocessing, for example the insertion of
artificial whitespace tokens such as "DUMMYPARAGRAPH" and
"DUMMYNEWLINE". In order to preserve formatting information in the
document, tokenization, simple "anti-post-processing"
normalization, alignment, and alignment-based filtering may be
implemented thereafter.
[0049] The step of alignment-based filtering may include labeling,
based on an alignment, where each token of the finished report is
evaluated and labeled as "Correct" ("C") or "Incorrect" ("I") and
smoothing having a sliding average and window size of 3. A sequence
of three labels may be evaluated as follows;
L.sub.i-1, L.sub.i, L.sub.i+1
[0050] Where the Label L.sub.i is re-evaluated as "Correct" if the
number of "Correct" tokens within the sliding window is greater
than the number of "Incorrect" tokens, and as "Incorrect"
otherwise. The detection of misalignments at the beginning and the
end of the document of 5 tokens and longer may be candidates for
headers and footers.
[0051] Use of anchoring tokens may include where the header/footer
boundary is shifted to the last/first "anchor" token of a
misalignment group. Such anchors may include whitespace tokens such
as the following: "\Paragraph", "\New-Line", "DUMMYPARAGRAPH",
"DUMMYNEWLINE".
[0052] For text-only-based classifiers the following features and
combinations of features may be implemented as follows: [0053] 1.
Line length (in tokens) [0054] 2. Colon rate (Number of ":\colon"
tokens divided by number of tokens in line); [0055] 3. Name rate
(Number of "NAME" tokens divided by number of tokens in line);
[0056] 4. Digit rate (Number of "DIGIT" tokens divided by number of
tokens in line); [0057] 5. Lexical features (N initial tokens of
each line); [0058] 6. Ascending and descending line offset
(position of the line in the document); and [0059] 7. Line
context.
[0060] In preferred embodiments, the present invention may be
implemented by classifying reports line by line. In some
embodiments line length, colon rate, name rate, and digit rate may
be evaluated together as such parameters may not be good enough
alone to be reliable header and footer predictors.
[0061] In still other embodiments line length, colon rate, name
rate, digit rate, and line offset may be evaluated where the line
offset may improve model performance when line features are used
without context. Models based on lexical features with context may
outperform models built with line length, colon rate, name rate,
digit rate, and line offset.
[0062] Initial tokens of each line may be implemented as features
where pure lexical features may perform better for certain site
specific domains. In other sites, accuracy results may be slightly
higher (about 0.4%) when lexical features are combined with other
line characteristics (line length, name rate, digit rate, and colon
rate).
[0063] Context features, such as features of an adjacent line
combined with the features of the target line of the reports, may
also be implemented as classification predictors. In some
embodiments different context window sizes may be implemented as
follows:
[0064] 1-1 (target line feature vector plus features of one line
before and one lie after);
[0065] 2-2 (target line feature vector plus features of two lines
before and two lines after);
[0066] 3-3 (target line feature vector plus features of three lines
before and three lines after); and
[0067] 4-4 (target line feature vector plus features of four lines
before and four lines after).
[0068] In one of the embodiments the optimal value for the context
window size was found to be 2-2, where the target line feature
vector includes features of two lines before and two lines
after.
[0069] Normalization of common data types may reduce lexical
variability of documents revealing their underlying structure. The
lines from two different but similar reports may look different and
as a consequence might be represented by different classification
feature values both before normalization and after.
[0070] However, in order to determine whether normalization is
helpful for the task, two sets of normalization rules may be
applied to the original training and test data. Classification
accuracy on the normalized data may also be compared to the numbers
obtained on the original non-normalized data. Normalization may be
especially effective and significantly improve performance in cases
of the limited feature sets such as the number of tokens, the name
rate, and the digit rate.
[0071] In addition, IGTree and k-nearest neighbor machine learning
algorithms may be implemented for various document conditions. In
certain cases for various feature sets, context size, normalization
levels, implementing the k-nearest neighbor algorithm (where k=2)
may provide unexpected results over implementing the IGTree
algorithm by slightly more than 1%.
[0072] In implementing the various embodiments of the present
invention examples have been provided in the medical domain.
However, it will be understood by those skilled in the art that the
present invention may also be implemented in other domains, such as
the legal field, engineering, and business domains as well. It will
also be understood by those skilled in the art that the invention
disclosed herein and embodied in the attached drawings may be
implemented with physical documents and systems as well as with
electronic documents and systems.
[0073] Specifically referring now to FIG. 1, there is shown a
diagram representative of the invention according to one embodiment
of the invention. The filtering system is shown generally as
reference numeral 10. Finished reports 15 are input to system 10
and decision 20 determines whether corresponding recognition output
is available. When finished reports and recognition output are
available, alignment-based filtering 25 is implemented for
detection of dictated an non-dictated sections of the finished
reports. Alignment-based filtering 25 is discussed in further
detail below in connection with FIG. 2.
[0074] Dictated sections of the reports 35 and non-dictated
sections of the reports 40 are output from the alignment base
filtering process 25 and loaded into training module 70, which
trains the population-based filters for non-dictated sections of
the report. Training module 70 is described in further detail below
in connection with FIG. 4.
[0075] Dictated sections 35 are loaded into ASR processes 50 which
can include LMID, LMA, AMA, automatic error correction, and speaker
evaluation. If there does not exist any recognition output system
10 determines whether population-based filters have been trained
30. If there are no filters and nothing is trained, no filtering is
required 45. The finished reports are used without any processing
or filtering and the reports are input directly to ASR processes
50.
[0076] If population-based site and/or speakers filters are
trained, population-based filtering 60 may be applied.
Population-based filtering 60 is discussed in further detail below
in connection with FIG. 3. Non-dictated sections of the report 75
are filtered and forwarded to training module 70 and to ASR
processes 50. In addition, non-dictated sections of the reports 80
are filtered and forwarded to training module 70.
[0077] Referring now to FIG. 2, there is shown a diagram
representative of alignment-based filtering according to embodiment
of the invention. Alignment-based filtering is represented by box
25 and may include multiple elements and/or processes therein.
Finished reports and recognition output pairs 105 are preprocessed
and memorized and finished reports are tokenized 110 as well.
Alignment of the recognition output against the tokenized finished
documents is then conducted 115. A label smoothing process is
performed 120 using sliding average techniques and where for one
embodiment the window size is about three (3). In alternative
embodiments the window size may be larger than three and other
smoothing or averaging techniques may be used.
[0078] Detection of all the long homogeneous sequences of
misaligned tokens 125 is then performed. Within the long
homogeneous sequences of misalignment, formatting anchors are
detected 130. Formatting anchors may include variants of tokens
indicating the start of new paragraph ("paragraph", "new
paragraph", "next paragraph") or new line ("new line", "next
line").
[0079] Boundaries between dictated and non-dictated sections of the
report are detected 135. In some embodiments the boundaries are
detected based on misalignment sequences and anchor tokens. As a
result the dictated sections are determined and outputted 35 and
inputted into ASR processes 50 as well as the training module 70.
Non-dictated sections of the report are output 40 to the training
module 70.
[0080] Referring now to FIG. 3, there is shown a diagram
representative of population-based filtering according to one
embodiment of the invention. Population-based filtering 60 is
represented by box 60 and may include multiple elements and/or
processes therein. Such elements may include annotated documents
105 which may be preprocessed, normalized and tokenized 110.
[0081] A pool of candidate headers, footers, and other non-dictated
sections can be identified by annotating a collection of documents.
These data can in turn be used to induce (construct) a classifier.
This approach suffers from several difficulties: (a) annotation is
time-consuming and expensive; and (b) non-dictated sections are
user-specific as some speakers tend to dictate signatures, cc
lists, titles or topics of the reports and while others do not.
Speaker habits cannot be identified without comparing text of the
reports to the voice recordings.
[0082] In addition, featurization 205 may be performed with
featurization rules 207 being input from classification model
configuration file 210.
[0083] It will be understood by those skilled in the art that
featurization is a process of transforming, converting, or
re-representing input data in a form which is appropriate and
amenable for computational processes, such as, for example, machine
learning. In this example, the system may identify a set of
attributes, characteristics, or features. These features may be
used to characterize a particular training example or examples when
using multiple or larger sets of data. As a result, when training
for example with the term `this`, the features sought after might
be the individual characters of the word: the fact that it begins
with a consonant and ends with a consonant, that it is a function
word, that the term is short, that the term is uncapitalized, that
the term contains a sibilant, that it is a monosyllabic word, and
so forth. These are attributes or characteristics of the word
`this`, which are not necessarily apparent from the raw data
itself, although such data might contain raw data elements in some
form, but are in a representation of the information in a
systematic, orderly, and input-relevant format that can be
processed, analyzed and used both for generation of training data
sets and evaluation of later instances using exactly the same
featurization process.
[0084] Featurization rules 207 include a set of classification
attributes which are expected to be used for document section
classification with prescribed values of some configurable
attribute parameters. Featurization rules can include instructions
on how a classification target, which may be a line of a specific
report, can be converted into a set of attributes matching the
appropriate classification model. The present invention is
beneficial in that exceptional results in document section
classification may be achieved by classifying line-by-line where
each line of the document is being the classification target. The
best classification features proved to be several initial lexemes
of each target line and several initial lexemes of the preceding
and following lines. In one embodiment, featurization rules may use
3 initial lexemes of the target line, 2 initial lexemes of the 2
previous lines and 2 initial lexemes of 3 following lines.
[0085] Continuing with FIG. 3, juxtaposed with element 210, there
are classification models 213 which outputs to a classifier 220A,
where classifier 220A incorporates an evaluation component.
Featurized data 223 from the featurization process 205 may be
loaded to the classifier 220A. Classification decisions 225 are
then forwarded such that filtering of the documents 230 may be
accomplished.
[0086] The dictated sections 75 are output to ASR processes 50 and
to training module 70. Non-dictated sections 80 are also output to
training module 70. Additionally, updated models 225 may be loaded
from training module into classification models 213.
[0087] Referring now to FIG. 4, there is shown a diagram
representative of training module 70 according one embodiment of
the invention. Training module 70 includes training data 300 (which
may include both dictated and non-dictated sections) that is input
into data evaluation module 305. Data evaluation module 305 is
discussed in further detail in connection with FIG. 5.
[0088] Training data 300 is then selectively preprocessed,
normalized and tokenized 110 before the featurization process 205
is performed. Here again featurization rules 207 may be loaded from
a classification model configuration file 210. In addition,
learning parameters and specifications 213 may be loaded into a
classifier 220B, which classifier includes a learning
component.
[0089] The featurized data 223 is also loaded into classifier 220B,
from which new or uploaded site/speaker classification models 310
may be derived.
[0090] Referring now to FIG. 5, there is shown a diagram
representative of a data evaluation module according to one
embodiment of the invention. Training data 300 is input into data
evaluation module represented in FIG. 5 by box 305. The system
determines whether the population-based models exist 350. If
site-specific population-based models are determined to exist,
system 10 then determines whether there is enough data to train
speaker-specific non-dictated section classification models 355. If
enough data exists, training data 300 is loaded directly to a
training module 375 containing training speaker-specific
population-based training models.
[0091] In the event site-specific population-based models do not
exist, the system determines whether there exists enough data to
train site models 360. If there exists enough data to train the
site models, training data 300 is directed to a training module 380
containing training site population-based models. In the event
there is not enough data to train site models, system 10 continues
to collect data for site models 370 until enough data is
gathered.
[0092] Training module 375 trains non-dictated section
classification models for each speaker based on data from
particular speakers where training module 380 is a site
population-based training module. If enough data does not exist for
a particular speaker, a site-specific non-dictated section
classification model is used.
[0093] It will be apparent to one of skill in the art that
described herein is a novel system and method for automatic
document section filtering. While the invention has been described
with reference to specific preferred embodiments, it is not limited
to these embodiments. The invention may be modified or varied in
many ways and such modifications and variations as would be obvious
to one of skill in the art are within the scope and spirit of the
invention and are included within the scope of the following
claims.
* * * * *