Conduent Business Services, LLC

Patent Trials and Appeals Board

Sep 17, 2021

2020001089 (P.T.A.B. Sep. 17, 2021)

UNITED STATES PATENT AND TRADEMARK OFFICE
UNITED STATES DEPARTMENT OF COMMERCE
United States Patent and Trademark Office
Address: COMMISSIONER FOR PATENTS
P.O. Box 1450
Alexandria, Virginia 22313-1450
www.uspto.gov
APPLICATION NO. FILING DATE FIRST NAMED INVENTOR ATTORNEY DOCKET NO. CONFIRMATION NO.
15/178,929 06/10/2016 Claude Roux 20160072US01-
CNDT3344US01
2926
144578 7590 09/17/2021
FAY SHARPE LLP / CONDUENT
1228 Euclid Avenue, 5th Floor
The Halle Building
Cleveland, OH 44115
EXAMINER
SONIFRANK, RICHA MISHRA
ART UNIT PAPER NUMBER
2674
NOTIFICATION DATE DELIVERY MODE
09/17/2021 ELECTRONIC
Please find below and/or attached an Office communication concerning this application or proceeding.
The time period for reply, if any, is set in the attached communication.
Notice of the Office communication was sent electronically on above-indicated "Notification Date" to the
following e-mail address(es):
Conduent.PatentDocketing@conduent.com
USPTO@faysharpe.com
PTOL-90A (Rev. 04/07)
UNITED STATES PATENT AND TRADEMARK OFFICE
_______________

BEFORE THE PATENT TRIAL AND APPEAL BOARD
_______________

Ex parte CLAUDE ROUX and JULIEN PEREZ
_______________

Appeal 2020-001089
Application 15/178,929
Technology Center 2600
_______________

Before LARRY J. HUME, JAMES W. DEJMEK, and
SCOTT E. BAIN, Administrative Patent Judges.

DEJMEK, Administrative Patent Judge.

DECISION ON APPEAL
Appellant1 appeals under 35 U.S.C. § 134(a) from a Final Rejection of
claims 1 and 4–24. Appellant has canceled claims 2 and 3. See Amdt. 2
(filed Apr. 27, 2018). We have jurisdiction over the remaining pending
claims under 35 U.S.C. § 6(b).
We reverse.

1 Throughout this Decision, we use the word “Appellant” to refer to
“applicant” as defined in 37 C.F.R. § 1.42 (2018). Appellant identifies
Conduent Business Services as the real party in interest. Appeal Br. 1.
Appeal 2020-001089
Application 15/178,929

2
STATEMENT OF THE CASE
Introduction
Appellant’s disclosed and claimed invention generally relates to
natural language generation in an automated dialog system. Spec. ¶ 1.
According to the Specification, dialog systems typically comprise three
parts: (i) a natural language understanding (NLU) module; (ii) a dialog
manager (DM); and (iii) a natural language generation (NLG) module.
Spec. ¶ 2. “The NLG module is used to generate a surface realization (i.e.,
grammatical text, understandable by people) of a dialog act, such as a
question, confirmation, or affirmation, expressed in a representational form.”
Spec. ¶ 2.
An exemplary dialog system also comprises a database containing so-
called text snippets. Spec. ¶ 3. Text snippets are short sequences of words
that may be assembled “to formulate a specific dialog act.” Spec. ¶ 3. “For
generating the surface realization, the text snippets often need to be
transformed into questions, and the task of the NLG module is to determine
how to phrase the question.” Spec. ¶ 3. A trained NLG model assigns an
utterance label to a text snippet from a knowledge base; the utterance label
being used “to guide the generation of the question from the text snippet.”
Appeal Br. 7
Claim 1 is illustrative of the subject matter on appeal and is
reproduced below with the disputed limitations emphasized in italics:
1. In an automated dialog system for conducting a dialog
with a human user, a method for natural language generation
comprising:
providing a natural language generation model which has
been trained to assign an utterance label to a text sequence that
is not in interrogatory form, the utterance label being based on
Appeal 2020-001089
Application 15/178,929

3
features extracted from the text sequence, the trained model
being a sequential decision model selected from a Conditional
Random Field model, a recurrent neural network model, and a
combination thereof, the utterance label being selected from a set
of utterance labels which have been learned by the natural
language generation model using a training set of text sequences
that are not in an interrogatory form, each of the learned utterance
labels including a sequence of at least one word, the sequence
including an auxiliary verb;
receiving a user utterance from the human user;
processing the user utterance to detect missing
information, the processing including selecting a new text
sequence from a knowledge base, wherein the new text sequence
is not in an interrogatory form, the knowledge base including
descriptions of problems and corresponding solutions, the new
text sequence being from one of the descriptions;
extracting features from the new text sequence;
assigning an utterance label from the learned utterance
labels to the new text sequence, based on the extracted features,
with the trained natural language generation model;
generating a natural language utterance in an interrogatory
form from the new text sequence, using the assigned utterance
label to guide the generation of the natural language utterance in
the interrogatory form; and
outputting the natural language utterance in the
interrogatory form to the human user, wherein the extracting and
generating are performed with a processor.

The Examiner’s Rejections
1. Claims 1, 4–9, 11–19, 22, and 24 stand rejected under
35 U.S.C. § 103 as being unpatentable over Mengle et al. (US 2016/0132501
A1; May 12, 2016) (“Mengle”); Venkatapathy et al. (US 9,473,637 B1;
Oct. 18, 2016) (“Venkatapathy”); Mabotuwana et al. (US 2017/0177795 A1;
June 22, 2017) (“Mabotuwana”); and Nina Dethlefs et al., Conditional
Appeal 2020-001089
Application 15/178,929

4
Random Fields for Responsive Surface Realisation using Global Features,
Proceedings of the 51st Annual Meeting of the Association for
Computational Linguistics, 1254–63 (2013) (“Dethlefs”). Final Act. 7–13.
2. Claim 10 stands rejected under 35 U.S.C. § 103 as being
unpatentable over Mengle, Venkatapathy, Mabotuwana, Dethlefs, and
Yuzhu Wang et al., A Light Rule-based Approach to English Subject-Verb
Agreement Errors on the Third Person Singular Forms, 29th Pacific Asia
Conference on Language, Information and Computation: Posters, 345–53
(2015) (“Wang”). Final Act. 14.
3. Claim 23 stands rejected under 35 U.S.C. § 103 as being
unpatentable over Mengle, Venkatapathy, Mabotuwana, Dethlefs, and
Abhinav Rastogi, Context Encoding LSTM CS224N Course Project, 1–7
(2015) (“Rastogi”). Final Act. 14–15.
4. Claims 20 and 21 stand rejected under 35 U.S.C. § 103 as being
unpatentable over Venkatapathy, Mengle, Mabotuwana, and Dethlefs. Final
Act. 15–17.

ANALYSIS2
In rejecting claim 1, the Examiner finds Mengle teaches, inter alia, a
natural language generation model which has been trained to assign an
utterance label to a text sequence. Final Act. 7 (citing Mengle ¶¶ 30, 50–51,
70, Figs. 1–4). The Examiner further finds Mabotuwana teaches a trained

2 Throughout this Decision, we have considered the Appeal Brief, filed
August 6, 2019 (“Appeal Br.”); the Reply Brief, filed November 12, 2019
(“Reply Br.”); the Examiner’s Answer, mailed October 29, 2019 (“Ans.”);
and the Final Office Action, mailed March 18, 2019 (“Final Act.”), from
which this Appeal is taken.
Appeal 2020-001089
Application 15/178,929

5
model being a sequential decision model and explains that one of ordinary
skill in the art would have been motivated to modify Mengle’s model with
the sequential decision model of Mabotuwana “so the model could be used
for future use.” Final Act. 8–9 (citing Mabotuwana ¶ 20). Still further, the
Examiner relies on Dethlefs to teach a particular sequential decision model
may be selected from a Conditional Random Field (CRF) model, a recurrent
neural network model, and a combination thereof. Final Act. 9 (citing
Dethlefs 11). The Examiner determines it would have been obvious to
modify the Mengle-Mabotuwana trained model to be a CRF model as taught
by Dethlefs because “CRFs are able to take the global utterance context into
account and are less constrained by local features than other realisers.”
Final Act. 9.
Appellant disputes that the ordinarily skilled artisan would have been
motivated to combine Mengle, Mabotuwana, and Dethlefs as suggested by
the Examiner, or that such a combination would teach the claimed trained
natural language generation model. Appeal Br. 13–16; Reply Br. 17–18. In
particular, Appellant asserts that rather than using any type of NLG model
for assigning a label to a text sequence, Mengle describes using triples (i.e.,
triplets) to identify missing information. Appeal Br. 13–14. Additionally,
Appellant argues that Mabotuwana does not teach a sequential decision
model, but describes a maximum entropy classifier. Appeal Br. 16.
Appellant asserts that, contrary to a sequential decision model, a maximum
entropy classifier treats the information in a document as a bag of words.
Appeal Br. 16; Reply Br. 17. Moreover, Appellant asserts that Examiner’s
proffered support to modify Mengle’s use of triples with Mabotuwana’s
maximum entropy classifier (i.e., so the model can be used for ‘future use’”)
Appeal 2020-001089
Application 15/178,929

6
lacks any rational underpinning. Appeal Br. 16. Additionally, regarding
Dethlef’s CRF model, Appellant argues “[u]sing a CRF model which has an
‘extended notion of context,’ as described in Dethlefs, would not be useful
in this case since the triples of Mengle’s database lack any context which
could be used.” Appeal Br. 14. In addition, Appellant asserts that because
Mabotuwana does not teach a sequential decision model, “there would have
been no reason to use the CRF model of Dethlefs in place of the maximum
entropy classifier of Mabotuwana.” Reply Br. 17.
We find Appellant’s arguments persuasive of Examiner error.
As identified by the Examiner, Mengle does not describe using a
trained natural language generation model to assign an utterance label to a
text sequence that is not in interrogatory form. Rather, Mengle generally
describes a system for using web resources to determine an answer for an
interrogative query. See Mengle, Abstract. Mengle describes an
interrogative query engine that generates interrogative queries to provide to
a search system. Mengle ¶ 50. More particularly, Mengle describes
receiving a query submitted by a client device, identifying missing
information from the submitted query, and re-writing the query as an
interrogative query based on missing information in an entity database.
Mengle ¶ 51. Mengle teaches the use of a triplet structure (e.g., (subject,
relationship, object)) to identify missing information from the submitted
query and generate (i.e., not select from a knowledge base) an interrogative
query. Mengle ¶ 51.
Mabotuwana, as relied on by the Examiner, describes a natural
language processing technique to extract narrative text from patient medical
imaging studies and reports. Mabotuwana ¶ 20, Abstract. Mabotuwana
Appeal 2020-001089
Application 15/178,929

7
describes using a maximum entropy classifier that assigns an end-of-
sentence character to the extracted content. Mabotuwana ¶¶ 20–24.
We agree with Appellant that Mabotuwana’s maximum entropy
classifier does not teach a sequential decision model, nor has the Examiner
provided objective evidence or technical explanation that Mabotuwana
teaches a sequential decision model. Additionally, we agree with Appellant
that, absent a trained natural language generation model being a sequential
decision model being described in Mengle or Mabotuwana, the Examiner
has not set forth articulated reasoning with rational underpinning as to why
an ordinarily skilled artisan would turn to Dethlefs to use a CRF model. See
In re Kahn, 441 F.3d 977, 988 (Fed. Cir. 2006) (cited with approval in KSR
Int’l Co. v. Teleflex, Inc., 550 U.S. 398, 418 (2007)). Although one of
ordinary skill in the art may understand that two references could be
combined as reasoned by the Examiner, this does not imply a motivation to
combine the references. Personal Web Techs., LLC v. Apple, Inc., 848 F.3d
987, 993–94 (Fed. Cir. 2017); see also Belden Inc. v. Berk-Tek LLC, 805
F.3d 1064, 1073 (Fed. Cir. 2015) (“[O]bviousness concerns whether a
skilled artisan not only could have made but would have been motivated to
make the combinations or modifications of prior art to arrive at the claimed
invention.”); InTouch Techs., Inc. v. VGO Commc’ns, Inc., 751 F.3d 1327,
1352 (Fed. Cir. 2014).
Because we find it dispositive that the Examiner has not shown by a
preponderance of evidence that prior art teaches or reasonably suggests the
recite trained natural language generation model, we do not address other
issues raised by Appellant’s arguments. See Beloit Corp. v. Valmet Oy, 742
Appeal 2020-001089
Application 15/178,929

8
F.2d 1421, 1423 (Fed. Cir. 1984) (finding an administrative agency is at
liberty to reach a decision based on “a single dispositive issue”).
For the reasons discussed supra, we are persuaded of Examiner error.
Accordingly, we do not sustain the Examiner’s rejection under 35 U.S.C.
§ 103 of independent claim 1. For similar reasons, we do not sustain the
Examiner’s rejection of independent claims 18 and 20, which recite
commensurate limitations. In addition, we do not sustain the Examiner’s
rejections of claims 4–17, 19, and 21–24, which depend directly or indirectly
therefrom.

CONCLUSION
We reverse the Examiner’s decision rejecting claims 1 and 4–24 under
35 U.S.C. § 103.

Appeal 2020-001089
Application 15/178,929

9
DECISION SUMMARY
Claims
Rejected
35 U.S.C.
§
Reference(s)/Basis Affirmed Reversed
1, 4–9, 11–
19, 22, 24
103 Mengle,
Venkatapathy,
Mabotuwana, Dethlefs
1, 4–9,
11–19,
22, 24
10 103 Mengle,
Venkatapathy,
Mabotuwana,
Dethlefs, Wang
10
23 103 Mengle,
Venkatapathy,
Mabotuwana,
Dethlefs, Rastogi
23
20, 21 103 Venkatapathy,
Mengle, Mabotuwana,
Dethlefs
20, 21
Overall
Outcome
1, 4–24

REVERSED