PTAB Affirms Patent Eligibility of Claims for Training a Spoken Language Understanding Classifier

In a decision issued earlier this month, the U.S. Patent and Trademark Office Patent Trial and Appeal Board reversed the final rejection of all twenty pending claims in U.S. Application No. 14/846,486, for which the real party in interest is Microsoft.  The claims at issue are directed to a system that trains a spoken language understanding (SLU) classifier based on user intent gleaned from user utterances (i.e., spoken natural language sentences and phrases, such as "send Mom an email").  In particular, the claimed invention involves collecting a variety of user utterances and semantically parsing the utterances (i.e., mapping the utterances into machine-understandable representations of their respective meanings) to generate a single graph that represents all the utterances in the form of nodes.  The claimed invention then involves clustering (i.e., grouping) the utterances by similar user intent, and using the resulting groups to train the SLU classifier.

The claims had been rejected under 35 U.S.C. § 101 as reciting patent ineligible subject matter in the form of a mathematical calculation.  Representative claim 1 is as follows, and the other independent claims recite similar limitations:

1.  A system for training a spoken language understanding (SLU) classifier, comprising:

one or more computing devices, said computing devices being in communication with each other via a computer network whenever there is a plurality of computing devices; and

a computer program having program modules executable by the one or more computing devices, the one or more computing devices being directed by the program modules of the computer program to,

receive a corpus of user utterances,

for each of the user utterances in the corpus,

semantically parse the user utterance, and


represent the result of said semantic parsing as a rooted semantic parse graph,


combine the parse graphs representing all of the user utterances in the corpus into a single corpus graph that represents the semantic parses of the entire corpus and comprises a root node that is common to the parse graph representing each of the user utterances in the corpus,

cluster the user utterances in the corpus into intent-wise homogeneous groups of user utterances, said clustering comprising finding subgraphs in the corpus graph that represent different groups of user utterances, each of said different groups having a similar user intent, each of the subgraphs being more specific than the root node alone and more general than the full semantic parses of the individual user utterances,

use the intent-wise homogeneous groups of user utterances to train the SLU classifier, and

output the trained SLU classifier.

With the two-part Alice framework in mind, the key point of contention between the Appellant and the Examiner was whether the claim limitations involving the feature of clustering utterances based on user intent (referred to by the Appellant as "utterance intent clustering") are similar to the claim limitations at issue in McRO, Inc. v. Bandai Namco Games Am. Inc., 837 F.3d 1299, 1316 (Fed. Cir. 2016), and thus whether the reasoning in McRO applies in the present case.  To this point, the Appellant argued that, similar to McRO, the claims focus on a specific improvement to a SLU subsystem that is not used by other SLU systems.  The Appellant also argued that the utterance intent clustering feature recited by the claims is used to determine user intent in a way that is not found in conventional spoken dialog computing systems.  For these reasons, the Appellant asserted that the claims amount to significantly more than any abstract idea.

The Examiner's arguments in support of the § 101 rejection were threefold.  First, the Examiner plainly disagreed that the reasoning in McRO is applicable to the claims at issue.  Second, the Examiner asserted that the utterance intent clustering is a mathematical calculation, and thus the improvement is not to a computer or other technology, but rather to the calculation itself.  Third, in citing to various portions of App. No. 14/846,486 for alleged support, the Examiner asserted that SLU classifiers themselves are mathematical calculations, and thus that any limitations directed to training a SLU classifier are merely field of use limitations and do not add significantly more to the abstract idea of utterance intent clustering.

But the Board disagreed with the Examiner on all three points.  The Board was quick to note that, although the portions of App. No. 14/846,486 cited by the Examiner might describe mathematical calculations, they do not discuss an SLU classifier, but rather discuss a method of developing the graph used to train the SLU classifier.  In addition, the Board stated that the last two steps of claim 1 are more than just field of use limitations.

Yet the Board's reasoning for finding the claims patent-eligible was focused on differentiating the claims from those in McRO.  In McRO, the patent-eligible claims were directed to a system of automating facial animation through the use of a specific set of rules.  The McRO court found the claimed rules-based process was not the same as facial animation processes conventionally performed by human animators, and thus that the rules-based process provided a technological improvement and limited the claims to a specific, non-abstract animation process.  In addition, the McRO court went as far to state that "processes that automate tasks that humans are capable of performing are patent eligible if properly claimed."

Here, although the claimed utterance intent clustering involves mathematical operations, the Examiner did not show that the utterance intent clustering is a process conventionally performed by humans.  In fact, as the Board was sure to note, Appellant's own specification describes how the claimed utterance intent clustering differs from (and improves upon) conventional processes where spoken dialog applications incorporate "a pre-determined set of domains and user intents that are manually designed by domain experts."  Before concluding that the claims are directed to patent-eligible subject matter, the Board provided an additional comparison to McRO:

Further, similar to McRO, the claims do not merely organize information into a new form.  Rather, the utterance intent clustering limitation recites a specific order of steps (parsing utterances, combining all utterances into one graphs with a common root node, and then clustering into intent-wise homogeneous groups) that renders the information in a specific format used to create the desired results.  McRO, 837 F.3d at 1315.  The utterance intent clustering limitation, which produces intent-wise homogeneous groups that are then used to train the SLU classifier, is improving a technological process, as it is improving a specific process by which an SLU classifier is trained.

This decision is somewhat reminiscent of another statement by the McRO court which was not mentioned here by the Board -- particularly that "processes that automate tasks that humans are capable of performing are patent eligible if properly claimed."  Proper claiming of such processes is a line delicately walked by applicants operating in a variety of fields, and the buzzword "machine learning" certainly comes to mind in reading this decision.  Classification and training are leading aspects of machine learning, and the Board's decision here gives further insight into how the Patent Office handles claims in this realm.

Ex parte Hakkani-Tur (PTAB 2018)
Panel: Administrative Patent Judges Allen R. MacDonald, Robert E. Nappi, and James W. Dejmek
Decision on Appeal by Administrative Patent Judge Nappi