DETAILED ACTION
This communication is in response to the Applicant Arguments/Remarks dated 6/6/2023. Claims 1, 4, 6-8, 11-15, 18-20 are pending in the application.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA . In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
Response to Arguments
Applicant's arguments filed 6/6/2023 have been fully considered but they are not persuasive. Regarding the applicant arguments on pages 9-10 in relating to the amended claim limitations, please see the new combination of references, with the newly columns and lines cited below. The newly cited reference Ghoshal et al. has been applied to teach the argued limitations.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1, 8, 15 are rejected under 35 U.S.C. 103 as being unpatentable over Ghoshal et al. (US 20200125575) in view of Huh et al. (US 11222027).
Specification, para. 20 teaches “the activity data may include search terms entered by users, search results returned to users, user actions taken in relation to the search results, such as whether search results were selected by the user, an amount of time the user viewed the resulting content, whether the user took an additional step, such as applying for a job, purchasing an item, making a reservation, or the like”.
As per claims 1, 8, 15, Ghoshal et al. (US 20200125575) teaches
a method comprising: generating a first multi-session embedding vector by combining a first single-session embedding vector representing a plurality of activity data derived from a plurality of activities undertaken by an end-user during a first activity session and a second single-session embedding vector representing a plurality of activity data derived from a plurality of activities undertaken by the end-user during a second activity session (fig. 4: topic vector space, content recommendation engine, content analysis and embedding, content management and classification system; para. 58: interactions between client device and content recommendation engine may be Internet-based web browsing sessions, or client-server application sessions, during which users access may input original authored content via the client device, and receive content recommendations from content recommendation engine in the form of additional content that is retrieved from the content repository and linked or embedded into the content authoring user interface at the client device; para. 77: extracting features and/or tags from the content in real-time (or near real-time) during the user's authoring session, vectorizing the authored content, and comparing the vector of the original authored content to one or more existing vector spaces in order to identify and retrieve related/associated content from one or more available content repositories; para. 96: the concept vectors for all the words in the text document can be combined to form a weighted concept vector for the text document. The content recommendation engine then may measure the similarity between each word concept vector and the text concept vector. All words above a certain threshold then may be selected as the "keywords" for the document; para. 204: interactions between client devices and the content recommendation system may be Internet-based web browsing sessions, or client-server application sessions, during which users access may input user content (e.g., search terms, original authored content/chats/interactions, etc.) via client devices, and may receive content item recommendations from the content recommendation system. Thus, the contents of a user’s interactions within sessions are vectorized and vectors are combined to form a weighted concept vector/multi-session embedding vector);
generating a first set of probability values for a set of candidate search terms based on the first multi-session embedding vector, the first set of probability values generated by providing the first multi-session embedding vector as input into a machine learning model trained to classify text based on a set of training multi-session embedding vectors generated from historical activity data; generating a recommended search query including one or more search terms with highest valued probability values from the set of candidate search terms; and causing presentation of the recommended search query within a user interface presented on a display of a client device (fig. 5: coffee 90%, tea 80%...flame 5%, feature vector prob. > z%, recommendations; figs. 37, 42: content recommendation system with ranked list of content items; para. 62: the resulting feature vectors 520 may further be narrowed to exclude those having a feature vector probability of less than z %, based on the output of the trained model, resulting in a subset of the retrieved feature vectors; para. 82: allow the content recommendation engine to determine a set of highest ranking feature vectors within the vector space(s) 430, which are most similar in features/characteristics/etc. to the feature vector generated in step 1206 based on the input received; para. 77: extracting features and/or tags from the content in real-time (or near real-time) during the user's authoring session, vectorizing the authored content, and comparing the vector of the original authored content to one or more existing vector spaces in order to identify and retrieve related/associated content from one or more available content repositories; para. 96: the concept vectors for all the words in the text document can be combined to form a weighted concept vector for the text document. The content recommendation engine then may measure the similarity between each word concept vector and the text concept vector. All words above a certain threshold then may be selected as the "keywords" for the document; para. 204: interactions between client devices and the content recommendation system may be Internet-based web browsing sessions, or client-server application sessions, during which users access may input user content (e.g., search terms, original authored content/chats/interactions, etc.) via client devices, and may receive content item recommendations from the content recommendation system. Thus, the contents of a user’s interactions within sessions are vectorized and vectors are combined to form a weighted concept vector/multi-session embedding vector);
Ghoshal does not teach wherein the first single-session embedding vector and the second single-session embedding vector are derived by averaging token embedding vectors for each activity detected during a respective activity session.
Huh et al. teaches wherein the first single-session embedding vector and the second single-session embedding vector are derived by averaging token embedding vectors for each activity detected during a respective activity session (figs. 3-4: display recommended search queries by concept marker/CM recommender; col. 9:10-12: CM recommender may be configured to identify and/or classify concept markers to be provided as suggestions to the user in response to the user query; col. 10:48-63: an average embedding for each concept marker may be generated by averaging the embeddings for each word in the concept; col. 11 :19-23: CM recommender may then consider a candidate concept marker, and may produce a vector by averaging a vector for just the candidate concept marker with a vector corresponding to the stored values associated with the query concept marker; col. 19:30-34: fig. 2, search results associated with the user query are returned. The search results may include documents ranked by probability of relevance, and a list of suggested concept markers). Thus, it would have been obvious to one or ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Ghoshal and Huh in order to effectively identifying user query intents in search sessions and providing users relevant search queries for the users to select as alternative searches if desire.
Claims 6-7, 12-14 and 19-20 are rejected under 35 U.S.C. 103 as being unpatentable over Ghoshal et al. (US 20200125575) in view of Huh et al. (US 11222027) and further in view of Bell et al. (US 20190392082).
As per claims 12, 19, Ghoshal teaches
wherein generating the first single-session embedding vector based on the first word embedding vector and the second word embedding vector (para. 58: interactions between client device and content recommendation engine may be Internet-based web browsing sessions, or client-server application sessions, during which users access may input original authored content via the client device, and receive content recommendations from content recommendation engine in the form of additional content that is retrieved from the content repository and linked or embedded into the content authoring user interface at the client device; para. 77: extracting features and/or tags from the content in real-time (or near real-time) during the user's authoring session, vectorizing the authored content, and comparing the vector of the original authored content to one or more existing vector spaces in order to identify and retrieve related/associated content from one or more available content repositories; para. 96: the concept vectors for all the words in the text document can be combined to form a weighted concept vector for the text document. The content recommendation engine then may measure the similarity between each word concept vector and the text concept vector. All words above a certain threshold then may be selected as the "keywords" for the document; para. 204: interactions between client devices and the content recommendation system may be Internet-based web browsing sessions, or client-server application sessions, during which users access may input user content (e.g., search terms, original authored content/chats/interactions, etc.) via client devices, and may receive content item recommendations from the content recommendation system. Thus, the contents of a user’s interactions within sessions are vectorized and vectors are combined to form a weighted concept vector/multi-session embedding vector)
Ghoshal et al. does not explicitly teach determining an average of at least the first token embedding vector and the second token embedding vector.
Huh teaches determining an average of at least the first token embedding vector and the second token embedding vector (figs. 3-4: display recommended search queries by concept marker/CM recommender; col. 9:10-12: CM recommender may be configured to identify and/or classify concept markers to be provided as suggestions to the user in response to the user query; col. 11:19-23: CM recommender may then consider a candidate concept marker, and may produce a vector by averaging a vector for just the candidate concept marker with a vector corresponding to the stored values associated with the query concept marker; col. 19:30-34: fig. 2, search results associated with the user query are returned. The search results may include documents ranked by probability of relevance, and a list of suggested concept markers). Thus, it would have been obvious to one or ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Ghoshal and Huh in order to effectively identifying user query intents in search sessions and providing users relevant search queries for the users to select as alternative searches if desire.
Ghoshal and Huh do not explicitly teach the limitation token.
Bell et al. teaches
based on the first token embedding vector and the second token embedding vector (figs. 4, 6; para. 36-37: identifying that it is a particular month (e.g., December) and that a particular item is sold in the particular month (e.g., tree ornaments). Accordingly, the scoring module may score query tokens or result candidates higher based on this determination. The training module is configured to receive input data points and include initial training phases in order to make predictions or classifications more robust, the learning module may have to analyze several years or iterations of data to make this inference strong over a threshold, as opposed to only analyzing one year. The learning module identifies patterns or associations based on the frequency of user selections indicated by the statistics (and/or other statistics) in a data structure and provides an associated output. Each token of the plurality of tokens is scored higher than other tokens if it includes a higher frequency of a particular user selection compared to other tokens). Thus, it would have been obvious to one or ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Ghoshal, Huh and Bell in order to effectively identifying word/tokens in user queries in search sessions and providing users the most relevant search queries for the users to select as alternative searches if desire.
As per claims 6, 13, 20, Ghoshal teaches
generating a search query sequence for the end-user the search query sequence including at least a first search query and a second search query executed by the end user, the first search query and the second search query being ordered sequentially in the search query sequence based on a real-time order in which each respective search query was executed; and generating a second recommended search query for the end-user based on the search query sequence, the second recommended search query generated by providing the search query sequence as input into a sequence to sequence model trained based on a set of training search query sequences generated from the historical activity data; and causing presentation of the second recommended search query within the user interface presented on the display of the client device (para. 77: extracting features and/or tags from the content in real-time (or near real-time) during the user's authoring session (which include queries/searches), vectorizing the authored content (also in real-time or near real-time), and comparing the vector of the original authored content to one or more existing vector spaces 430 in order to identify and retrieve related/associated content from one or more available content repositories; para. 82: allow the content recommendation engine 425 to determine a set of highest ranking feature vectors within the vector space(s) 430, which are most similar in features/characteristics/etc. to the feature vector generated based on the input received; fig. 5: coffee 90%, tea 80%...flame 5%, feature vector prob. > z%, recommendations; figs. 37, 42: content recommendation system with ranked list of content items; para. 62: the resulting feature vectors may further be narrowed to exclude those having a feature vector probability of less than z %, based on the output of the trained model, resulting in a subset of the retrieved feature vectors; para. 204: interactions between client devices and the content recommendation system may be Internet-based web browsing sessions, or client-server application sessions, during which users access may input user content (e.g., search terms, original authored content/chats/interactions, etc.) via client devices, and may receive content item recommendations from the content recommendation system. Thus, the contents of a user’s interactions including queries within user sessions are vectorized and vectors are combined to form a weighted concept vector/multi-session embedding vector or series/sequence of queries).
Huh also teaches at figs. 3-4: a sequence of queries are displayed to the user based on user history; col. 16:46-53: training module may be configured to learn from analysis that goes beyond analyzing user queries in isolation, but may learn from an entire interactive user session. For example, training module may analyze a user session from the initial query to the end of the session, analyzing e.g., the initial query, selected concept markers, relevant documents, subsequent queries, etc.; col. 16:46-53: training module may be configured to learn from analysis that goes beyond analyzing user queries in isolation, but may learn from an entire interactive user session. For example, training module 154 may analyze a user session from the initial query to the end of the session, analyzing e.g., the initial query, selected concept markers, relevant documents, subsequent queries, etc.; col. 21:47-56). Thus, it would have been obvious to one or ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Ghosal and Huh in order to effectively providing users relevant search queries for the users to select as alternative searches if desire.
However, Ghosal and Huh does not explicitly teach search query sequence.
Bell teaches
the search query sequence including at least a first search query and a second search query executed by the end user (fig. 10: user selection history data store; para. 22: storing user selection statistics (e.g., click frequency, purchase frequency, skip frequency, etc.) in a statistics data structure and generating a learning model (e.g., a word embedding vector model) based on the user selection statistics for use in executing queries for one or more resources; para. 27, 45: in response to the table being populated at a first time, and a query (e.g., "Toy cars") being issued at a second time (which is subsequent to the first time), a scoring module (e.g., the scoring module 130) may score "remote control car" higher than "push car" when scoring individual tokens and/or search result candidates for the particular query issued; para. 107-108). Thus, it would have been obvious to one or ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Ghoshal, Huh and Bell in order to effectively identifying users query intents and providing users the most relevant search queries for the users to select as alternative searches if desire.
As per claims 7, 14, Ghoshal and Huh does not explicitly teach said claims.
Bell et al. teaches
wherein the first search query was executed by the end-user during the first activity session and the second search query was executed by the end-user during the second activity session (para. 36-37: identifying that it is a particular month (e.g., December) and that a particular item is sold in the particular month (e.g., tree ornaments). Accordingly, the scoring module may score query tokens or result candidates higher based on this determination. The training module is configured to receive input data points and include initial training phases in order to make predictions or classifications more robust, the learning module may have to analyze several years or iterations of data to make this inference strong over a threshold, as opposed to only analyzing one year. The learning module identifies patterns or associations based on the frequency of user selections indicated by the statistics (and/or other statistics) in a data structure and provides an associated output. Each token of the plurality of tokens is scored higher than other tokens if it includes a higher frequency of a particular user selection compared to other tokens; para. 44-45: storing user selection statistics (e.g., click frequency, purchase frequency, skip frequency, etc.) in a statistics data structure and generating a learning model (e.g., a word embedding vector model) based on the user selection statistics for use in executing queries for one or more resources…each of the user selections represent a history of user selections associated with only a particular user). Thus, it would have been obvious to one or ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Ghoshal, Huh and Bell in order to effectively identifying users query intents in multiple sessions and providing users the most relevant search queries for the users to select as alternative searches if desire.
Claims 4, 11, 18 are rejected under 35 U.S.C. 103 as being unpatentable over Ghoshal et al. (US 20200125575) in view of Huh et al. (US 11222027) and further in view of Bell et al. (US 20190392082) and Min et al. (US 20190122655).
As per claims 4, 11, 18, Ghoshal and Huh do not explicitly teach said claims.
Bell teaches
wherein the token embedding vectors for each activity are generated, a token obtained from the activity data for a corresponding activity undertaken by the end-user (fig. 8: pre-trained word embedding vector model(s) and Re-train the word embedding vector model(s) as the users continue to search/interact with the system. Thus, the word embedding vector models are continuously be trained on users’ activities, historical logs on multiple or all sessions in order to recognize users’ interests, needs, etc. in order to provide relevant search results to the users – See fig. 6: increment a statistics data structure(s) for each of the token(s) based on the user selection(s), set/modify learning model(s) based on the incremental statistics data structure(s); fig. 10: user selection history data store; para. 33: A "word embedding vector model" as described herein maps data (e.g. the identifier selections 104/query 108) to or orients one or more points or vectors in vector space (e.g., real numbers in a 2D or 3D graph model) according to a context similarity to other data points or vectors in vector space. For example, a word embedding vector model in some embodiments includes a Word2Vec model; para. 48: the vector space represents a "retrained" or trained embedding. A retrained or trained word embedding model is an embedding that receives training feedback after it has received initial training session(s) and is optimized or generated for a specific data set (e.g. scoring one or more tokens, etc.)). Thus, it would have been obvious to one or ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Ghoshal, Huh and Bell in order to effectively identifying users query intents in multiple sessions and providing users the most relevant search queries for the users to select as alternative searches if desire.
Ghoshal, Huh and Bell do not explicitly teach by providing to a word embedding encoder.
Min teaches
by providing to a word embedding encoder a token obtained from the activity data for a corresponding activity undertaken by the end-user (para. 49-53: To investigate the modeling capacity of word embeddings, we consider a type of models with no additional compositional parameters to encode natural language sequences, termed as SWEM (simple word embedding-based models). Among them, the simplest strategy is to compute the element-wise average over word vectors of all sequence tokens; figs. 4-5). Thus, it would have been obvious to one or ordinary skill in the art before the effective filing date of the claimed invention to combine the teachings of Ghoshal, Huh, Bell and Min et al. in order to effectively manage inputs and outputs in variable length sequences in identify user query intents in search sessions and providing users relevant search queries to select as alternative searches if desire.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. Kenny et al. (US 20200342348) teaches at para. 49: the server computing device 106 also captures and stores encoded text from prior user interactions, in order to build a set of training data that is then used by the classification model training module 110 to train the plurality of classification models 108a-108c to predict cost and user engagement, as described herein. For example, as users at many different client computing devices (such as device 102) establish communication sessions with server computing device 106 and client computing device 103, the user interactions generated by the client computing device 103 (e.g., service agent and/or chat bot) are captured and vectorized as described above.
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any extension fee pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to LINH BLACK whose telephone number is (571)272-4106. The examiner can normally be reached 9AM-5PM EST M-F.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Tony Mahmoudi can be reached on 571-272-4078. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/LINH BLACK/Examiner, Art Unit 2163 8/26/2023
/TONY MAHMOUDI/Supervisory Patent Examiner, Art Unit 2163