DETAILED ACTION
This final action is in response to the amendment and remarks filed on 04/07/2026 for application 18/069,771.
Claims 1, 4-5, 7, 10, 13-14, 16, and 19 have been amended.
Claims 1-20 remain pending in the application. Claims 1, 10, and 19 are independent claims.
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 .
Claim Rejections - 35 USC § 103
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 (i.e., changing from AIA to pre-AIA ) 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.
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 1-20 are rejected under 35 U.S.C. 103 as being unpatentable over Jiang et al. (“Interpretable contrastive word mover’s embedding”, available via arXiv 1 Nov 2021), hereinafter Jiang, in view of Skianis et al. (“Boosting Tricks for Word Mover’s Distance”, published 2020), hereinafter Skianis, Ji et al. (“Does the Magic of BERT Apply to Medical Code Assignment? A Quantitative Study”, available via arXiv 26 Oct 2021), hereinafter Ji, and Liu et al. (“MetBERT: a generalizable and pre-trained deep learning model for the prediction of metastatic cancer from clinical notes”, available via conference May 2022), hereinafter Liu.
Regarding claim 1, Jiang teaches A computer-implemented method for generating a classification prediction (“This paper shows that a popular approach to the supervised embedding of documents for classification, namely, contrastive Word Mover’s Embedding, can be significantly enhanced by adding interpretability. This interpretability is achieved by incorporating a clustering promoting mechanism into the contrastive loss” [Jiang Abstract]; “Our key idea is to learn a set of anchors C(y) ∈ R d×p for some p and for each class y ∈ [1 : Y ] in the representation space. Anchors offer two advantages. First, they provide for direct and simple NN classification. Second, using anchors we can learn words that have discriminatory power for particular classes, thereby enabling interpretability” [Jiang pages 2-3 Problem formulation and approach]), the computer-implemented method comprising:
for each input reference text data object of one or more input reference text data objects with respect to a set of candidate target text data objects, generating, using a computing entity and a classification machine learning model, a set of maximal word similarity scores, (“Given labeled data (X(m) , w(m) ), m = 1, 2, · · · , M, we seek to learn a representation Z = f(X) such that a Nearest Neighbor (NN)-type classifier in the representation space accurately classifies the document. Using NN in representation space requires a notion of similarity or distances between documents. We use the WMD Kusner et al. (2015), defined as follows” [Jiang page 2 Problem formulation and approach]; “Given this set-up, our approach is to learn the anchors C(y) , y ∈ [1 : Y ] via contrastive learning (see Figure 1) Chen et al. (2020); Khosla et al. (2020) in the representation space. In contrastive learning, one defines triplets (µm, µym c , µym’ c), where µm is the representation of a document m with label ym, m’ /= m, and µym C , µym’ C are representation of anchors of C(ym) and C(ym’) respectively. Assuming a uniform measure on the support of the anchor points, we can use the WMD W(µm, µym c) and W(µm, µym’ c ) as similarity, i.e. contrastive measure. We will contrast each document with all the anchors. Thus, we will have Y − 1 triplets (µm, µym c , µym’ c ) for each document… Given M documents, in order to train the model parameters, C(y) , y ∈ [1 : Y ] and A, we minimize the following triplet loss function Hermans et al. (2017):
PNG
media_image1.png
91
657
media_image1.png
Greyscale
” [Jiang page 3 Proposed approach]; In the disclosed contrastive approach, each document representation µ(m) for m=1…M (i.e., input reference text data object) is compared to a set of anchor points C(y) (i.e., candidate target text data objects, one for each class) utilizing WMD as a contrastive measure, and thereby determining a set of scores
PNG
media_image2.png
43
365
media_image2.png
Greyscale
(i.e., similarity scores) for each document) wherein:
(i) the classification machine learning model has been pre-trained based at least in part on embeddings associated with training data (“We are given M documents each belonging to one of the Y classes. Each document D(m) with label y (m) is represented by two sets, {a m i } n i=1 and {b m i } n i=1, where n is the number of unique words in one document, a m i is the i-th word, b m i is the number of times a m i appears in D(m)…Using pre-trained word embeddings from GLoVe Pennington et al. (2014), D(m) is represented as a tuple (X(m) , w(m) ), where X(m) = [x (m) 1 , · · · , x (m) n ] ∈ R d×n and x (m) i ∈ R d is the embedding for the word a m I” [Jiang page 2 Problem formulation and approach])
(ii) the classification machine learning model has been fine-tuned using labeled training data based at least in part on a maximal word similarity-based contrastive loss function associated with the training data (“Given M documents, in order to train the model parameters, C(y) , y ∈ [1 : Y ] and A, we minimize the following triplet loss function Hermans et al. (2017):
PNG
media_image1.png
91
657
media_image1.png
Greyscale
” [Jiang page 3 Proposed approach]; After obtaining pre-trained word embeddings, model parameters are further fine-tuned based on the disclosed contrastive loss function)
(iii) each maximal word similarity score in the set of maximal word similarity scores comprises a maximal value of a transition cost value associated with one or more reference words of the input reference text data object and one or more target words of a target text data object in the set of candidate target text data objects (“D(m) is represented as a tuple (X(m) , w(m) ), where X(m) = [x (m) 1 , · · · , x (m) n ] ∈ R d×n and x(m)i ∈ Rd is the embedding for the word ami…Given the representations of two documents, (Z(m) , w(m) ), (Z(m0 ) , w(m’ ) ) when seen as empirical measures
PNG
media_image3.png
70
295
media_image3.png
Greyscale
,
PNG
media_image4.png
77
323
media_image4.png
Greyscale
, the WMD between µm and νm’ is defined as Kusner et al. (2015))
PNG
media_image5.png
88
572
media_image5.png
Greyscale
…Here d(z(m)i , z(m’)j) is referred to as the ground cost” [Jiang page 2 Problem formulation and approach]; Contrastive measure WMD is calculated based on distance cost measure
PNG
media_image6.png
30
167
media_image6.png
Greyscale
corresponding to each word in the document µm, wherein minimization of the distance cost measure maximizes valuation of similarity) wherein the transition cost value is determined based at least in part on: (a) a word-wise flow data object for the input reference text data object and the target text data object , wherein the word-wise flow data object comprises a word-wise flow value for each word pair of a plurality of word pairs, wherein the word pair comprises a reference word and a target word (see
PNG
media_image7.png
25
50
media_image7.png
Greyscale
in
PNG
media_image8.png
27
101
media_image8.png
Greyscale
equation [Jiang page 2 Problem formulation and approach]), and (b) a word-wise similarity value for each word pair (see
PNG
media_image9.png
32
113
media_image9.png
Greyscale
in
PNG
media_image8.png
27
101
media_image8.png
Greyscale
equation [Jiang page 2 Problem formulation and approach]);
generating, using the computing entity, a classification output based at least in part on the set of maximal word similarity scores (“NN classification: Once the model is trained, in order to do NN classification on a test document, we first embed it in the representation space using the learned parameters, A via equation (1), compute the WMD distances between the representation and the anchors C(y) , y ∈ [1 : Y ]. The class represented by the anchor with the minimum distance is declared as the label” [Jiang page 2 Proposed approach]); and
initiating, using the computing entity, the performance of one or more prediction-based actions based at least in part on the classification output ([Jiang page 2 Proposed approach] as detailed above; Based on the calculated WMD distances between the representation and the anchors (i.e., classification output), a label is declared)
The classification machine learning model of Jiang relies on a GLoVe architecture to generate embeddings [Jiang page 2 Problem formulation and approach], and is not expressly taught as being pre-trained or fine-tuned using cross-domain training data (i.e., comprising data from both a source domain and a target domain, wherein the model is pre-trained at least in part on embeddings associated with source domain training data and target domain training data and fine-tuned using source domain training data).
In the same field of endeavor, Skianis teaches a framework of applying a word mover’s distance (WMD) metric to address natural language processing (NLP) tasks (“Word embeddings have opened a new path in creating novel approaches for addressing traditional problems in the natural language processing (NLP) domain. However, using word embeddings to compare text documents remains a relatively unexplored topic—with Word Mover’s Distance (WMD) being the prominent tool used so far. In this paper, we present a variety of tools that can further improve the computation of distances between documents based on WMD” [Skianis Abstract]) wherein the classification machine learning model may comprise an architecture (such as ELMo or BERT) capable of generating contextualized embeddings (“Recent work by [24] introduced ELMo, a novel type of deep contextualized word representations. These vectors represent internal states of a deep bidirectional language model (biLM), which is pre-trained on a large text corpus. In their work, the vectors used are derived from a bidirectional LSTM that is trained with a coupled language model (LM) objective. Essentially, for every word there is a vector every time it is found around a context. In our work, we replace Google’s pretrained vectors with ELMo, to test how it can affect measuring distances” [Skianis page 767 Deep Contextualized Word Representations]; “Measuring similarity between two documents that share words, appearing in different context, can make comparison harder. Thus, the problem of polysemy should also be addressed. In order to address that, topical word embeddings [20] can be applied. Thus, a “topical” WMD, based on topics rather than documents alone, would be a promising direction step. Fine-tuning ELMo or BERT [10] for distance computation can be a future direction” [Skianis page 770 Conclusion and Future Work])
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have substituted the GLoVe architecture of Jiang for a contextualized language model (e.g. BERT), as suggested by Skianis, because both Jiang and Skianis are directed towards applying WMD frameworks to NLP tasks. Given that contextualized language models, such as BERT, are well-known in the field of natural language processing, a person of ordinary skill in the art would recognize the value of making a substitution expressly suggested by Skianis for “boosting” WMD classification performance (“We see that in three datasets, using ELMo as vectors reduced the knn classification error dramatically, with its expressive contextualized power” [Skianis page 770 Experiments]), thereby allowing the “topical” WMD framework (similarly mentioned in [Skianis page 770 Conclusion and Future Work], as detailed above) of Jiang to further benefit from the additional information provided by contextualized embeddings (rather than GLoVe embeddings) prior to fine-tuning (“This is expected, as ELMo is proved to boost many diverse NLP tasks, even with a simple averaging without any fine-tuning. To the best of our knowledge, our work is the first to incorporate deep contextualized word representations as an input for distance computation” [Skianis page 767 Deep Contextualized Word Representations]).
In the same field of endeavor, Ji teaches a framework of applying pre-trained, fine-tuned NLP models to document classification tasks (“This paper conducts a comprehensive quantitative analysis of various contextualized language models’ performances, pretrained in different domains, for medical code assignment from clinical notes. We propose a hierarchical fine-tuning architecture to capture interactions between distant words and adopt label-wise attention to exploit label information” [Ji Abstract]) wherein the cross-domain classification machine learning model may comprise a BERT architecture that has been pre-trained (“RQ1: What kind of BERT pretraining works best? We adopt domain-specific corpora and BERT variants pretrained with different domain adaptation illustrated in Fig. 1, and compare their performance on the MIMIC-III benchmark” [Ji page 2 Introduction]) based at least in part on embeddings associated with source domain training data and target domain training data (“We study three types of domains: 1) general domains such as book corpora and general Wikipedia articles; 2) domains that are closely related to the target clinical domain; 3) the target clinical domain. Assigning ICD codes from clinical notes is a task in the clinical domain. We consider biomedical and health-related social domains as candidate domains closely related to the clinical domain. Inspired by domain- and task-adaptive pretraining (Gururangan et al., 2020), we investigate different ways of pretraining models for medical code assignment: 1) pretrain only on general domains and immediately transfer to the target clinical domain; 2) continue pretraining on close domains and clinical domains such as the biomedical domain, and transfer to the target clinical domain 3) pretrain on close domains from scratch and transfer to the target clinical domain; 4) pretrain on mixed domains and further fine-tune on the target domain” [Ji page 3 Pretraining Domains])
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have incorporated a BERT model pre-trained based at least in part on embeddings associated with source domain training data and target domain training data as taught by Ji into the combination of Jiang and Skianis because both Jiang and Ji are directed towards applying pre-trained, fine-tuned NLP models to document classification tasks, and both Skianis and Ji are further directed towards applications of contextualized language models (such as BERT). Given that Ji attempts to discover a means of successfully adapting pre-trained BERT architectures to the problem of medical code assignment by utilizing various fine-tuning strategies (“RQ2: What kind of BERT fine-tuning formulation works best for long notes? We employ classical finetuning, develop a hierarchical architecture for long clinical notes, and consider label-aware feature representation” [Ji page 2 Introduction]), a person of ordinary skill in the art would thereby recognize the BERT-boosted WMD framework of Jiang and Skianis as implementing an additional un-tested BERT fine-tuning technique with potential applicability to the problem disclosed in Ji.
However, the combination does not expressly teach fine-tun[ing] using source domain training data.
In the same field of endeavor, Liu teaches a framework of applying pre-trained, fine-tuned NLP models to document classification tasks (“Here, we fine-tuned multiple state-of-the-art BERT-based models using discharge summaries from the open MIMIC-III dataset and derived MetBERT, a novel model tailored to predict cancer metastasis from clinical notes” [Liu Abstract]; “Pre-trained weights of all the fine-tuned BERT models were provided by the HuggingFace library 12. For each model, we added a dropout layer for regularization and a fully connected classification layer (which outputs scores) for fine-tuning. The softmax and argmax functions were used for mapping scores to probabilities and class assignment, respectively” [Liu page 333 Methods]) that fine-tune[s] the model using source domain training data, particularly for application to unseen target tasks (“We started by utilizing the publicly available MIMIC-III dataset to fine-tune five different BERT-based models and then evaluated the performance of the best fine-tuned-model (which we call MetBERT) in an independent dataset prepared from the Epic system at Spectrum Health” [Liu page 332 Introduction]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have incorporated the cross-domain classification machine learning model being fine-tuned using source domain training data as taught by Liu into the combination because both Jiang and Liu are directed towards applying pre-trained, fine-tuned NLP models to document classification tasks, and both Ji and Liu are further directed towards applications of pre-trained, fine-tuned BERT models to medical code assignment. Given that Ji already suggests pre-training BERT model types using a mix of available domains ([Ji page 3 Pretraining Domains] as detailed above), incorporating the teachings of Liu would further expand the overall WMD framework to adapt for scenarios with no labelled target domain data available for fine-tuning (“However, training a language model using deep neural networks from scratch requires a huge amount of labeled data and computing resources, which are usually lacking in small institutes” [Liu page 331 Introduction]; “Although MetBERT was fine-tuned using MIMIC III dataset, the high performance (AUC = 0.94) on Spectrum Health dataset suggested that it is generalizable to other datasets, and such feature enables a potential usage of MetBERT for early diagnosis of metastatic cancer. Training a model using open datasets and deploying to an internal system is very appealing in many small institutes including ours where the initial labeled data is often scarce” [Liu page 337 Conclusion]).
Regarding claim 2, the combination of Jiang, Skianis, Ji, and Liu teaches the limitations of parent claim 1, and Jiang further teaches wherein each input reference text data object comprises text data originating from a target domain data source (“For all datasets, the ground cost in WMD is set to squared Euclidean and we use the Sinkhorn algorithm for computing it Peyré and Cuturi (2018). The various hyper-parameters are set using cross-validation….Public Datasets: 2 Information about the public datasets is shown in Table 1. In Table 2 we show the results from WMD Kusner et al. (2015) , supervised-WMD (S-WMD) Huang et al. (2016) and our method” [Jiang page 4 Evaluation]; see Table 1: Public dataset characteristics including different datasets (e.g., BBCSPORT, TWITTER, RECIPE, etc.) each comprising #Test data (i.e., input reference text data objects) with avg word count (i.e. having text data) [Jiang page 4]).
Regarding claim 3, the combination of Jiang, Skianis, Ji, and Liu teaches the limitations of parent claim 1, and Jiang further teaches generating, using the computing entity, a ranked similarity list for each input reference text data object based at least in part on the set of maximal word similarity scores (“Figure 2 shows how our model leads to interpretability. Under the contrastive loss, the difference between WMD W(µm, µym c ) and W(µm, µ ym0 c ) will be maximized. This forces the important words for a given class to be close to the anchor of its corresponding class in the representation space and further from the anchors of the other classes…Then we define the importance value of zt for class y as
PNG
media_image10.png
38
690
media_image10.png
Greyscale
. Larger I(zt, y) means that word at is important for class y” [Jiang page 4 Interpretation]; The disclosed framework allows for the individual words of a document (i.e., reference text data object) to be ranked (via importance value) on their importance to classification prediction based on the WMD contrastive loss)
Regarding claim 4, the combination of Jiang, Skianis, Ji, and Liu teaches the limitations of parent claim 1, and Skianis further teaches wherein the cross-domain classification machine learning model comprises at least one of (i) a Bidirectional Encoder Representation from Transformers (BERT) layer or (ii) a Long Short-Term Memory (LSTM) layer ([Skianis page 767 Deep Contextualized Word Representations] as detailed in claim 1 above).
Regarding claim 5, the combination of Jiang, Skianis, Ji, and Liu teaches the limitations of parent claim 1, and Jiang further teaches wherein the maximal word similarity-based contrastive loss function is configured to: (i) maximize the maximal word similarity score for positive training text data object pairs comprising a training reference text data object and an assigned target text data object and (ii) minimize the maximal word similarity score for negative training text data object pairs comprising the training reference text data object and the unassigned target text data object (“Assuming a uniform measure on the support of the anchor points, we can use the WMD W(µm, µym c ) and W(µm, µ ym’ c ) as similarity, i.e. contrastive measure. We will contrast each document with all the anchors. Thus, we will have Y − 1 triplets (µm, µym c , µ ym’ c ) for each document…Given M documents, in order to train the model parameters, C(y) , y ∈ [1 : Y ] and A, we minimize the following triplet loss function Hermans et al. (2017):
PNG
media_image11.png
100
754
media_image11.png
Greyscale
” [Jiang page 3 Proposed approach]; Given WMD is a contrastive measure, the disclosed loss function minimizes distance
PNG
media_image12.png
57
148
media_image12.png
Greyscale
(i.e., maximizes similarity) for positive pairs and maximizes distance
PNG
media_image13.png
48
193
media_image13.png
Greyscale
(i.e., minimizes similarity) for negative pairs)
Regarding claim 6, the combination of Jiang, Skianis, Ji, and Liu teaches the limitations of parent claim 1, and Liu further teaches wherein the source domain training data comprises labeled training data and unlabeled training data (“However, training a language model using deep neural networks from scratch requires a huge amount of labeled data and computing resources, which are usually lacking in small institutes. Utilizing pre-trained model weights followed by fine-tuning for downstream domain-specific tasks has therefore become a popular approach as it often results in better performance with lower training overhead on smaller datasets…With the success of BERT based models in general domain, clinical researchers have developed models such as BioBERT (pre-trained using data from PubMed abstracts and PMC full-text articles), clinicalBERT (pre-trained with MIMIC-III data), BlueBERT(pre-trained with the combination of MIMIC-III and PubMed data), and PubmedBERT (pre-trained using domain-specific pretraining from scratch on PubMed data)” [Liu pages 331-332 Introduction]; “Our training data was compiled from MIMIC-III, a publicly available (and de-identified) EHR dataset consisting of clinical information of over 40,000 patients. All the 1,610 expert-annotated discharge summaries used for model fine-tuning came from previous work…We fine-tuned five types of BERT-based models (BERT, BlueBERT, BioBERT, ClinicalBERT, and PubmedBERT) and assessed their performances by measuring precision, recall, and F-1 scores of positive samples in the training dataset” [Liu page 333 Methods and Results])
Regarding claim 7, the combination of Jiang, Skianis, Ji, and Liu teaches the limitations of parent claim 1, and Ji further teaches assigning a target text data object to each input reference text data object in response to association with a threshold-satisfying maximal word similarity score (“In the first approach, we truncate clinical notes to 512 tokens, take the final hidden state of the first token [CLS] as the pooled representation of the truncated note (denoted as C ∈ Rdh ), and apply a fully connected network (FCN) as the classifier with sigmoid activation to predict output probabilities…The learning objective function adopts the binary cross entropy loss denoted as:
PNG
media_image14.png
104
578
media_image14.png
Greyscale
where yi ∈ {0,1} is the ground-truth label, yˆi is the sigmoid score for prediction, and m is the number of ICD codes” [ “We evaluate the metrics of precision at k, where k = 5 for MIMIC-III subset with top-50 frequent codes and k = 8,15 for full sets of MIMIC-III, given the observation that most medical documents are assigned no more than 20 code” [Ji page 6 Results]; The disclosed multi-label classification (i.e., assigning multiple target objects (e.g., codes) to a single reference object (e.g., document)) approach utilizes sigmoid activation to determine output probabilities for each ICD code (i.e., similarity scores indicating likelihood of code assignment to the given document), and thereby assigns predicted codes in accordance with a threshold-satisfying sigmoid score)
Regarding claim 8, the combination of Jiang, Skianis, Ji, and Liu teaches the limitations of parent claim 1, and Jiang further teaches wherein the transition cost value is maximized in accordance with a maximization constraint requiring that a sum of each word-wise flow value for a particular target word be equal to a document-wide word weight value for the particular target word in the target text data object (“the WMD between µm and νm’ is defined as Kusner et al. (2015))
PNG
media_image5.png
88
572
media_image5.png
Greyscale
such that
PNG
media_image15.png
78
387
media_image15.png
Greyscale
and
PNG
media_image16.png
70
259
media_image16.png
Greyscale
” [Jiang page 2 Problem formulation and approach])
Regarding claim 9, the combination of Jiang, Skianis, Ji, and Liu teaches the limitations of parent claim 8, and Jiang further teaches wherein the document-wide word weight value is determined based at least in part on: (i) a term frequency value of the particular target word in the target text data object, or (ii) a sum of each term frequency value for the one or more target words in the set of candidate target text data objects ([Jiang page 2 Problem formulation and approach] as detailed above)
Regarding claims 10-18, they are system/apparatus claims that largely correspond to the method of claims 1-9, which are already taught by the combination of Jiang, Skianis, Ji, and Liu as detailed above. Jiang further teaches An apparatus for generating a classification prediction, the apparatus comprising at least one processor and at least one memory including program code, the at least one memory and the program code configured to, with the processor, cause the apparatus to at least: perform the claimed functions (see [Jiang page 4 Evaluation] and [Jiang page 7 Codes and Implementation]; Implementing the disclosed approach on public datasets inherently requires a computer with adequate processing and memory capabilities for executing the associated code). Consequently, claims 10-15 and 17-18 are rejected for the same reasons as claims 1-6 and 8-9.
Regarding claims 19-20, they are product claims that largely correspond to the method of claims 1-2, which are already taught by the combination of Jiang, Skianis, Ji, and Liu as detailed above. Jiang further teaches A computer program product for generating a classification prediction, the computer program product comprising at least one non-transitory computer-readable storage medium having computer- readable program code portions stored therein, the computer-readable program code portions configured to: perform the claimed functions (see [Jiang page 4 Evaluation] and [Jiang page 7 Codes and Implementation]; Implementing the disclosed approach on public datasets inherently requires a computer with adequate processing and memory capabilities for executing the associated code). Consequently, claims 19-20 are rejected for the same reasons as claims 1-2.
Response to Amendment and Arguments
The amendment filed 04/07/2026 has been entered.
Applicant’s amendment to the disclosure with respect to resolving drawing objections has been considered, and the objections are consequently withdrawn.
Applicant’s amendment to the claims with respect to resolving claim objections and indefiniteness rejections under 35 U.S.C. 112(b) has been considered, and the objections and rejections are consequently withdrawn.
The remarks filed 04/07/2026 have been fully considered.
Applicant’s remarks [Remarks pages 9-11] traversing the non-eligible subject matter rejections under 35 U.S.C. 101 set forth in the office action mailed 01/07/2026, have been considered and are persuasive in part.
Upon further consideration, the examiner at least agrees that the recited procedure of “(a) cross-domain pre-training on embeddings from both source and target domain training data; (b) fine-tuning using a "maximal word similarity-based contrastive loss function"-a specific, non-generic loss function tied to the Word Mover's Similarity framework and (c) the specific computation of maximal word similarity scores via transition cost values, word-wise flow data objects, and word-wise similarity values” adequately reflects improvements detailed in the specification with respect to operation of a machine learning model. While the described procedure may invoke limitations that on their own are interpretable through the lens of well-known, routine, and conventional concepts in the field of natural language processing (transfer learning, pre-training/fine-tuning models, word mover’s distance (WMD)), the overall recited procedure reflects an unconventional ordered combination of features (particularly the “maximal word similarity-based contrastive loss function”) that is thereby interpretable as an improvement over conventional technology.
Consequently, the rejections are withdrawn.
Applicant’s remarks traversing the obviousness rejections under 35 U.S.C. 103 set forth in the office action mailed 01/07/2026 have been considered but are not persuasive.
Applicant alleges that the cited combination of references (Jiang, Skianis, Ji, and Liu) fail to teach or suggest all elements of the independent claims.
The examiner respectfully disagrees. Applicant is directed towards the grounds of rejection under 35 U.S.C. 103 with respect to amended claims 1-20 set forth above. Applicant’s arguments are further summarized and addressed below.
Applicant argues [Remarks pages 11-12] that the claims require maximization while Jiang teaches minimization – particularly, the claims require maximization of a transition cost value (a similarity measure), while Jiang teaches minimization of transport cost (a distance measure).
In response, it is the examiner’s view that minimization of a distance measure, as Jiang is interpreted to teach by the applicant and which the examiner does not necessarily refute, is indeed broadly interpretable as an equivalent optimization objective to maximization of a similarity measure as it is detailed in the claims. The claims merely recite “a maximal value of a transition cost value”, and a person of ordinary skill in the art would reasonably interpret minimization of a distance measure as directly resulting in a maximization of a corresponding similarity measure (i.e., transition cost), particularly given that Jiang particularly describes the WMD metric as an implicit measure of similarity (see [Jiang page 3 Proposed approach], as detailed above), wherein minimizing distance naturally maximizes similarity, and vice versa. Applicant’s characterization of the described maximization as more than a semantic distinction and rather a “fundamentally different technical approach” is not reasonable, because the claim does not set forth any actual limitations that would distinguish the recited concept of “similarity” based functions, scores, values etc. from merely being the conceptual counterpart of distance based functions, scores, and values. The claims do not set forth any particular procedure for determining similarity measures that would then invalidate the currently set forth interpretation.
Applicant argues [Remarks pages 12-13] that the cited combination fails to teach a cross-domain contrastive loss function, because Jiang operates entirely within a single domain, while Liu’s approach is standard transfer learning and thereby fundamentally different.
In response, the examiner notes that applicant’s argument appears to disregard the additional references (particularly, Ji) that were cited in the combination to teach the claim. Ji expressly teaches utilization of cross-domain training data (i.e., comprising data from both a source domain and a target domain) for training, and would thereby upon incorporation into the combination result in a “cross-domain” loss function as described in the claims. It is unreasonable to suggest that mere broad recitation of the claimed loss function as being “associated with source domain training data and target domain training data” would require any modification to the loss function itself, or anything more than what was taught via incorporation of Ji into the cited combination.
Applicant argues [Remarks page 13] that the four-reference combination reflects impermissible hindsight because the motivation to combine Ji is speculative, and that the examiner therefore relied on knowledge gleaned from applicant’s disclosure.
In response, the examiner notes that the applicant’s argument appears to disregard the explicit evidence that was cited to explain why a person of ordinary skill in the art would recognize the described applicability of prior references (as detailed in rejection above – Ji already attempts to discover a means of successfully adapting pre-trained BERT architectures to the problem of medical code assignment by utilizing various fine-tuning strategies (“RQ2: What kind of BERT fine-tuning formulation works best for long notes? We employ classical finetuning, develop a hierarchical architecture for long clinical notes, and consider label-aware feature representation” [Ji page 2 Introduction]) in order to characterize the motivation as speculative. Additionally, applicant does not cite to any evidence or explanation, beyond mere conclusory statements, to explain their claim that the examiner relied on knowledge gleaned from applicant’s disclosure (rather than knowledge obtained from the references themselves and/or knowledge that would be apparent to one of ordinary skill in the art).
Applicant has not presented further arguments with respect to the dependent claims. As such, amended claims 1-20 stand rejected under 35 U.S.C. 103.
Conclusion
THIS ACTION IS MADE FINAL. 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 nonprovisional extension fee (37 CFR 1.17(a)) 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 mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to VIJAY M BALAKRISHNAN whose telephone number is (571) 272-0455. The examiner can normally be reached 10am-5pm EST Mon-Thurs.
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, JENNIFER WELCH can be reached on (571) 272-7212. 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.
/V.M.B./
Examiner, Art Unit 2143
/JENNIFER N WELCH/Supervisory Patent Examiner, Art Unit 2143
Prosecution Insights