DETAILED ACTION
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 .
Status of Claim
The office action is being examined in response to the application filed by the applicant on 11 February 2026.
Claim 1-20 are pending and have been examined.
This action is made FINAL.
Response to Arguments
35 U.S.C. § 101
Applicant's arguments filed 11 February 2026, on pages 10-13, with regards to 35 U.S.C. § 101 have been fully considered but they are not persuasive.
The applicants’ arguments, on pages 11-12, asserting that the independent claims improve functioning, training, and retraining of a model by improving word-for word legal clause extraction, is not probative. The claims recite a step by step process of using general purpose models and general purpose computing structures, i.e. the additional elements, to perform functional limitations. The claims do not recite limits to how the limitations are performed, but for the additional elements. The arguments assert improvements that are rooted in the characterized data, using off the shelf classification models or LLM’s while implementing off the shelf LoRA techniques to perform the efficient fine-tuning and matrix evaluations according to metric thresholding.
The instant specification supports this finding in at least [0032-0041], “the system 100 may use (pre-existing) techniques described herein to efficiently train an accurate machine learning model for legal clause extraction … machine learning model may be an example of an off-the-shelf pre-trained sequence-to-sequence LLM that the system 100 fine-tunes on legal contracts. Further, the specification discloses in [0060], “an open-source LLM may be initially trained on a relatively large corpus of data publicly available on the internet;” and in [0064] “perform a parameter-efficient fine-tuning … based on LoRA.” [0065] discloses that the LoRA model inherently performs the encoder/decoder matrices pair tuning to a particular accuracy threshold. The LoRA technique is inherently utilized to improve efficiency and accuracy of machine learning models to improve finding a particularly characterized data.
Merely applying existing machine learning processes to existing machine learning models, using particularly characterized data to direct the model to return particularly characterized data faster than another off the shelf model that does not use the LoRA encoder/decoder matrices and accuracy thresholding methods, does not effectuate an improvement to the functioning of the machine learning model in a manner that is indicative of a practical application under 35 U.S.C. § 101. This is effectively generally linking the use of the abstract ideas to the particular model environment using off the shelf modelling techniques, i.e. the claim is a drafting effort to monopolize the exception and the off the shelf modelling techniques applied in open source LLM’s.
On page 12, the applicants’ arguments, asserting that the general-purpose nature of the computing structures was made without adequate explanation, is not persuasive. The claims do not recite specific user devices, non-transitory computer readable mediums, processors, memories, or apparatus. Under the broadest reasonable interpretation, these could be any appropriate commuting structures, of which the specification does not refute. The statement mischaracterizes the Panel, where the specification clearly discloses, as discussed above, that the instant invention implements at least off the shelf open source LLM’s and off the shelf LoRA and other (chunking) techniques. These existing models are not patentable. Instead, the applicant must be able to elucidate how the claims in light of the specification, when abstract ideas are found, are comprised of additional elements that fulfill a category under MPEP 2106.05(a)-(h).
The applicants’ arguments on page 12, asserting that claim 1 is non-conventional and not routine, are not probative, nor germane to the discussions of patentability under 35 U.S.C. § 101 for the presently presented or previously presented claim language. The Examiner did not identify any claim language as insignificant extra-solution activity, nor identify any limitations as well understood, routine, or conventional claim language. Since no limitations or claim language in the claims were identified in either category by the Office as part of the 35 U.S.C. § 101 rejection, the applicants’ arguments asserting that claim 1 is non-conventional and not routine, are not probative nor germane to overcome the 35 U.S.C. § 101 rejection citing unrelated issues, rules, analyses, and conclusions, i.e. the reasons for the rejection are wholly unrelated to these page 12 arguments.
Please find the adjusted 35 U.S.C. § 101 rejection below to reflect the amendments.
The 35 U.S.C. § 101 rejection is Maintained.
35 U.S.C. § 103
The applicants’ arguments, on pages 14-16, asserting that Bonfante, Han, Higgins, and Chen do not teach or suggest all of the features of the amended independent claims 1, 19, and 20, either alone or in any combination, are not persuasive. First, the obviousness rejection of claims 1, 19, and 20 rely upon obviousness over Bonfante in view of Han, in further view of Chen. Since Higgins is only implemented in connection with Bonfante, Han, and Chen, in the rejection for claims 7 and 8, refuting Higgins teaching or suggesting any features of the independent claims, alone or in combination is not probative in relation to the specific rejection being made against the independent claims. Further, the applicants’ assertions regarding each prior art disclosure, rely on restrictive interpretations of the disclosures achieved through selective omissions that do not account for the full scope of the rejection presented by the Examiner.
The applicants’ arguments on, pages 14-15, fail to address the full scope of the 35 U.S.C. § 103 obviousness rejection presented for claims 1, 19, and 20 over the prior art of Bonfante in view of Han. The applicant’s assertions that Bonfante does not teach or suggest the claim language, including the applicants' arguments regarding Bonfantes’ alleged silence regarding any claimed evaluation metric and failure to teach or suggest the claimed model updating are not persuasive. The applicants’ arguments, based on Bonfantes’ [abstract], [0027], and [0040], fail to address the remainder of the ten paragraphs or parentheticals asserted by the Examiner, as obvious over Bonfante in view of and Han. Bonfante explicitly discloses both training and retraining, i.e. updating, a machine learning model in at least [0027] and [0029]. The present rejection is also in further view of Chen.
The applicants’ arguments against Bonfante, fail at two points of issue. First, the Applicant argues a narrative that is based on only the three out of many of the articulated disclosure ¶’s, chosen to assert a limited argued narrative without paying respect to the full set of disclosed rejection ¶’s and parentheticals.
Bonfante discloses “iteratively updating the machine learning model, based at least in part on an on an evaluation metric corresponding to one to one mappings, an accuracy, using at least the common substring between the first legal clause and the second legal clause of the mappings,” as presented in the updated rejection below. This presentation is comprised of a new presentation of claim limitations integrated with the previous claim presentation, effectively altering the meaning of the claims. However, there are still no limits placed on the evaluation metric. That means that the claims to not place limits on what an evaluation metric may be, how the iterative updating of weight matrices may occur, how they are based on the metric, how the accuracy is determined, or how the evaluation metric may fail to or may satisfy an accuracy threshold. The claims also place no limits on how the calculating using longest common substrings occurs. In fact, since automatically updating the pairs of weight matrices is based at least in part on both the loss function and an evaluation metric, and the longest common substring is related to a frequency of one or more words, it would be reasonable to assert that these could be based on the same set of tokens of the one or more words, and the related token weights, which may also be based on frequency. The aforementioned section of the new claim limitations are disclosed by Bonfante in the full scope of cited paragraphs and parentheticals, beyond the scope of the three paragraphs asserted by the applicants’ arguments.
The longest common substring is an extension of the one to one mappings, according to the instant specification [0052]-[0054], where matching may at least be calculated using vectors and vector distances, i.e. the instant invention “may perform clause alignment to dynamically determine one-to-one mappings between the set of ground truths 315 and the set of legal clauses … [the instant invention] may use a … process (e.g., …alignment techniques) to determine the one-to-one mappings.” Further, since vectors and vector calculations may be implemented in clause mapping and as part of identifying the text and phrase matching, word for word, the one to one mappings are lexically synonymous to identifying common substrings, as disclosed. That is, Bonfante describes using common substrings, i.e. historical legal obligations, to identify common substrings, e.g. legal obligations, based on the loss function, that assigns token weights to word tokens, which are also word vectors. Said word vectors are calculated, i.e. they are a metric that is evaluated based on calculation of common substrings of word for word, e.g. one-to-one mappings, of words or phrases. Therefore, Bonfante is not silent to an evaluation metric. The instant specification, therefore, discloses and the claims recite a direct correlation between the vectors of word to word text identification and the loss function. That is, since the loss function may be based on word frequency, and is used to update the weight matrices, one could make the correlation that the iterative updating of the weight matrices based on an evaluation metric appears grammatically interchangeable with said loss function, which is in fact also an evaluation metric, without described differentiation. This is especially significant when the specification fails to differentiate the iterative updating of weight matrices pairs when evaluated by either said loss function, based on frequency of words in the document, and said evaluation metric corresponding to one to one mappings of, both disclosed in the instant specification as the matching of legal clause phrases from the training data and legal clause phrases in the evaluative data, i.e. they are both generally based on frequency of words in the document, e.g. Term Frequency-Inverse Document Frequency (TF-IDF), where word or phrase importance falls as the frequency increases.
That is, the rank of the words or phrases decreases as the frequency increases, such that a ranking is merely an evaluation metric. Bonfante discloses a plurality of evaluation metrics that may be implemented in updating the machine learning model. In at least [0032] and [0033], the prior art discloses an accuracy metric and a metric of how common or uncommon each [legal] obligation is, i.e. an evaluation metric of term frequency as in TF-IDF, that is used at least in part, for retraining/updating the machine learning model. Said accuracy metric is disclosed as corresponding to the accuracy of the identified legal obligations, i.e. mappings of known legal clauses to identified legal clauses between historic and new documents. Therefore, the prior art of Bonfante does disclose the argued claim limitations contrary to the applicants’ assertions. In fact, the newly articulated limitations and the limitations that were rolled into the sixth clause of claims 1, 19, and 20, under the broadest reasonable interpretation, place no limits on the evaluation metric. Therefore, Bonfante’s accuracy and ranking metrics are implemented in combination with Han and Chen to show the full scope of the argued limitations.
The amended clause 6, recited in claims 1, 19, and 20, recites a wherein clause to further define the evaluation metric. Said evaluation metric “is calculated,” i.e. calculating the evaluation metric is not a positively recited feature. The wherein clause, wherein the evaluation metric is calculated using at least a longest common substring between the first legal clause and the second legal clause of the mapping, merely further defines the intended result such that calculating an evaluation metric is not positively recited in the claim. Instead, the wherein clause merely defines the origin of the evaluation metric, that it is calculated. Regardless, the Examiner has shown the prior art citations that disclose a calculated evaluation metric.
This same pattern is repeated on page 15 with regards to Han, and on page 16 with regards to Chen. On page 15, the obviousness rejection in view of Han is also refuted over only two amended claim limitations that are connected in a single claim clause, the same limitations the applicant asserted for Bonfante. On page 16, the applicants’ arguments that Higgins and Chen do not overcome the deficiencies of Bonfante and Han, which were not presented nor even suggested by the Examiner, are Moot. The original presentation of claims 1, 19, and 20, were only rejected as obvious over Bonfante in view of Han, such that neither Chen nor Higgins were addressed in terms of the independent claims in the first action on the merits. The further assertion, on page 16, that amended independent claim 1, 19, and 20 are allowable over Bonfante, Han, Higgins, and Chen is Moot, again, because Chen and Higgins were not applied as prior art to the previous independent claim presentation.
The applicant asserts, on page 15, that Han does not teach nor suggest said new claim limitations, are not persuasive. Again, the applicants’ assertions cherry pick only three truncated citations from three different prior art ¶’s from Han, one of which was not even presented in the rejection of the previous claim set, [0020], [0046], and [0047], along with a cherry picked interpretation based summary of said prior art ¶’s. The applicants’ assertions that Han does not overcome the deficiencies of Bonfante, either independently, or in combination, to teach or suggest all of the features, are driven by the applicants’ truncated parentheticals strategically combined in an attempt to appear to teach away from the instant amended claims, without clear evidence, all while avoiding the actual cited ¶’s, citations, and parentheticals. The applicant also avoids directly addressing the Examiner’s rejection through application of the prior art to the appropriate previously presented claims and claim limitations, making arguments based on a skewed or isolated reading of prior art as applied only to particular new claim limits. For instance, the applicant asserts a summarized interpretation that defines Han as describing a disclosure where an expert provides arrays of legal terms that occur in legal clauses, and the expert performs “word by word search in the array for tokens of a sentence in a legal document, and [trains] a ML model based on relevant sentences.” The applicant emphasizes a rendition of steps that are only performed by a human expert, discounting the full scope of the prior art of Han. These arguments are not probative and are also not germane to the actual rejections made.
Since the applicants’ arguments fail to address the full scope of the rejection, the arguments are not persuasive for two particular reasons cited as form paragraphs:
Applicant's arguments fail to comply with 37 CFR 1.111(b) because they amount to a general allegation that the claims define a patentable invention without specifically pointing out how the language of the claims patentably distinguishes them from the references.
Applicant's arguments do not comply with 37 CFR 1.111(c) because they do not clearly point out the patentable novelty which he or she thinks the claims present in view of the state of the art disclosed by the references cited or the objections made. Further, they do not show how the amendments avoid such references or objections.
The applicant does not ascertain the differences between the prior art and the claims at issue, resolved at the level of ordinary skill in the pertinent art. Further the applicant does not present objective evidence to support the assertions that Bonfante or Han do not disclose or teach particular claim limitations as asserted in the arguments at hand.
Higgins is only disclosed as obvious over claims 7 and 8. Therefore, the applicants’ argument, on page 16, that Higgins does not teach or suggest, alone or in combination, claims 1, 19, and 20, is Moot because Higgins was not historically disclosed, and is not presently disclosed as being obvious over claims 1, 19, and 20.
Therefore, the new 35 U.S.C. § 103 rejection is presented below, reflecting at least the argued claim amendments: an evaluation metric that is calculated using at least a longest common substring…, iteratively updating pairs of weight matrices, an evaluation metric corresponding to one ton one mappings…failing to satisfy an accuracy threshold, and wherein the evaluation metric is calculated using at least a longest common substring between the first legal clause and the second legal clause of the mapping.
Applicant's arguments filed 11 February 2026, on page 16, with regards to 35 U.S.C. § 103 for the dependent claims, have been fully considered but they are persuasive. Therefore, the rejection has been withdrawn. However, upon further consideration, a new ground(s) of rejection is made in view of the claim amendments.
For the same reasons as disclosed for the independent claim 1 as applied to dependent claims 2-18, the claims are not allowable over the prior art as disclosed below in the full rejection. The arguments do not present additional arguments that address the individual claims nor the prior art ¶’s and parentheticals as applied in the rejection, therefore, there are no additional arguments to discuss. The new rejections are presented below for the dependent claims 2-18.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claim 1-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Independent Claim:
Regarding Claims 1, 19, and 20: The claims recite: receiving a document, receiving legal clause indications, updating a loss function by assigning weights based on word frequency, updating a first and/or second pair of weight matrices based on the loss function, inputting and outputting characterized data, determining one to one mappings between first and second clauses based on a vector embeddings procedure, and update pairs of weight matrices iteratively based on an evaluation metric repeating upon threshold failure, These limitations recite an abstract idea in the category of “mental processes” including observations, evaluations, judgments, and opinions. The claims do not recite limits to how the models perform the functions. Under the broadest reasonable interpretation of a person having ordinary skill in the art, there is nothing in the claims that preclude the steps from being performed in the human mind because the limitations describe the steps a statistician would have historically performed to analyze data using statistical methods iteratively to parse and compare data until a threshold is met. Further this is a mental process a computer scientist has performed for decades, training a model using particular statistical methods of fine-tuning to create a particular outcome model that fits an overall need of the model such that it may be implemented on new data. The only clarifying difference would be the particularly characterized data input and expected characterized data outputs implemented in training and fine-tuning the model to reach the expected characterized outcome by the human performing a historically similar task.
These same claim limitations also recite abstract ideas in the category of "Certain Methods of Organizing Human Activity" in the subcategory of managing personal behavior. As discussed, the claims recite a mental process that a statistician should follow when training a model with characterized data to perform classification tasks using off the shelf machine learning models.
The claims recite: update a loss function by assigning weights to tokens based on word frequency, update first, second, or both pairs of weight matrices based on the loss function, iteratively update one or both of the pairs of weight matrices based on evaluation metric failing to satisfy a threshold, and the evaluation metric is calculated using at least a longest common substring between clauses of the mapping. These claim functions recite abstract ideas in the category of mathematical concepts in the sub-categories of mathematical relationships, mathematical equations, and mathematical calculations. The claims recite relationships between functions, matrices, weights, metrics, and thresholds, of which are implemented in equations written in prose, and calculated also written in prose.
The claims recite an abstract idea.
Step 2A Prong 2: The claims recite additional elements including a machine learning model and first and second user devices. The model and the computing structures are recited in the claim, with a high level of generality, , e.g. the claims recite general-purpose models and computing structures. The claims do not place any discernable limits on how the steps are performed but for generally linking the receiving data functions to the general-purpose computing structures, or generally linking automatically and iteratively updating pairs of weight matrices to the machine learning model. The claims place no discernable limits on how the updating a loss function by assigning token weights, or determining one to one mappings based on a vector embedding procedure steps are performed. These steps, wholly, recite abstract ideas. Claims 19 and 20 implement these abstract ideas using general-purpose computing structures. Claim 1 differs, however, because these limitations are merely steps, without reciting the structures that be used to implement the steps.
The matrices manipulations, and one to one mappings and associated vector embedding procedures are not tied to a specific computational method, i.e. the model, or technological improvement. Instead, they reflect a high-level abstraction of manipulating mathematical objects conceptually. Moreover, the claims do not recite any details of how these functions are implemented in practice, thus, the claim merely recites automation of matrix manipulations for fine tuning data, and automating a vector embedding to determine one to one mappings of characterized data. The training of a model using historically implemented fine tuning and data mapping procedures, historically performed by statisticians or computer scientists, using an existing model or computer as tools, are ineligible without delivering concrete technical improvements.
There is nothing to preclude any of these limitations from being performed as a mental process, or with the aid of a pen and paper, general-purpose computing structures, or general-purpose models, but for the use of said generally linked and recited additional elements, i.e. the computing structures or machine learning model, discussed further below. That is, the machine learning model is recited without specifying how the model functions to achieve the results based claim limitations. Said model is recited without specifying the particular model implemented. According to the claim language under the broadest reasonable interpretation, the claims may implement any machine learning model or general-purpose computing structures, off the shelf, to perform the associated steps. These recitations amount to “apply it,” mere instructions to apply the abstract idea using generic machine learning models and general purpose computing devices.
The computing structures and models are generally linked to the use of the abstract ideas, without a disclosure in the specification of advances or improvements to the functioning or technologies of computing structures, machine learning models, or improvements to the technical field of machine learning model training to perform classifications via matrices, loss functions, weighting thresholds, vector embedding procedures, text recognition, classifying characterized data, contextual data understanding, or the legal field. Therefore, since the claims merely generally link the additional elements to the use of abstract ideas, use these additional elements as tools to perform the abstract ideas, and said additional elements are not revealed as improvements in the specification, the claims apply or use the abstract ideas without revealing another meaningful use or application. The claims are merely a drafting effort designed to monopolize the judicial exceptions.
The claim also recites claim limitations that are merely data or groups of data, i.e. mere characterizations of data. Characterized data is non-functional descriptive information.
The claims recite: receive a document and receive indications, which are transmitting data. The specification does not reveal advances to the technologies of data transmission, especially where the claims fail to limit how the transmission occurs.
The claims, as a whole, are not indicative of additional elements that integrate the abstract idea into a practical application.
Step 2B: The analysis above for Step 2A Prong 2 is commensurate with the analysis for this Step 2B, such that the claims, as a whole, do not include additional elements that are sufficient to amount to significantly more than the judicial exception when taken individually and in combination (MPEP 2106.05).
Dependent Claim:
Regarding claim 2: This claim is merely transmitting data to display on a device in a distributed computing system. The specification does not reveal advances to transmitting or displaying data. The claim is focused on the descriptive nature of the data being transmitted and displayed. Therefore, it is not indicative of a practical application.
Regarding claim 3 This claim is merely storing data that was returned from the machine learning model. The model does not perform any functions. Therefore, the machine learning model does not integrate the claim into a practical application or amount to significantly more. The specification does not reveal advances to data storage techniques or to the storage hardware, databases, or database architecture. The claim is focused on the descriptive nature of the data being stored. Therefore, it is not indicative of a practical application.
Regarding claim 4: This claim recites limitations that further embody the abstract idea categories of the independent claim: generate a second document, generate additional documents or clauses, which are additional functions in the abstract ideas categories of “mental processes” and "Certain Methods of Organizing Human Activity" for managing personal behaviors because they further perform an existing human task of creating documents/clauses. Insofar as there is a storing element, the specification does not reveal advances to storing, databases, or database architecture. The additional element is the multi-tenant database system, recited in the claim and disclosed in the specification, at a high level of generality, as one or many various general-purpose database systems. The data in the claim is merely non-functional descriptive information, i.e. characterized data, that is not patentably distinct. The specification does not reveal that the invention makes advances to databases or database architecture. The claim focuses on the descriptive nature of the data and the functional claim limitations without detailing how the claim performs the functions. These additional elements cannot be relied upon to integrate the abstract ideas of the claim into a practical application or significantly more.
Regarding claim 5: This claim is merely transmitting data to a device in a distributed computing system. The specification does not reveal advances to transmitting data. The claim is focused on the descriptive nature of the data being transmitted. Therefore, it is not indicative of a practical application.
Regarding claim 6: This claim recites limitations that further embody the abstract idea categories of the independent claim: legal clauses are as associated to tenant identifier, which is an additional function in the abstract ideas categories of “mental processes” and "Certain Methods of Organizing Human Activity" for managing personal behaviors because it further performs an existing human task of associating/matching data. Insofar as there is a storing element, the specification does not reveal advances to storing, databases, or database architecture. The additional element is the multi-tenant database system, recited in the claim and disclosed in the specification, at a high level of generality, as one or many various general-purpose database systems. The data in the claim is merely non-functional descriptive information, i.e. characterized data, that is not patentably distinct. The specification does not reveal that the invention makes advances to databases or database architecture. The claim focuses on the descriptive nature of the data and the functional claim limitations without detailing how the claim performs the functions. These additional elements cannot be relied upon to integrate the abstract ideas of the claim into a practical application or significantly more.
Regarding claim 7: This claim recites limitations that further embody the abstract idea categories of the independent claim: determine portions of documents, determine size of document, determine context window size of machine learning model, which are additional functions in the abstract ideas categories of “mental processes” and "Certain Methods of Organizing Human Activity" for managing personal behaviors because the claim further performs an existing human task of making decisions and determining what to put into a computer. The additional element is the machine learning model, recited in the claim and disclosed in the specification, at a high level of generality, as one or many available open-source general-purpose machine learning models, that does not perform and functions in this claim. The data in the claim is merely non-functional descriptive information, i.e. characterized data, that is not patentably distinct. The specification does not reveal that the invention makes advances to machine learning models. The claim focuses on the descriptive nature of the data and the functional claim limitations without detailing how the claim performs the functions. These additional elements cannot be relied upon to integrate the abstract ideas of the claim into a practical application or significantly more.
Regarding claim 8: This claim recites limitations that further embody the abstract idea categories of the independent claim: determine a start and an end of a document first portion, find new line, a full stop, a second header, and white space in the document, which are additional functions in the abstract ideas categories of “mental processes” and "Certain Methods of Organizing Human Activity" for managing personal behaviors because the claim further performs an existing human task of making decisions and determining what sections of a document. There are no additional elements in this claim, therefore there is no additional element that can integrate the claim into a practical application or significantly more. The data in the claim is merely non-functional descriptive information, i.e. characterized data, that is not patentably distinct. The claim focuses on the descriptive nature of the data and the functional claim limitations without detailing how the claim performs the functions. These additional elements cannot be relied upon to integrate the abstract ideas of the claim into a practical application or significantly more.
Regarding claims 9: This claim is merely receiving data with a particular characterization. The additional element is a JavaScript Object Notation (JSON) array, which is a general-purpose data structure holding data, the data being non-functional descriptive information. This element does not perform any functions, so it is not an additional element that is indicative of a practical application or significantly more. The specification does not reveal advances to receiving or characterizing data. The claim is focused on the descriptive nature of the data being transmitted and displayed. Therefore, it is not indicative of a practical application.
Regarding claim 10: This claim recites limitations that further embody the abstract idea categories of the independent claim: generate a JSON array, which is an additional function in the abstract ideas categories of “mental processes” and "Certain Methods of Organizing Human Activity" for managing personal behaviors because the claim further performs an existing human task of populating a data structure with characterized data. The additional element is the JSON array, recited in the claim and disclosed in the specification, at a high level of generality, as a general purpose data structure holding characterized data, which does not perform and functions in this claim. The data in the claim is merely non-functional descriptive information, i.e. characterized data, that is not patentably distinct. The specification does not reveal that the invention makes advances to databases, database architecture, or JSON data formatting. The claim focuses on the descriptive nature of the data and the functional claim limitations without detailing how the claim performs the functions. These additional elements cannot be relied upon to integrate the abstract ideas of the claim into a practical application or significantly more.
Regarding claims 11-13: These claims recite limitations that further embody the abstract idea categories of the independent claim. The limitations from claims 11-13 recite the prose reflecting the application of a LoRA model. Update the current layer through updating the attention and feed forward layers in claim 11: define weight matrices associations to the attention and/or feed forward layers, iteratively multiply matrix of weights of layers to iteratively update said weight matrices comprises: check matrix sizes equivalencies between determined layers and current weight matrices, and apply weight matrices to current weight matrices for layers. Iteratively repeating claim 11 in claim 12: iteratively update the matrices pairs for at least each document. For claim 13: refrain from updating the current weight matrices for the layers during iterative updating, which are additional functions in the abstract ideas categories of “mathematical concepts,” more specifically, “mathematical relationships” and “mathematical calculations.” These limitations are also “mental processes” and "Certain Methods of Organizing Human Activity" for managing personal behaviors because the claims perform and automate an existing human task of matrix factorization using low rank matrix decomposition that isolates small sections of the problem in order to solve for large data sets more efficiently.
The matrices manipulations are not tied to a specific computational method or technological improvement, but instead reflect a high-level abstraction of manipulating mathematical objects conceptually. Moreover, the claim does not recite any details of how these functions are implemented in practice, thus, the claim merely recites automation of matrix manipulations that were historically performed by hand.
The additional element is a machine learning model, recited in the claim and disclosed in the specification at a high level of generality. The specification discloses ¶ [0033] “The system 100 may perform efficient fine-tuning of the machine learning model using a low-rank adaptation (LoRA) technique and freezing weights… to … reduce … compute overhead.” Claim 11 outlines the general function of one iteration of a LoRA model, while claim 12 iterates and claim 13 recites practice of weight freezing that is inherent with the particular model function. The specification also discloses in ¶ [0036] “The machine learning model225 may be an example of an LLM (e.g., an artificial neural network), a classical machine learning model, or any other machine learning model. And in ¶ ¶ [0061] “the machine learning model400 (e.g., a pre-trained open-source LLM), i.e. the models are not only general-purpose models, but the models are freely available to the public for use and therefore are not a patentably distinct models.
The data in the claim, including parameters of the matrices, are merely non-functional descriptive information, i.e. characterized data, that are not patentably distinct. Therefore, using a publicly available model to update matrices, associate matrices with layers, multiply matrix weights, determine weight matrices, equilibrate matrices sizes, apply weight matrices to current weight matrices for layers, or determine an updated model, which are any publicly available, open source models, even assuming the use of specific characterized data, is merely applying the models as a tool to implement the abstract ideas, i.e. adding the words “apply it.” The specification does not reveal that the invention makes advances to databases or database architecture. The claim focuses on the descriptive nature of the data and the functional claim limitations without detailing how the claim performs the functions. These additional elements cannot be relied upon to integrate the abstract ideas of the claim limitations individually and as a combination.
Regarding claim 14: This claim recites limitations that further embody the abstract idea categories of the independent claim: determine one-to-one mappings, and determine mappings based on a string match, an edit distance, or a unigram overlap analysis, which are additional functions in the abstract ideas categories of “mental processes” and "Certain Methods of Organizing Human Activity" for managing personal behaviors because the claim further performs an existing human task of mapping data to other data. These manipulations are not tied to a specific computational method or technological improvement in the specification, but instead reflect a high-level abstraction of manipulating data objects, one to one, conceptually. Moreover, the claim does not recite any details of how these functions are implemented in practice, thus, the claim merely recites data comparison manipulations that were historically performed by hand. There are no additional elements in the claim. The data in the claim is merely non-functional descriptive information, i.e. characterized data, that is not patentably distinct. Since there are no additional elements, the claim cannot be integrated into a practical application or amount to significantly more.
Regarding claims 15 and 16: These claims recite limitations that further embody the abstract idea categories of the independent claim: claim 15 recites: determine a false positive error based on one to one mappings, and apply the false positive error to updating; claim 16 recites: determine a false negative error based on one to one mappings, and apply the false negative error to updating, which are additional functions in the abstract ideas categories of “mental processes” and "Certain Methods of Organizing Human Activity" for managing personal behaviors because the claim further performs an existing human task of reviewing one-to-one mappings and finding false positives and negatives to update the mappings with. The additional element is the machine learning model, recited in the claims and disclosed in the specification, at a high level of generality, as one or many available open-source general-purpose machine learning models, that does not perform and functions in this claim. The data in the claim is merely non-functional descriptive information, i.e. characterized data, that is not patentably distinct. The specification does not reveal that the invention makes advances to machine learning models or to fine-tuning the models. The claim focuses on the descriptive nature of the data and the functional claim limitations without detailing how the claim performs the functions. These additional elements cannot be relied upon to integrate the abstract ideas of the claim, individually and in combination, into a practical application or significantly more.
Regarding claim 17: This claim recites limitations that do not embody the abstract idea categories of the independent claim. The claim merely defines how the token weights may be assigned, without positively reciting any functional steps. The additional element is the machine learning model, recited in the claims and disclosed in the specification, at a high level of generality, as one or many available open-source general-purpose machine learning models, that does not perform and functions in this claim. The data in the claim is merely non-functional descriptive information, i.e. characterized data, that is not patentably distinct. The specification does not reveal that the invention makes advances to machine learning models or to fine-tuning the models. The claim focuses on the descriptive nature of the data and the descriptive nature of what occurs when weights are assigned, without actually performing the assigning. Since there are no abstract ideas, the claim cannot be integrated into a practical application, nor amount to significantly more.
Regarding claim 18: This claim recites limitations that further embody the abstract idea categories of the independent claim: from the document, determine individuals, entities, or both, which are additional functions in the abstract ideas categories of “mental processes” and "Certain Methods of Organizing Human Activity" for managing personal behaviors because the claim further performs an existing human task of identifying data in a document. The manner of recitation of a natural language processing analysis in the claim could reasonably be interpreted as another abstract idea of analyzing natural language in both the “mental processes” and "Certain Methods of Organizing Human Activity" categories for managing personal behaviors of analyzing text based on the language presented to find the required elements by hand. The manner of recitation could also reasonably be interpreted to be a Natural Language Processor (NLP) Model or algorithm. For the purposes of compact prosecution, the Examiner is interpreting this to be an NLP model. Therefore, the additional element is an NLP model, recited in the claim and disclosed in the specification at a high level of generality. The specification discloses that the models of this instant application may be one or many available open-source general-purpose machine learning models that does not perform any functions as the determining is merely “based on” the analysis type. The data in the claim is merely non-functional descriptive information, i.e. characterized data, that is not patentably distinct. The specification does not reveal that the invention makes advances to machine learning models, to natural language processing, or to fine-tuning the models. The claim focuses on the descriptive nature of the data and the functional claim limitations without detailing how the claim performs the functions. These additional elements cannot be relied upon to integrate the abstract ideas of the claim, individually and in combination, into a practical application or significantly more.
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.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claim 1-6 and 9-20 are rejected under 35 U.S.C. § 103 as being unpatentable over Bonfante, US-20240070794-A1, in view of Han, US-20220004713-A1, in further view of Chen, US-20220383126-A1.
Regarding claims 1, 19, and 20: Bonfante discloses:
receiving, from a first user device, a document; [0014] (system enables receipt of document from at least a first user, through their device);
receiving, from a second user device, an indication of a first set of legal clauses within the document, wherein a legal clause comprises a name and text indicating a legal significance of the legal clause; (Per instant application specification ¶ [0029], significance is “an example of a specific point or provision in a law or legal document”), [0017-0022] (second users, via individuals’ devices, analyze documents and identify legal obligations, i.e. legal clauses, where the clauses may be tagged as legal obligations or attributes, and sent, received, managed and stored by the system), [0031] (name and description of the document sections are presented);
inputting a least a portion of the document into the machine learning model, the machine learning model outputting a second set of legal clauses responsive to at least the portion of the document input into the machine learning model; [0035] (input document into the machine learning model and the model identifies legal obligations, i.e. the set of legal clauses, responsive to the input);
determining a plurality of one-to-one mappings between the first set of legal clauses and the second set of legal clauses based at least in part on a vector embedding procedure for the first set of legal clauses and the second set of legal clauses, wherein a mapping of the plurality of one-to-one mappings comprises a first legal clause from the first set of legal clauses and a second legal clause from the second set of legal clauses; and
iteratively updating the machine learning model, based at least in part on an on an evaluation metric corresponding to one to one mappings, an accuracy, using at least the common substring between the first legal clause and the second legal clause of the mappings. [0027] (models may be retrained for new data), [0029] (machine learning models updated over time, i.e. iterations), [0032] (identified legal clauses are evaluated based on an accuracy of the identified legal clauses and other evaluation metrics, against the text of historical legal obligations; and [0033] (where the evaluation data, including the accuracy, i.e. evaluation metrics, are input back into the model for model retraining; legal documents and identified legal obligations corresponding to evaluation metrics data are designated historical contract documents and historical legal obligations used to retrain the model to identify legal clauses and documents)), [0037] (historical legal obligations are evaluated according to a metric of how common or uncommon each legal text is, where:) [0040] “machine-learned model 300 learns to correlate a presence of one or more contract attributes within the historical contract documents 315 and portions of text corresponding to legal obligations 325 within the historical contract documents,” (where contract attributes are considered negative training sets and legal obligations are considered positive training sets, such that text corresponding to legal obligations are identified based on historical positive training sets), [0041] (the machine learning model identifies portions of the text that correspond between the historical document legal clauses (first legal clause) and contract document (second legal clauses), (i.e. the machine learning model is updated based at least in part on an accuracy evaluation metric, where the metric corresponds to mapping of one string of text to another string of text to identify accurate text strings), and [0047-0051] (portions of text from the training set of historical documents is compared to portions of text from an input document to identify portions of the input text that match portions of text from the training set using machine learning models, i.e. text mapping);
Where Bonfante does not disclose, Han teaches:
updating a loss function based at least in part on a frequency of one or more words of the document by assigning token weights to one or more tokens of the one or more words, based at least in part on a frequency of one or more words of the document; [0044] (vectors/tensors are assigned to each word in a document, binary classifications are applied to map distinct legal clauses), [0045] (the loss function is updated and parameters are adjusted, where parameters are synonymous with weights, “The loss value represents a degree of accuracy of the output of the ML model, a degree of difference between the output of the ML model and an expected output of the ML model,”), [0046] (tokens and parameter values “combine legal expertise with lexical features to determine a relevant set of words,”” wr,c can be provided as legal terms that occur to the clause and c is a string representation of the clause … The function ƒ first stores wr,c into a hash structure, and a legal document D is parsed into an array of sentences S=[s1, . . . , sn]. In some examples, wr,c is a array mapping from legal terms to an integer indices … Each sentence si∈S is parsed into individual tokens. Each token conducts a lookup in wr,c for a match. In some examples, a token search in wr,c is by exact match.” i.e. longest common substrings), [0047] “In some examples, a sentence is first identified as a relevant sentence based on a Regular Expression match with predefined keywords and sentence structures.”
automatically updating a machine learning model based at least in part on the loss function; [0045] “a loss value is determined based on a loss function. The loss value represents a degree of accuracy of the output of the ML model … if the loss value does not meet an expected value (e.g., is not equal to zero), parameters of the ML model (e.g., parameters of the function being learned) are adjusted, and another iteration of training is performed … this process is repeated until the loss value meets the expected value.”
determining a plurality of one-to-one mappings between the first set of legal clauses and the second set of legal clauses based at least in part on a vector embedding procedure for the first set of legal clauses and the second set of legal clauses, wherein a mapping of the plurality of one-to-one mappings comprises a first legal clause from the first set of legal clauses and a second legal clause from the second set of legal clauses; and [0044-0048] (system maps words, phrases, or legal phrases, word by word by assigning tensors/vectors, i.e. “a set of real-valued vectors (i.e., a tensor) corresponding to each word,” i.e. based on vector embeddings, are mapped using “a set of binary classifications that map onto the status of distinct legal clauses,” such that exact legal clauses are mapped word by word to determine exact common substrings. words and determines if the first legal clause from a document matches a second legal clause from a plurality of documents),
iteratively updating the machine learning model, based at least in part on an evaluation metric corresponding to the plurality of one to one mappings failing to satisfy an accuracy threshold, wherein the evaluation metric is calculated using at least a longest common substring between the first legal clause and the second legal clause of the mapping; [0044] (analysis method, i.e. determination, i.e. calculation of at least an evaluation metric, such that [0045] (the machine learning model is iteratively trained, i.e. iteratively updated, based on a loss value, i.e. an evaluation metric, that represents an accuracy metric determined based on a loss function that represents an accuracy of the output as compared to the expected output, said evaluation metric, i.e. the loss value, corresponds to the one to one mappings, and is calculated based on longest common substrings, where the longest common substrings are implemented in determining said evaluation metric, of the ML model; “if the loss value (i.e. the evaluation metric) does not meet an expected value… (i.e. fails to satisfy the accuracy metric threshold) parameters of the ML model (e.g., parameters of the function being learned, [i.e., the first, second, or both pairs of weight matrices]) are adjusted, and another iteration of training is performed. In some instances, this process is repeated until the [evaluation metric] meets the expected value (i.e. the metric succeeds at satisfying the accuracy threshold”)), [0046] (text strings are matched by vector values broken into tokens, where the highest number of matching tokens represents the longest string of values in the evaluation, i.e. the model implements at least a longest common substring between the first and second legal clauses of the mapping to calculate the evaluation metric);
Where Bonfante does not disclose and Han does teach, Chen further teaches overlapping applicability leading to the improvement of Chen over Han:
updating a loss function based at least in part on one or more words of the document by assigning token weights to one or more tokens of the one or more words based at least in part on the one or more words in the document; [0031] “. Training may include the use of a loss function using standard backpropagation, calculating a gradient for every parameter and updating weights by subtracting the gradients,”
Where Bonfante does not disclose and Han does not teach, Chen also teaches:
automatically updating a first pair of weight matrices of a machine learning model, a second pair of weight matrices of the machine learning model, or both; [0026] (update weight matrices associated with either layer of machine learning model; weight matrices are updated according to weighting constraints, which are equivalent to an evaluation metric), [0048] (two distinct layers), [0032-0033] (both the first and second pairs of weight matrices are updated);
iteratively updating the first pair of weight matrices of the machine learning model, the second pair of weight matrices of the machine learning model, or both, based at least in part on an evaluation metric corresponding to mappings failing to satisfy an accuracy threshold, wherein the evaluation metric is calculated [0042] (matrix results are compare to an error metric, i.e. looking for accuracy of the LoRA function of the model, i.e. the model is optimized until the error threshold is reached, where the an error implicitly discloses an accuracy metric, that corrects the weights iteratively while the model no longer indicates an error, while the threshold is not satisfied, when accuracy is at a threshold, the values may be implemented back into the main model);
It would have been obvious to a person having ordinary skill in the art to have recognized, before the effective filing date, to combine the improvement from the teachings of Chen with the base combined prior art disclosures of Bonfante and Han, using the application of known techniques to yield a predictable result that achieves an improvement upon the base device.
Regarding claim 2: Bonfante discloses, and Han and Chen teach: The method of claim 1,
Bonfante discloses: further comprising: transmitting, for display at a third user device, an indication of the evaluation metric, the first set of legal clauses, the second set of legal clauses, a plurality of longest common substring results for the plurality of one-to-one mappings, or any combination thereof. [0014] (system enables one, two, three, or more users to receive document data from one, two, three, or more other users, each through their devices), [0022] (one or more legal obligations may be transmitted), [0027] (presents legal obligations to users), [0028] “a third party,” [0042] (the system displays the legal obligations found by the model, the documents, as well as information about the found obligations, which may include the determined metrics and string matching results), [0044] (display includes risks, i.e. the evaluation metric), [0052] (system generates a display interface to display the above to one, two, three, or more users via their devices).
Regarding claim 3: Bonfante discloses, and Han and Chen teach: The method of claim 1,
Bonfante discloses: further comprising: storing a plurality of legal clauses output by the updated machine learning model. [0025] (the system stores legal obligations output from the machine learning model in databases).
Regarding claim 4: Bonfante discloses, and Han and Chen teach: The method of claim 3,
Bonfante discloses: further comprising: generating a second document for a tenant of a multi-tenant database system based at least in part on one or more legal clauses of the stored plurality of legal clauses associated with the tenant; [0015] (multi-party system), [0025] (system with databases), [0014] (party is a user, organization, etc., where a party may include more than one user), [0014] “A document management system enables a party (e.g., individuals, organizations, etc.) to create and send documents to one or more receiving parties for negotiation, collaborative editing, electronic execution (e.g., via electronic signatures), contract fulfillment, archival, analysis, and more,”
storing the second document for the tenant; and [0018] (store documents);
generating one or more additional documents, one or more additional legal clauses, or both for the tenant based at least in part on the stored second document for the tenant and the one or more legal clauses of the stored plurality of legal clauses associated with the tenant. [0014] “A document management system enables a party (e.g., individuals, organizations, etc.) to create and send documents to one or more receiving parties for negotiation, collaborative editing, electronic execution (e.g., via electronic signatures), contract fulfillment, archival, analysis, and more,” and [abstract] (the system trains on historical legal documents to generate legal clauses based on a plurality of stored documents, i.e. the document created, i.e. the only document created, becomes a historical document and the process repeats).
Regarding claim 5: Bonfante discloses, and Han and Chen teach: The method of claim 3,
Bonfante discloses: further comprising: transmitting, to a fourth user device associated with a tenant of a multi-tenant database system, a suggested legal clause based at least in part on the stored plurality of legal clauses and a legal district associated with the tenant, a geographic location associated with the tenant, a request associated with the tenant, or any combination thereof. [0014] (one, two, three, four, or more users transmit documents via their devices to any number of new users who receive the documents via their devices, i.e. a multi-tenant system); [0045] (the suggested legal clauses are presented based on rank, based in part on stored historical legal clauses, which may include the tenant’s geographic location, priorities of the entity, jurisdiction, or a plurality of alternate reasons).
Regarding claim 6: Bonfante discloses, and Han and Chen teach: The method of claim 3,
Bonfante discloses: wherein each legal clause of the plurality of legal clauses is stored with an association to a tenant identifier of a multi-tenant database system. [0025] (legal clauses are stored in association to information about the users including client device identifiers).
Regarding claim 9: Bonfante discloses, and Han and Chen teach: The method of claim 1,
wherein receiving the indication of the first set of legal clauses within the document comprises: comprising the first set of legal clauses, [0029], significance is “an example of a specific point or provision in a law or legal document”), [0017-0022] (second users, via individuals’ devices, analyze documents and identify legal obligations, i.e. legal clauses, where the clauses may be tagged as legal obligations or attributes, and sent, received, managed and stored by the system);
Where Bonfante does not disclose, and Chen and Higgins do not teach: Han teaches: receiving a JavaScript Object Notation (JSON) array. [0030] and [0046] (data is parsed into JSON arrays).
It would have been obvious to a person having ordinary skill in the art to have recognized, before the effective filing date, to combine the improvement from the teachings of Han with the base disclosure of Bonfante using the application of known techniques to yield a predictable result that achieves an improvement upon the base device.
Regarding claim 10: Bonfante discloses, and Han and Chen teach: The method of claim 1,
Bonfante discloses: further comprising: generating comprising the first set of legal clauses based at least in part on the indication of the first set of legal clauses within the document. [0017-0022] (second users, via individuals’ devices, analyze documents and identify legal obligations, i.e. legal clauses, where the clauses may be tagged as legal obligations or attributes, and sent, received, managed and stored by the system), [0037] (historical legal obligations are ranked according to a risk value metric of how common or uncommon each obligation is, where:) [0040] “machine-learned model 300 learns to correlate a presence of one or more contract attributes within the historical contract documents 315 and portions of text corresponding to legal obligations 325 within the historical contract documents,” [0041] (the machine learning model identifies portions of the text that correspond between the historical document legal clauses (first legal clause) and contract document (second legal clauses), (i.e. the machine learning model is updated based at least in part on the evaluation metric, where the metric is part of the correlation of, e.g. mapping, of one string in one document to another), and [0047-0051] (portions of text from the training set of historical documents is compared to portions of text from an input document to identify portions of the input that match portions of text from the training set using machine learning models);
Where Bonfante does not disclose, and Chen and Higgins do not teach, Han teaches: generating a JavaScript Object Notation (JSON) array; [0030] and [0046] (data is parsed into JSON arrays).
It would have been obvious to a person having ordinary skill in the art to have recognized, before the effective filing date, to combine the improvement from the teachings of Han with the base disclosure of Bonfante using the application of known techniques to yield a predictable result that achieves an improvement upon the base device.
Regarding claim 11: Bonfante discloses, and Han and Chen teach: The method of claim 1,
57. Where Bonfante does not disclose and Han and Higgins do not teach, Chen teaches:
The first pair of weight matrices is associated with an attention layer of the machine learning model; the second pair of weight matrices is associated with a feed forward layer of the machine learning model; wherein the machine learning model comprises one or more attention layers, one or more feed forward layers, or both; [0026] (update weight matrices associated with either layer of machine learning model), [0048] (two distinct layers), [0026] (weight matrices are updated according to weighting constraints, which are equivalent to an evaluation metric);
and iteratively updating the first pair of weight matrices, the second pair of weight matrices, or both, comprises: multiplying the first pair of weight matrices to determine a first weight matrix and the second pair of weight matrices to determine a second weight matrix, wherein a first size of the first weight matrix is equal to a second size of a first current weight matrix for the attention layer, and wherein a third size of the second weight matrix is equal to a fourth size of a second current weight matrix for the feed forward layer; and [0045] (update the matrix pairs, i.e. a plurality of matrix pairs are updated using a tuning matrix multiplied by the unchanged base matrix), [0055] (the size of the base model weight matrices for each layer are equal to the sizes of the current weight matrices for each layer), and [0060] (for each LoRA iteration, r x d and d x r make a matrix d x d);
applying the first weight matrix to the first current weight matrix for the attention layer and the second weight matrix to the second current weight matrix for the feed forward layer to determine the updated machine learning model. [0048] (iteratively run and update model including using LoRA matrix weightings on each layer).
It would have been obvious to a person having ordinary skill in the art to have recognized, before the effective filing date, to combine the improvement from the teachings of Chen with the base disclosure of Bonfante combined with the base teachings of Han using the application of known techniques to yield a predictable result that achieves an improvement upon the base disclosure and teaching combination.
Regarding claim 12: Bonfante discloses, and Han and Chen teach: The method of claim 11,
Where Bonfante does not disclose and Han and Higgins dos not teach, Chen teaches: wherein iteratively updating first pair of weight matrices, the second pair of weight matrices, or both further comprises: iteratively updating the first pair of weight matrices, the second pair of weight matrices, or both based at least in part on a plurality of documents. [0048] (iteratively run and update model including using LoRA matrix weightings on each layer).
It would have been obvious to a person having ordinary skill in the art to have recognized, before the effective filing date, to combine the improvement from the teachings of Higgins with the base disclosure of Bonfante combined with the base teachings of Han using the application of known techniques to yield a predictable result that achieves an improvement upon the base disclosure and teaching combination.
Regarding claim 13: Bonfante discloses, and Han and Chen teach: The method of claim 12,
Where Bonfante does not disclose and Han and Higgins do not teach, Chen teaches: further comprising: refraining from modifying the first current weight matrix for the attention layer and the second current weight matrix for the feed forward layer during the iterative updating. [0027] (the first weight of each discrete layer is set to zero).
It would have been obvious to a person having ordinary skill in the art to have recognized, before the effective filing date, to combine the improvement from the teachings of Higgins with the base disclosure of Bonfante combined with the base teachings of Han using the application of known techniques to yield a predictable result that achieves an improvement upon the base disclosure and teaching combination.
Regarding claim 14: Bonfante discloses, and Han and Chen teach: The method of claim 1,
Where Bonfante does not disclose, Han teaches: wherein determining the plurality of one-to-one mappings comprises: determining the plurality of one-to-one mappings further based at least in part on a String match analysis, an edit distance analysis, a unigram overlap analysis, or any combination thereof for the first set of legal clauses and the second set of legal clauses. [0044-0048] (system maps words, phrases, or legal phrases to vectors using a procedure that incorporates vector embedding words and determines if the first legal clause from a document matches a second legal clause from a plurality of documents); [0046] (text strings are matched by vector values broken into tokens, where the highest number of matching tokens represents the longest string of values in the evaluation);
It would have been obvious to a person having ordinary skill in the art to have recognized, before the effective filing date, to combine the improvement from the teachings of Han with the base disclosure of Bonfante using the application of known techniques to yield a predictable result that achieves an improvement upon the base device.
Regarding claim 15: Bonfante discloses, and Han and Chen teach: The method of claim 1,
Where Bonfante does not disclose, Han teaches: wherein determining the plurality of one-to-one mappings comprises: determining a false positive error for the machine learning model based at least in part on a third legal clause of the second set of legal clauses failing to map to a fourth legal clause of the first set of legal clauses based at least in part on the plurality of one-to-one mappings, wherein updating the machine learning model is further based at least in part on the false positive error. [0045] (difference between output and expected output) and [0052] (negative feedback on the model with respect to accuracy), [0057] (accurate predictions from the model may be approved, inaccurate predictions, i.e. false positives or false negatives, may be denied, both of which are used for training the model, e.g. updating the model is synonymous with training the model in this prior art.)
It would have been obvious to a person having ordinary skill in the art to have recognized, before the effective filing date, to combine the improvement from the teachings of Han with the base disclosure of Bonfante using the application of known techniques to yield a predictable result that achieves an improvement upon the base device.
Regarding claim 16: Bonfante discloses, and Han and Chen teach: The method of claim 1,
Where Bonfante does not disclose, and Chen and Higgins do not teach, Han teaches: wherein determining the plurality of one-to-one mappings comprises: determining a false negative error for the machine learning model based at least in part on a third legal clause of the first set of legal clauses failing to map to a fourth legal clause of the second set of legal clauses based at least in part on the plurality of one-to¬ one mappings, wherein updating the machine learning model is further based at least in part on the false negative error. [0045] (difference between output and expected output) and [0052] (negative feedback on the model with respect to accuracy), [0057] (accurate predictions from the model may be approved, inaccurate predictions, i.e. false positives or false negatives, may be denied, both of which are used for training the model, e.g. updating the model is synonymous with training the model in this prior art.)
It would have been obvious to a person having ordinary skill in the art to have recognized, before the effective filing date, to combine the improvement from the teachings of Han with the base disclosure of Bonfante using the application of known techniques to yield a predictable result that achieves an improvement upon the base device.
Regarding claim 17: Bonfante discloses, and Han and Chen teach: The method of claim 1,
Where Bonfante does not disclose, and Chen and Higgins do not teach: Han teaches: wherein:
The token weights are assigned to the one or more tokens further based at least in part on the one or more words and a corpus of legal language associated with a plurality of legal clauses.[0045] (parameters are adjusted, where parameters are synonymous with weights), [0046] (tokens and parameter values “combine legal expertise with lexical features to determine a relevant set of words”),
It would have been obvious to a person having ordinary skill in the art to have recognized, before the effective filing date, to combine the improvement from the teachings of Han with the base disclosure of Bonfante using the application of known techniques to yield a predictable result that achieves an improvement upon the base device.
Regarding claim 18: Bonfante discloses, and Han and Chen teach: The method of claim 1, further comprising:
determining, from the document, one or more individuals, one or more entities, or both based at least in part on a natural language processing analysis of the document. [0025] (legal clauses are stored in association to information about the users including individuals and entities), [0040] (each machine learning model disclosed in the prior art, “linear support vector machines (linear SUM), boosting for other algorithms (e.g., AdaBoost), neural networks, logistic regression, naïve Bayes, memory based learning, random forests, bagged trees, decision trees, boosted trees, or boosted stumps,” have been known to incorporate natural language processing to identify data).
Claim 7-8 are rejected under 35 U.S.C. § 103 as being unpatentable over Bonfante, US-20240070794-A1, in view of Han, US-20220004713-A1, in further view of Chen, US-20220383126-A1, and in further view of Higgins, US-20240037682-A1.
Regarding claim 7: Bonfante discloses, and Han and Chen teach: The method of claim 1, further comprising:
Where Bonfante does not disclose and Han and Chen do not teach, Higgins teaches: determining a plurality of portions of the document for inputting separately into the machine learning model based at least in part on a size of the document and a context window size of the machine learning model. [0040] (data that exceeds a size threshold will be chunked, based on data size and context window size).
It would have been obvious to a person having ordinary skill in the art to have recognized, before the effective filing date, to combine the improvement from the teachings of Higgins with the base disclosure of Bonfante combined with the base teachings of Han using the application of known techniques to yield a predictable result that achieves an improvement upon the base disclosure and teaching combination.
Regarding claim 8: Bonfante discloses, and Han and Chen teach: The method of claim 7,
Where Bonfante does not disclose and Han and Chen do not teach, Higgins teaches: wherein determining the plurality of portions of the document comprises: determining a start of a first portion of the document, an end of the first portion of the document, or both based at least in part on a new line in the document, a full stop in the document, a section header in the document, a white space search of the document, or any combination thereof. [0118] (tools are utilized to visually determine portions of a document like: “Tableau, QlikView, Power BI, Looker, TIBCO Spotfire, SAP Lumira, IBM Cognos Analytics, Microsoft Excel (with Power View and Power Pivot), Google Data Studio, High charts, It may also be interacted with, layered and/or viewed through various lenses”), [0121] (determine the location of text in a document like page, line, etc., are detected and depicted) [0123] (flag and tag portions of text), [0124] (objects are detected, i.e. a start of a portion, an end of a portion, new lines in a document, full stops in a document, section headers, whitespaces, where these sections of a document are objects that may be visually identified, determined, flagged, saved, and utilized for any document positioning needs.)
It would have been obvious to a person having ordinary skill in the art to have recognized, before the effective filing date, to combine the improvement from the teachings of Higgins with the base disclosure of Bonfante combined with the base teachings of Han using the application of known techniques to yield a predictable result that achieves an improvement upon the base disclosure and teaching combination.
Conclusion
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 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 ANGELA HATCH whose telephone number is (571)270-1393. The examiner can normally be reached 10:00-6:00.
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, Nathan Uber can be reached at (571)270-3923. 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.
ANGELA HATCH
Examiner
Art Unit 3626
/ANGELA HATCH/Examiner, Art Unit 3626
/NATHAN C UBER/Supervisory Patent Examiner, Art Unit 3626
Prosecution Insights