Here's what actually issued. On October 3, 2023, International Business Machines Corporation was granted US11775839B2, "Frequently asked questions and document retrieval using bidirectional encoder representations from transformers (BERT) model trained on generated paraphrases," inventors including Yosi Mass and Haggai Roitman. The CPC codes are G06N 3/088, G06N 3/045 (network architectures), and retrieval code G06F 16/24578.

The mechanism is retrieval with a twist. BERT, the bidirectional transformer encoder, turns text into embeddings you can match for semantic similarity — the backbone of modern search and of the retrieval step in retrieval-augmented generation. The claimed twist is training data: the model is trained on generated paraphrases, synthetic rewordings of questions, so it learns to match a user's phrasing to a stored FAQ or document even when the wording differs.

“An example system includes a processor to receive a query. The processor can retrieve ranked candidates from an index based on the query.”— U.S. Patent No. 11,775,839 source

The independent claim turns that twist into a full pipeline, and the order of operations is the invention. Claim 1's processor first receives a query and retrieves ranked candidates from an index. Then comes the distinctive part: it "fine-tunes a generative pretrained transformer trained on question-answer pairs of a concatenated frequently asked question (FAQ) dataset using randomly sampled sequences of the concatenated FAQ dataset." In other words, a generative model (a GPT-style transformer) is first specialized on the FAQ corpus. That generative model is then used to "automatically generate paraphrases for questions in the FAQ dataset" — and crucially, it generates them "based on input answers from the FAQ dataset." So the system feeds an answer in and asks the generative model to produce plausible question phrasings for it, manufacturing synthetic question variants the original FAQ never listed.

Those synthetic paraphrases are not trusted blindly. Claim 1 next "filters the automatically generated paraphrases to match a same FAQ as their generation questions using the index" — each generated paraphrase is run back through the retrieval index, and only paraphrases that still retrieve the correct FAQ entry survive. That filter is a self-consistency check: it discards hallucinated or off-target rewordings before they can poison training. The filtered paraphrases then "fine-tune a Bidirectional Encoder Representations from Transformers (BERT) query-question (Q-q) model" to match queries to FAQ questions, and that fine-tuned BERT model re-ranks the originally retrieved candidates, which are returned to the user. The whole arc — generate paraphrases from answers, filter them by retrieval consistency, fine-tune a retriever on the survivors — is a concrete data-augmentation recipe, not a vague claim to "using BERT for search."

The dependent claims add the training and re-ranking specifics. Claim 2 trains the Q-q BERT model "using triplets comprising a question, a positive paraphrase, and a negative paraphrase" — classic contrastive learning, where the model is pulled toward valid rewordings and pushed away from invalid ones. Claim 3 describes a final re-ranking that combines several re-rankers "using an unsupervised late-fusion," and names them: the fine-tuned BERT query-question model, a separate BERT query-answer (Q-a) model, and a passage-based re-ranker. Claim 4 makes the fusion concrete — it "sums candidate scores assigned for each candidate" by those three re-rankers — and claim 5 adds "an unsupervised query expansion step" over the summed pool. Claim 14 even describes the generation stopping condition: a special token is input with each answer, the generative transformer produces tokens until that special token is reached, and all generated tokens become the paraphrase. These are engineering details with teeth: triplet training, three named re-rankers, summed late fusion, query expansion.

That data-augmentation angle is the heart of the invention. Generating paraphrases from answers to make a retriever robust to phrasing variation — and then filtering them through the index for consistency — is a concrete method, and it is directly relevant to enterprise help-desk and knowledge-base products, exactly IBM's commercial turf. As retrieval-augmented systems proliferate, retrieval quality becomes a competitive surface, and a method that improves it by synthesizing and vetting its own training data is worth owning.

On scope, the line holds: granted B2, enforceable, but the claims cover this specific paraphrase-generate-filter-fine-tune-and-late-fuse retrieval method, including the answer-conditioned generation, the index-based filtering, the triplet training, and the named three-way re-ranker fusion. They do not lock up BERT, semantic retrieval, or retrieval-augmented generation generally. Claim 1 defines the boundary, and that boundary is the pipeline, not the components.

It is worth drawing out why the index-based filtering step is the quiet linchpin of the whole method. A generative transformer asked to invent question phrasings from an answer will inevitably produce some rewordings that drift — questions that read fluently but point at a different FAQ entry, or at no entry at all. Training a retriever on those bad paraphrases would teach it the wrong associations, eroding rather than improving accuracy. Claim 1 closes that hole by running every generated paraphrase back through the same retrieval index and keeping only those that "match a same FAQ as their generation questions." The filter uses the system's own retriever as a judge of its own synthetic data, so the augmentation is self-correcting: only paraphrases that already behave like good queries survive to train the next, better retriever. That self-consistency check, combined with the contrastive triplet training in claim 2 — question, positive paraphrase, negative paraphrase — is what turns a noisy generative process into reliable supervision. And the three-way late fusion in claims 3 and 4 hedges the final ranking across complementary signals: the query-question model that benefits from all this paraphrase augmentation, a query-answer model that matches questions against answer text directly, and a passage-based re-ranker, their scores simply summed. The architecture is layered defenses against the failure modes of synthetic data and single-signal retrieval.

The takeaway: US11775839B2 is retrieval IP arriving as the RAG pattern goes mainstream — a method that improves a retriever by generating question paraphrases from answers, filtering them for consistency, and fusing three re-rankers, held by an incumbent with a deep enterprise-search business.