Data Analysis and Visualization

A financial analyst wants to extract quarterly revenue figures from a digitally-created PDF report - one that was exported directly from a spreadsheet application and contains selectable text. Which statement best describes the role of OCR in this scenario?

• AOCR is essential because all PDFs store content as images and cannot be parsed without it.
• BOCR should be applied first to improve table structure detection before the embedded text is read.
• COCR is not required and no further processing is needed because the PDF text stream preserves full table structure automatically.
• DOCR is not required because the embedded text layer can be extracted directly, though layout parsing may still be needed to correctly associate table cells with their headers.check_circle Correct

Why A is wrong: This conflates scanned PDFs with digitally-created ones. Digitally-created PDFs embed an encoded text stream that direct extraction libraries can access without any OCR step.

Why B is wrong: Running OCR on a PDF that already has an embedded text layer is redundant and risks introducing character substitution errors that do not exist in the source file.

Why C is wrong: PDF text streams encode reading-order character sequences but do not reliably preserve two-dimensional table structure. Layout analysis is typically required to reconstruct row and column relationships from positional data.

Why D is correct: Digitally-created PDFs include an encoded text stream, so OCR is unnecessary for character recovery. However, the PDF format does not natively encode table semantics, so a layout or table-extraction step is often still needed to correctly align cells with column and row headers.

A text classification model struggles to generalise to paraphrased user queries. Which augmentation approach best addresses this by creating new training examples that retain the original meaning and label?

• ABack-translating each sentence through an intermediate language to produce semantically equivalent paraphrasescheck_circle Correct
• BRandomly deleting up to 40% of words from each sentence before training
• CReplacing the sentence encoder with a larger pre-trained model at inference time
• DAdding Gaussian noise to the token embedding vectors during training as a regularisation step

Why A is correct: Back-translation translates the sentence into another language and then back into the source language, producing a natural restatement that preserves meaning and therefore label, directly addressing lexical and syntactic variation.

Why B is wrong: Aggressive random deletion at that rate frequently destroys the meaning of a sentence, which may alter the correct label and introduce noise rather than useful variation.

Why C is wrong: Swapping the encoder is a model architecture change, not a data augmentation technique; it does not add new training examples and does not directly expand the diversity of surface forms in the dataset.

Why D is wrong: Embedding noise is a regularisation strategy applied in latent space rather than a text-level augmentation; it does not create new readable training sentences that cover paraphrase variation at the surface level.

At which stage of a typical supervised training workflow is data augmentation conventionally applied?

• AOnly to the held-out test set, so that final evaluation reflects the model's behaviour on transformed inputs it will encounter in deployment.
• BTo the training set, frequently generated on the fly each epoch so the model sees different variants of the same samples across epochs.check_circle Correct
• CTo the validation set, so that early-stopping decisions are based on the model's response to randomly transformed examples.
• DEqually across the training, validation, and test sets so that all three partitions share an identical input distribution.

Why A is wrong: Tempting because robustness to transformed inputs matters at deployment, but augmenting the test set changes the benchmark and inflates or distorts reported metrics; standard practice keeps the test set fixed.

Why B is correct: Correct: augmentation is applied to training data, commonly sampled randomly each epoch, so the same source image yields different variants over time and broadens the training signal.

Why C is wrong: Tempting because validation guides training decisions, but introducing random augmentation makes validation metrics noisy and non-reproducible; the validation set is normally kept unaugmented.

Why D is wrong: Tempting because matching distributions sounds principled, but augmenting evaluation sets corrupts the measurement of true generalisation; augmentation is reserved for training data.

A developer runs OCR on faded photocopies and the recognition accuracy is poor. A colleague suggests adding a pre-processing step that converts each greyscale scan into a clean two-tone image, separating dark text pixels from the light background before recognition. Which image pre-processing technique does this describe?

• ABinarisation, which thresholds the greyscale pixels into foreground text and background to sharpen character edges.check_circle Correct
• BRotation, which turns the page by a fixed angle so the text reads left to right before recognition begins.
• CCropping, which removes the page margins so only the central text block is passed to the recognition engine.
• DUpscaling, which increases the pixel dimensions of the scan so that small characters occupy more pixels overall.

Why A is correct: Correct: binarisation applies a threshold that splits greyscale pixels into black foreground and white background, giving the engine crisp character shapes.

Why B is wrong: Tempting because orientation matters for OCR, but rotation corrects angle and does nothing to separate text pixels from the background by tone.

Why C is wrong: Tempting because trimming clutter can help, but cropping only selects a region; it does not convert greyscale into a clean two-tone image.

Why D is wrong: Tempting because resolution affects OCR, but upscaling adds pixels through interpolation and never reduces the image to two separated tones.

A product team asks an LLM to summarise a 5,000-word market research report into a 200-word executive brief. After reviewing outputs, they find some summaries omit critical risk findings buried in the middle of the report. Which technique best mitigates this selective-omission problem while preserving brevity?

• ASwitch to a text-to-image model to render the report as an infographic, then summarise the visual output
• BAsk the model to first produce a bullet-point extraction of all risk findings, then use those bullets as explicit context when writing the final summarycheck_circle Correct
• CIncrease the summarisation temperature setting to 1.5 so the model explores a broader range of content
• DEmbed the report and retrieve the top-five most similar chunks to the query 'executive summary', then concatenate those chunks as the output

Why A is wrong: Text-to-image models generate images from descriptions; they do not read or comprehend long documents. Routing report text through an image model does not address content completeness in a written summary.

Why B is correct: A two-step chain - extract first, then summarise with extracted content in context - forces the model to surface mid-document facts before condensing. The extraction step acts as a structured retrieval pass, preventing salience bias that favours text near the start or end of the input.

Why C is wrong: Raising temperature increases randomness and diversity of word choices; it does not guarantee coverage of specific content sections. High temperature is more likely to introduce hallucinations than to improve completeness.

Why D is wrong: Retrieval via semantic similarity returns chunks relevant to the query term, but 'executive summary' as a query will likely surface introductory text rather than mid-document risk findings. This approach does not guarantee the completeness needed for an executive brief.