Session Dataset

A datsetsSessionDataset is a DocumentArray, in which each Document has matches Document. The labels are stored under matched Document’s .tags['finetuner_label']. The label can be either 1 (denoting similarity of match to its parent Document) or -1 (denoting dissimilarity of the match from its parent Document).

The word “session” comes from the scenario where a user’s click-through introduces an implicit yes/no on the relevance, hence an interactive session.

Batch construction

A SessionDataset works with datsetsSessionSampler.

Here the batches are simply constructed by putting together enough root documents and their matches (we call this a session) to fill the batch according to batch_size parameter. An example of a batch of size batch_size=8 made of two sessions is show on the image below.

../../../_images/session-dataset.png

Examples

Toy example

Here is an example of a session dataset

from jina import Document, DocumentArray

from finetuner.tuner.dataset import SessionDataset, SessionSampler

ds = DocumentArray(
    [Document(content='shirt'), Document(content='shoe')]
)

ds[0].matches = [
    Document(content='red shirt', tags={'finetuner_label': 1}),
    Document(content='red shoe', tags={'finetuner_label': -1}),
]
ds[1].matches = [
    Document(content='black shoe', tags={'finetuner_label': 1}),
    Document(content='black pants', tags={'finetuner_label': -1}),
]

sds = SessionDataset(ds)
for b in SessionSampler(sds.labels, batch_size=2):
    print([sds[bb] for bb in b])

[('shirt', (0, 0)), ('red shoe', (0, -1))]
[('shoe', (1, 0)), ('black pants', (1, -1))]

We got 2 batches here.

Covid QA data

Covid QA data is a CSV that has 481 rows with the columns question, answer & wrong_answer.

../../../_images/covid-qa-data.png

To convert this dataset into a SessionDataset, we build each Document to contain the following relevant information:

  • .text: the question column

  • .matches: the generated positive & negative matches Document

    • .text: the answer/wrong_answer column

    • .tags['finetuner_label']: the match label: 1 or -1.

Matches are built with the logic below:

  • only allows 1 positive match per Document, it is taken from the answer column;

  • always include wrong_answer column as the negative match. Then sample other documents’ answer as negative matches.