gsarti
commented
1 year ago Hi @seemavishal,
Thanks for reaching out! Do you have a specific use-case in mind for saliency heatmaps for the summarization task? I guess that you might want to use this to verify which parts of the inputs are attended by the model during summarization. In this case, I might suggest:
- Using the ContiguousSpanAggregator (
spans) class to aggregate source tokens at sentence level, to obtain a per-source-sentence importance score for every generated token that is more easily interpretable. - Prefer faster attribution methods like
saliencyandattentionto make the attribution quicker. While these methods might not always faithfully reflect model importance, they can be useful approximations if efficiency is required.
Here is an example applied to headline generation:
import inseq
model = "Michau/t5-base-en-generate-headline"
article = (
"Very early yesterday morning, the United States President Donald Trump reported he and his wife First Lady Melania Trump tested positive for COVID-19. "
"Officials said the Trumps' 14-year-old son Barron tested negative as did First Family and Senior Advisors Jared Kushner and Ivanka Trump. "
"Trump took to social media, posting at 12:54 am local time (0454 UTC) on Twitter, \"Tonight, [Melania] and I tested positive for COVID-19. "
"We will begin our quarantine and recovery process immediately. We will get through this TOGETHER!\" Yesterday afternoon Marine One landed on "
"the White House's South Lawn flying Trump to Walter Reed National Military Medical Center (WRNMMC) in Bethesda, Maryland. "
"Reports said both were showing \"mild symptoms\". Senior administration officials were tested as people were informed of the positive test. "
"Senior advisor Hope Hicks had tested positive on Thursday. Presidential physician Sean Conley issued a statement saying Trump has been given zinc, "
"vitamin D, Pepcid and a daily Aspirin. Conley also gave a single dose of the experimental polyclonal antibodies drug from Regeneron Pharmaceuticals. "
"According to official statements, Trump, now operating from the WRNMMC, is to continue performing his duties as president during a 14-day quarantine. "
"In the event of Trump becoming incapacitated, Vice President Mike Pence could take over the duties of president via the 25th Amendment of the US Constitution. "
"The Pence family all tested negative as of yesterday and there were no changes regarding Pence's campaign events"
)
model = inseq.load_model(model, "attention")
out = model.attribute(article)
# Use periods as end-of-sentence indices for aggregation
ends = [i + 1 for i, t in enumerate(out[0].source) if t.token == "."] + [len(out[0].source) - 1]
starts = [0] + [i + 1 for i, t in enumerate(out[0].source) if t.token == "."]
source_spans = list(zip(starts, ends))
out.aggregate("spans", source_spans=source_spans).show()Alternatively, you can also aggregate the generated tokens in the target side to obtain an aggregated per-source-sentence importance score for the whole generation:
# Added after the code above
target_spans=[(0, len(out[0].target) - 1)]
out.aggregate("spans", source_spans=source_spans, target_spans=target_spans).show()Hope it helps!
SVC04
How can i utilize inseq for the problem of text generation containing long piece of text.
Hello, Thank you for providing such a great library for interpretability of generation tasks. I am trying to use inseq to explain output of text summarization generated from hugging face transformer model. Since the article used for summarization is too long the saliency heatmap visualization would be too long and computationally expensive as well. Any suggestion on this that how can i utilize inseq for summarization interpretability.