Extract top N key words from a Word or PDF document

[grantjenkins]

I've got a flow that triggers when a new file is created in a SharePoint Library (will be either Word or PDF documents only - 95% of them will be PDF). What I'm trying to do it extract the top 50 (for example) key words out of the document. There could be several pages in the document as well as tables, images, etc. There is no particular format of the documents. I'm not looking for specific key words - but rather use AI to determine the top N key words. As an example, if I asked Copilot to give me the top 50 key words from a specific document, it would provide the list of key words it deems to be the most relevant.

I know a Word document under the hood is made up of several documents (zipped), and we can unzip to get the individual files, then use xpath expressions to extract out the text from the xml file. But looking for a cleaner approach using some of the AI connectors/actions available if possible.

I've been working with AI Builder, but still having trouble achieving this. Extracting from text has a limitation of 5000 characters and also need to cater for the raw PDF/Word format, etc. Any guidance, references, etc. would be much appreciated.

IMPORTANT: I'm not looking to use any third-party connectors for this. Premium Microsoft connectors are fine though.

[Jondoesflow]

What about creating a custom connector to Azure Open AI. 

pass the file their and just write a prompt to do what I needs to do

https://techcommunity.microsoft.com/t5/educator-developer-blog/enhancing-your-power-apps-with-ai-cre...

---------------------------------------------------------------------------


If I have helped you with an answer, please accept this post as a solution and thumbs up my reply. Thanks !

Blog - JonDoesFlow
Twitter - JonDoesFlow
YouTube - JonDoesFlow

[Chriddle]

Maybe a PA Cloud Flow is not the best way to do this, but I wanted to satisfy my curiosity how it performs.

With the given text (20497 words) it takes a bit more than one minute.

This is the (probably naive) result:

The file content is a PDF that I created from this text file via Word's PDF export.

Maybe the easiest way to handle Word files as input would be to convert them to PDF with the connector "Convert Word to PDF" (but I have not tested this) 

Text (Compose)

join(
	xpath(
		xml(json(concat('{"Root":', outputs('Recognize_text_in_an_image_or_a_PDF_document')?['body/responsev2'], '}'))),
		'//text/text()'
	),
	' '
)

Get and combine all texts from the connectors output

TextWithReplacements (Compose)

replace(
	replace(
		replace(
			replace(
				replace(
					replace(
						replace(
							replace(
								replace(
									replace(
										outputs('Text'),
										'.', ' '
									),
									'!', ' '
								),
								'?', ' '
							),
							',', ' '
						),
						'[', ' '
					),
					']', ' '
				),
				'(', ' '
			),
			')', ' '
		),
		'"', ' '	
	),
	'''', ' '
)

Replace all characters that act as word separators with spaces.

You will probably need to add some more.

NonWords (Compose)

Spoiler

[
  "",
  "a",
  "of",
  "the",
  "in",
  "and",
  "to",
  "is",
  "was",
  "it",
  "from",
  "for",
  "at",
  "had",
  "by",
  "have",
  "been",
  "which",
  "be",
  "that",
  "as",
  "on",
  "were",
  "not",
  "all",
  "before",
  "i",
  "this",
  "s",
  "out",
  "has",
  "could",
  "into",
  "them",
  "most",
  "up",
  "with",
  "one",
  "since",
  "first",
  "more",
  "again",
  "we",
  "their",
  "are",
  "its",
  "_",
  "itself",
  "between",
  "both",
  "they",
  "or",
  "you",
  "no",
  "but"
]

List of words that are not considered keywords. This certainly needs to be supplemented.

You'll probably find a suitable list for your language on the web.

FilteredWords (Filter array)

From

split(outputs('TextWithReplacements'), ' ')

Filter

outputs('NonWords')

does not contain

toLower(item())

FilteredWordsXml (Compose)

xml(json(concat('{"Root":{"Item":', body('FilteredWords'), '}}')))

Counts (Select)

From

union(
	xpath(
		xml(outputs('FilteredWordsXml')),
		'//Item/text()'
	),
	json('[]')
)

Map Word

item()

Map Count

xpath(
	outputs('FilteredWordsXml'),
	concat('count(//Item[.="', item(), '"])')
)

Result (Compose)

take(
	reverse(
		sort(
			body('Counts'),
			'Count'
		)
	),
	50
)

[grantjenkins]

Hi @Chriddle Thanks for the solution - appreciated, but unfortunately it won't work in our case since the PDFs are generated from our Word documents so you get the following output (not plain text).

We would also have documents in various languages (that I didn't mention before) including Chinese, Mongolian, Zulu, etc. so would need to handle those too (XML/XPath wouldn't work in those cases).

[grantjenkins]

@Jondoesflow Thanks - might be the way to go. I don't believe we have Chat GPT enabled (allowed) in our environment given the sensitive data that we have. Need to check where the data will be processed (needs to be completely within our tenant). I'll provide some feedback on how we go after chatting with our AI team.

[Chriddle]

As I wrote, I also generated the PDF from a Word document.

Doesn't the connector "Recognize text in an image or a PDF document" work with your PDF?

Why wouldn't this work with Chinese, etc.?

Of course, you'll need to find "non-word" lists (and word separators if necessary) for all of these languages, but that should be doable.

[grantjenkins]

I think if you create a Word document that includes text, tables, images, etc. (not just a text document) then save as PDF you won't get plain text coming from your Get file content. And we don't have scanned PDFs either. The connector Recognize text in an image or a PDF document fails for me.

[Chriddle]

Of course the file content is not plain text, it's a PDF.

I just created such a Word document, exported it to PDF and put its file content into Recognize text in an image or a PDF document. Work fine 😉

What's your issue? Is there an error?