I am still very new to Retrieve and Rank, and Document Conversion services, so I have been playing around with that lately.
I encountered a problem where when I upload a large document (100+ pages) - Retrieve and Rank would help me automatically break it up into answer units, which is great and helpful.
However, some questions only require ONE small line in the big chunks of answer units, is there a way that I can manually break further down the answer units that Retrieve and Rank service has provided me?
I heard that you can do it through JavaScript, but is there a way to do it through the UI?
I am contemplating to manually break up the huge doc into multiple smaller documents, but that could potentially lead to 100s of them - which is probably the last option that I'd resort to.
Any help or suggestions is greatly appreciated!
Thank you all!
First off, one clarification:
Retrieve and Rank does not break up your documents into answer units. That is something that the Document Conversion Service does when your conversion target is ANSWER_UNITS.
Regarding your question:
I don't fully understand exactly what you're trying to do, but if the answer units that are produced by default don't meet your requirements, you can customize different steps of the conversion process to adjust the produced answer units. Take a look at the documentation here.
Specifically, you want to make sure that the heading levels (for Word, PDF or HTML, depending on your document type) are defined in a way that
they detect the start of each answer unit. Then, make sure that the heading levels that you defined (h1, h2, h3, etc.) are included in the selector_tags list within the answer_units section.
Once your custom Document Conversion Service configuration produces the answer units you are looking for, you will be ready to send them to Retrieve and Rank to be indexed.
Related
Context:
I have a set of documents in SharePoint. I have set up Azure Cognitive Search (Standard tier) with data sources (SharePoint), index and indexers. I have also added a semantic configuration.
Outcome:
Ask a question, and have the search find and return relevant sections from the documents. I will use these sections to feed into OpenAI to construct a cohesive result.
I would like to replicate this Microsoft demo: https://www.youtube.com/watch?v=3t3qZu1Dy1k&t=572s It seems to me to create this 'demo' each document content is very small and they could easily be combined to pass into OpenAI.
My experience so far:
The results return the documents and rank them, which seems OK - however it returns a short 'caption' and the full text. The caption is not necessarily related to my question - and can therefore not be used for the next step. The full document is far too big to be used in OpenAI.
I have managed to get Semantic answers - however the question has to be so precise to get a result, and the associated text is limited.
What I would like:
I would like the search to return sub-sections of the document, where the results of my question may be. If that is not supported, I feel I need an entirely new approach.
Any ideas? Thanks in advance for your time.
The demo you refer to works by feeding documents to Azure Cognitive Search. A query is then formulated as a question that uses the Semantic Search functionality to return a set of potential semantic answers extracted from the content in the index.
These potential semantic answers are then fed as a prompt to OpenAI's text completion service: https://beta.openai.com/docs/guides/completion
First, you must ensure you can get good semantic answers. Inspect the content you have indexed and verify that it contains content that could semantically be an answer to the questions you test with. Good content should have declarations of facts. I.e., statements that could be used verbatim as an answer to a question. Examples:
The capital of France is Paris.
Forecast for 2022 is expected to be 22%.
The semantic functionality in Azure Search will only respond with a text section containing a potential answer to your question. If you can't get this step to work, you have to work on improving that. Either via semantic configuration, choice of content, or by making sure you process your content so that the items in your index contain the relevant content in the correct properties.
Ensure your content is indexed and mapped to properties in a sensible way
Work with the semantic configuration until you get sensible results
Once the previous two steps are ok, submit to OpenAI
I have tested the semantic text on two different data sets. Both were a combination of website content, PDF- and Word documents, etc. The topic and volume of content were essentially the same. From one data set, I could get excellent semantic answers. But, the other data set was disappointing.
My conclusion was that the content in the good data set was formulated and structured in a way that fits a semantic scenario. The other data set would often have logic and meaning presented in tables and layouts. As a human reading the content on paper, you would understand it. But, semantically, it would not make as much sense.
We regularly need to perform a handful of relatively simple tests against a bunch of MS Word documents. As these checks are currently done manually, I am striving for a way to automate this. For example:
Check if every page actually has a page number and verify that it is correct.
Verify that a version identifier in the page header is identical across all pages.
Check if the document has a table of contents.
Check if the document has a table of figures.
Check if every figure has a caption.
et cetera. Is this reasonably feasible using PowerShell in conjunction with a Word API?
Powershell can access Word via its object model/Interop (on Windows, at any rate) and AIUI can also work with the Office Open XML OOXML) API, so really you should be able to write any checks you want on the document content. What is slightly less obvious is how you verify that the document content will result in a particular "printed appearance". I'm going to start with some comments on the details first.
Just bear in mind that in the following notes I'm just pointing out a few things that you might have to deal with. If you're examining documents produced by an organisation where people are already broadly speaking following the same standards, it may be easier.
Of the 5 examples you give, without checking the details I couldn't say exactly how you would do them, and there could be difficulties with all of them, but for example
Check if every page actually has a page number and verify that it is correct.
Difficult using either OOXML or the object model, because what you would really be checking is that the header for a particular section had a visible { PAGE } field code. Because that field code might be nested inside other fields that say "if don't display this field code", it's not so easy to be sure that there would be a page number.
Which is what I mean by checking the document's "printed appearance" - if, for example, you can use the object model to print to PDF and have some mechanism that lets PS inspect the PDF's content, that might be a better approach.
Verify that a version identifier in the page header is identical across all pages.
Similar problem to the above, IMO. It depends partly on how the version identifier might be inserted. Is it just a piece of text? Could it be constructed from a number of fields? Might it reference Document Properties or Variables, or Custom XML content?
Check if the document has a table of contents.
Perhaps enough to look for a TOC field that does not have certain options, such as a \c option that a Table of Figures would contain.
Check if the document has a table of figures.
Perhaps enough to check for a TOC field that does have a \c option, perhaps with a specific parameter such as "Figure"
Check if every figure has a caption.
Not sure that you can tell whether a particular image is "a Figure". But if you mean "verify that every graphic object has a caption", you could probably iterate through the inline and floating graphics in the document and verify that there was something that looked like a Word standard caption paragraph within a certain distance of that object. Word has two standard field code patterns for captions AFAIK (one where the chapter number is included and one where it isn't), so you could look for those. You could measure a distance between the image and the caption by ensuring that they were no more than a predefined number of paragraphs apart, or in the case of a floating image, perhaps that the paragraph anchoring the image was no more than so many paragraphs away from the caption.
A couple of more general problems that you might have to deal with:
- just because a document contains a certain feature, such as a ToC field, does not mean that it is visible. A TOC field might have been formatted as not visible. Even harder to detect, it could have been formatted as colored white.
- change tracking. You might have to use the Word object model to "accept changes" before checking whether any given feature is actually there or not. Unless you can find existing code that would help you do that using the OOXML representation of the document, that's probably a strong case for doing checks via the object model.
Some final observations
for future checks, perhaps worth noting that in principle you could create a "DocumentInspector" that users could call from Word BackStage to perform checks on a document. Not sure you can force users to run it, or that you could create it in PS, but perhaps a useful tool.
longer term, if you are doing a very large number of checks, perhaps worth considering whether you could train a ML model to try to detect problems.
I have a unique problem and I'm not aware of any algorithm that can help me. Maybe someone on here does.
I have a dataset compiled from many different sources (teams). One field in particular is called "type". Here are some example values for type:
aple, apples, appls, ornge, fruits, orange, orange z, pear,
cauliflower, colifower, brocli, brocoli, leeks, veg, vegetables.
What I would like to be able to do is to group them together into e.g. fruits, vegetables, etc.
Put another way I have multiple spellings of various permutations of a parent level variable (fruits or vegetables in this example) and I need to be able to group them as best I can.
The only other potentially relevant feature of the data is the team that entered it, assuming some consistency in the way each team enters their data.
So, I have several million records of multiple spellings and short spellings (e.g. apple, appls) and I want to group them together in some way. In this example by fruits and vegetables.
Clustering would be challenging since each entry is most often 1 or two words, making it tricky to calculate a distance between terms.
Short of creating a massive lookup table created by a human (not likely with millions of rows), is there any approach I can take with this problem?
You will need to first solve the spelling problem, unless you have Google scale data that could allow you to learn fixing spelling with Google scale statistics.
Then you will still have the problem that "Apple" could be a fruit or a computer. Apple and "Granny Smith" will be completely different. You best guess at this second stage is something like word2vec trained on massive data. Then you get high dimensional word vectors, and can finally try to solve the clustering challenge, if you ever get that far with decent results. Good luck.
I am working with some really large databases of newspaper articles, I have them in a MySQL database, and I can query them all.
I am now searching for ways to help me tag these articles with somewhat descriptive tags.
All these articles is accessible from a URL that looks like this:
http://web.site/CATEGORY/this-is-the-title-slug
So at least I can use the category to figure what type of content that we are working with. However, I also want to tag based on the article-text.
My initial approach was doing this:
Get all articles
Get all words, remove all punctuation, split by space, and count them by occurrence
Analyze them, and filter common non-descriptive words out like "them", "I", "this", "these", "their" etc.
When all the common words was filtered out, the only thing left is words that is tag-worthy.
But this turned out to be a rather manual task, and not a very pretty or helpful approach.
This also suffered from the problem of words or names that are split by space, for example if 1.000 articles contains the name "John Doe", and 1.000 articles contains the name of "John Hanson", I would only get the word "John" out of it, not his first name, and last name.
Automatically tagging articles is really a research problem and you can spend a lot of time re-inventing the wheel when others have already done much of the work. I'd advise using one of the existing natural language processing toolkits like NLTK.
To get started, I would suggest looking at implementing a proper Tokeniser (much better than splitting by whitespace), and then take a look at Chunking and Stemming algorithms.
You might also want to count frequencies for n-grams, i.e. a sequences of words, instead of individual words. This would take care of "words split by a space". Toolkits like NLTK have functions in-built for this.
Finally, as you iteratively improve your algorithm, you might want to train on a random subset of the database and then try how the algorithm tags the remaining set of articles to see how well it works.
You should use a metric such as tf-idf to get the tags out:
Count the frequency of each term per document. This is the term frequency, tf(t, D). The more often a term occurs in the document D, the more important it is for D.
Count, per term, the number of documents the term appears in. This is the document frequency, df(t). The higher df, the less the term discriminates among your documents and the less interesting it is.
Divide tf by the log of df: tfidf(t, D) = tf(t, D) / log(df(D) + 1).
For each document, declare the top k terms by their tf-idf score to be the tags for that document.
Various implementations of tf-idf are available; for Java and .NET, there's Lucene, for Python there's scikits.learn.
If you want to do better than this, use language models. That requires some knowledge of probability theory.
Take a look at Kea. It's an open source tool for extracting keyphrases from text documents.
Your problem has also been discussed many times at http://metaoptimize.com/qa:
http://metaoptimize.com/qa/questions/1527/what-are-some-good-toolkits-to-get-lda-like-tagging-of-my-documents
http://metaoptimize.com/qa/questions/1060/tag-analysis-for-document-recommendation
If I understand your question correctly, you'd like to group the articles into similarity classes. For example, you might assign article 1 to 'Sports', article 2 to 'Politics', and so on. Or if your classes are much finer-grained, the same articles might be assigned to 'Dallas Mavericks' and 'GOP Presidential Race'.
This falls under the general category of 'clustering' algorithms. There are many possible choices of such algorithms, but this is an active area of research (meaning it is not a solved problem, and thus none of the algorithms are likely to perform quite as well as you'd like).
I'd recommend you look at Latent Direchlet Allocation (http://en.wikipedia.org/wiki/Latent_Dirichlet_allocation) or 'LDA'. I don't have personal experience with any of the LDA implementations available, so I can't recommend a specific system (perhaps others more knowledgeable than I might be able to recommend a user-friendly implementation).
You might also consider the agglomerative clustering implementations available in LingPipe (see http://alias-i.com/lingpipe/demos/tutorial/cluster/read-me.html), although I suspect an LDA implementation might prove somewhat more reliable.
A couple questions to consider while you're looking at clustering systems:
Do you want to allow fractional class membership - e.g. consider an article discussing the economic outlook and its potential effect on the presidential race; can that document belong partly to the 'economy' cluster and partly to the 'election' cluster? Some clustering algorithms allow partial class assignment and some do not
Do you want to create a set of classes manually (i.e., list out 'economy', 'sports', ...), or do you prefer to learn the set of classes from the data? Manual class labels may require more supervision (manual intervention), but if you choose to learn from the data, the 'labels' will likely not be meaningful to a human (e.g., class 1, class 2, etc.), and even the contents of the classes may not be terribly informative. That is, the learning algorithm will find similarities and cluster documents it considers similar, but the resulting clusters may not match your idea of what a 'good' class should contain.
Your approach seems sensible and there are two ways you can improve the tagging.
Use a known list of keywords/phrases for your tagging and if the count of the instances of this word/phrase is greater than a threshold (probably based on the length of the article) then include the tag.
Use a part of speech tagging algorithm to help reduce the article into a sensible set of phrases and use a sensible method to extract tags out of this. Once you have the articles reduced using such an algorithm, you would be able to identify some good candidate words/phrases to use in your keyword/phrase list for method 1.
If the content is an image or video, please check out the following blog article:
http://scottge.net/2015/06/30/automatic-image-and-video-tagging/
There are basically two approaches to automatically extract keywords from images and videos.
Multiple Instance Learning (MIL)
Deep Neural Networks (DNN), Recurrent Neural Networks (RNN), and the variants
In the above blog article, I list the latest research papers to illustrate the solutions. Some of them even include demo site and source code.
If the content is a large text document, please check out this blog article:
Best Key Phrase Extraction APIs in the Market
http://scottge.net/2015/06/13/best-key-phrase-extraction-apis-in-the-market/
Thanks, Scott
Assuming you have pre-defined set of tags, you can use the Elasticsearch Percolator API like this answer suggests:
Elasticsearch - use a "tags" index to discover all tags in a given string
Are you talking about the name-entity recognition ? if so, Anupam Jain is right. it;s research problem with using deep learning & CRF. In 2017, the name-entity recognition problem is force on semi-surprise learning technology.
The below link is related ner of paper:
http://ai2-website.s3.amazonaws.com/publications/semi-supervised-sequence.pdf
Also, The below link is key-phase extraction on twitter:
http://jkx.fudan.edu.cn/~qzhang/paper/keyphrase.emnlp2016.pdf
So I am working on a webservice to access our weather forecast data (10000 locations, 40 parameters each, hourly values for the next 14 days = about 130 million values).
So I read all about RESTful services and its ideology.
So I understand that an URL is adressing a ressource.
But what is a ressource in my case?
The common use case is that you want to get the data for a couple of parameters over a timespan at one or more location. So clearly giving every value its own URL is not pratical and would result in hundreds of requests. I have the feeling that my specific problem doesn't excactly fit into the RESTful pattern.
Update: To clarify: There are two usage patterns of the service. 1. Raw data; rows and rows of data for several locations and parameters.
Interpreted data; the raw data calculated into symbols (Suns & clouds, for example) and other parameters.
There is not one 'forecast'. Different clients have different needs for data.
The reason I think this doesn't fit into the REST-pattern is, that while I can actually have a 'forecast' ressource, I still have to submit a lot of request parameters. So a simple GET-request on a ressource doesn't work, I end up POSTing data all over the place.
So I am working on a webservice to access our weather forecast data (10000 locations, 40 parameters each, hourly values for the next 14 days = about 130 million values). ... But what is a ressource in my case?
That depends on the details of your problem domain. Simply having a large amount of data is not a good reason to avoid REST. There are smart ways and dumb ways to model and expose that data.
As you rightly see, your main goal at this point should be to understand what exactly a resource is. Knowing only enough about weather forecasting to follow the Weather Channel, I won't be much help here. It's for domain experts like yourself to make that call.
If you were to explain in a little more detail the major domain concepts you're working with, it might make it a little easier to give specific advice.
For example, one resource might be Forecast. When weatherpeople talk about Forecasts, what words keep coming up? When you think about breaking a forecast down into smaller elements, what words do you use to describe the pieces?
Do this process recursively, and you'll probably be able to make a list of important terms. Don't forget that these terms can describe things or actions. Think about what these terms really mean, what data you can use to model them, how they can be aggregated.
At this point you'll have the makings of something you can start building a RESTful system around - but not before.
Don't forget that a RESTful system is not a data dump wrapped in HTTP - it's a hypertext-driven system.
Also don't forget that media types are the point of contact between your server and its clients. A media type is only limited by your imagination and can model datasets of any size if you're clever about it. It can contain XML, JSON, YAML, binary elements such as a Bloom Filter, or whatever works for the problem.
Firstly, there is no once-and-for-all right answer.
Each valid url is something that makes sense to query, think of them as equivalents to providing query forms for people looking for your data - that might help you narrow down the scenarios.
It is a matter of personal taste and possibly the toolkit you use, as to what goes into the basic url path and what parameters are encoded. The debate is a bit like the XML debate over putting values in elements vs attributes. It is not always a rational or logically decided issue nor will everybody be kind in their comments on your decisions.
If you are using a backend like Rails, that implies certain conventions. Even if you're not using Rails, it makes sense to work in the same way unless you have a strong reason to change. That way, people writing clients to talk to Rails-based services will find yours easier to understand and it saves you on documentation time ;-)
Maybe you can use forecast as the ressource and go deeper to fine grained services with xlink.
Would it be possible to do something like this,Since you have so many parameters so i was thinking if somehow you can relate it to a mix of id / parameter combo to decrease the url size
/WeatherForeCastService//day/hour
www.weatherornot.com/today/days/x // (where x is number of days)
www.weatherornot.com/today/9am/hours/h // (where h is number of hours)