if we want to evaluate a classifier of NLP application with data that are annotated with two annotators, and they are not completely agreed on the annotation, how is the procedure?
That is, if we should compare the classifier output with just the portion of data that annotators agreed on? or just one of the annotator data? or the both of them separately and then compute the average?
Taking the majority vote between annotators is common. Throwing out disagreements is also done.
Here's a blog post on the subject:
Suppose we have a bunch of annotators and we don’t have perfect agreement on items. What do we do? Well, in practice, machine learning evals tend to either (1) throw away the examples without agreement (e.g., the RTE evals, some biocreative named entity evals, etc.), or (2) go with the majority label (everything else I know of). Either way, we are throwing away a huge amount of information by reducing the label to artificial certainty. You can see this pretty easily with simulations, and Raykar et al. showed it with real data.
What's right for you depends heavily on your data and how the annotators disagree; for starters, why not use only items they agree on and see what then compare the model to the ones they didn't agree on?
Related
I have this classification scenario below in which Im getting a very low F1, precision, recall and other metrics.
Target is multiclass (about ~200 classes) which is highly imbalanced
I only use company names as classifier (mostly 1-2 words which have max of 8 words), no other fields (like description, etc.)
Training data ~ 100k+ records
Preprocessing: numeric and special characters and stopwords removal
I have very low resources for processing (thats why when I try to use oversampling techniques like smote, distance_smote for multiclass, etc., I always get memory error)
Tried using different vectorization/embedding/tokenizer like word2vec, tfidf, fasttext, bert, roberta, etc. but to no avail
Tried using (and fine-tuning) different algorithms (networks, svm, trees, boosting, etc.) but also getting low scores.
I also did cost-sensitive learning (using class weights) but it only decreased my scores.
Tried all options that I know but scores are not increasing. Can you recommend other options here or do you think any part of the process that may be wrong/discarded? Thank you!
Distribution of target labels:
Sample observations
There is essentially no way to know that 'Exxon' is an oil company, and 'Apple' a computer company, and 'McDonalds' a fast-food chain, just from their company names.
Even if you have a list of every other company in the world, by name and type, that's not enough to make the deduction for these last 3. Only other outside info – like a few sentences about them, or other data – could classify them.
In fact, while company names sometimes describe their exact field-of-commerce, often they're totally arbitrary, as that gives them more freedom to range over many products/services, or create their own unique associations with the name (aka branding).
So I strongly suspect your (unshown) names & (unshown) labels are just too arbitrary for the data you're using to get very good at the task you're attempting.
Is there a real-world situation where someone will only have a company name – no other info, or research options – and benefit from correctly guessing the class? If so, more specifics about the situation might help generate more specific tactical recommendations. But mainly such recommendations will be: get richer data about the targets of the classification.
You might squeeze a little more out of vague trends in corporate naming via better preprocessing/feature-extraction. You may want to keep numbers, special-characters, & punctuation in some form, as they might include extra slight hints. Using subwords (character n-grams) might also reveal some shared word-roots used even in made-up names.
In the context of bounded model checking, one describes the system as a State Transition System and the properties that need to be checked.
When one needs to provide multiple system descriptions and properties to the Model Checker Tool, it can become tedious to write the property by hand. In my case, I use some temporal logic.
How does one automate the process of translating/parsing the system description and deriving verifiable properties from it (ideally, a set of Initial states, Transitions, Set of States).
For example, consider the Microwave Example given here Given such a system description, how can I arrive at the specifications in an efficient manner?
There is no such open source tool that I know of, that can do this. Any approaches in terms of ideas, theories are welcome.
You can't automatically derive LTL formulae from automata as you suggest, because automata are more expressive than LTL formulae.
That leaves you with mainly two options: 1. find a verification tool which accepts specifications that are directly expressed as automata (I'm not sure which ones do, but I suspect it is worth checking SPIN and NuSMV for this feature.), or 2. use a meta-specification language that makes the writing of specifications easier; for example, https://www.isp.uni-luebeck.de/salt (doi: 10.1007/11901433_41) or IEE1850/PSL. While PSL is more a language definition for tool-implementors, SALT already offers a web front-end that translates your input straight into LTL.
(By the way, I find your approach methodologically challenging though: you're not supposed to derive formulae from your model, but from your initial system description as it is this very model which you're going to verify. But I am not a 100% sure, if I understood this point in your question correctly.)
I think properties of a system, e.g. Microwave system, come from technical and common sense expectation and requirements, not the model. E.g. microwave is supposed to cook the food. But it is not supposed to cook with door open. Nevertheless a repository of typical LTL pattern can be useful to define properties. It also lists properties along with more familiar regex and automata properties.
If you certain you still want to translate automata to LTL automatically check
https://mathoverflow.net/questions/96963/translate-a-buchi-automaton-to-ltl
Kansas Specification Property Repository
http://patterns.projects.cs.ksu.edu/documentation/patterns.shtml
I read about how FoundationDB does its network testing/simulation here: http://www.slideshare.net/FoundationDB/deterministic-simulation-testing
I would like to implement something very similar, but cannot figure out how they actually did implement it. How would one go about writing, for example, a C++ class that does what they do. Is it possible to do the kind of simulation they do without doing any code generation (as they presumeably do)?
Also: How can a simulation be repeated, if it contains random events?? Each time the simulation would require to choose a new random value and thus be not the same run as the one before. Maybe I am missing something here...hope somebody can shed a bit of light on the matter.
You can find a little bit more detail in the talk that went along with those slides here: https://www.youtube.com/watch?v=4fFDFbi3toc
As for the determinism question, you're right that a simulation cannot be repeated exactly unless all possible sources of randomness and other non-determinism are carefully controlled. To that end:
(1) Generate all random numbers from a PRNG that you seed with a known value.
(2) Avoid any sort of branching or conditionals based on facts about the world which you don't control (e.g. the time of day, the load on the machine, etc.), or if you can't help that, then pseudo-randomly simulate those things too.
(3) Ensure that whatever mechanism you pick for concurrency has a mode in which it can guarantee a deterministic execution order.
Since it's easy to mess all those things up, you'll also want to have a way of checking whether determinism has been violated.
All of this is covered in greater detail in the talk that I linked above.
In the sims I've built the biggest issue with repeatability ends up being proper seed management (as per the previous answer). You want your simulations to give different results only when you supply a different seed to your random number generators than before.
After that the biggest issue I've seen seems tends to be making sure you don't iterate over collections with nondeterministic ordering. For instance, in Java, you'd use a LinkedHashMap instead of a HashMap.
I am curious what are the methods / approaches to overcome the "cold start" problem where when a new user or an item enters the system, due to lack of info about this new entity, making recommendation is a problem.
I can think of doing some prediction based recommendation (like gender, nationality and so on).
You can cold start a recommendation system.
There are two type of recommendation systems; collaborative filtering and content-based. Content based systems use meta data about the things you are recommending. The question is then what meta data is important? The second approach is collaborative filtering which doesn't care about the meta data, it just uses what people did or said about an item to make a recommendation. With collaborative filtering you don't have to worry about what terms in the meta data are important. In fact you don't need any meta data to make the recommendation. The problem with collaborative filtering is that you need data. Before you have enough data you can use content-based recommendations. You can provide recommendations that are based on both methods, and at the beginning have 100% content-based, then as you get more data start to mix in collaborative filtering based.
That is the method I have used in the past.
Another common technique is to treat the content-based portion as a simple search problem. You just put in meta data as the text or body of your document then index your documents. You can do this with Lucene & Solr without writing any code.
If you want to know how basic collaborative filtering works, check out Chapter 2 of "Programming Collective Intelligence" by Toby Segaran
Maybe there are times you just shouldn't make a recommendation? "Insufficient data" should qualify as one of those times.
I just don't see how prediction recommendations based on "gender, nationality and so on" will amount to more than stereotyping.
IIRC, places such as Amazon built up their databases for a while before rolling out recommendations. It's not the kind of thing you want to get wrong; there are lots of stories out there about inappropriate recommendations based on insufficient data.
Working on this problem myself, but this paper from microsoft on Boltzmann machines looks worthwhile: http://research.microsoft.com/pubs/81783/gunawardana09__unified_approac_build_hybrid_recom_system.pdf
This has been asked several times before (naturally, I cannot find those questions now :/, but the general conclusion was it's better to avoid such recommendations. In various parts of the worls same names belong to different sexes, and so on ...
Recommendations based on "similar users liked..." clearly must wait. You can give out coupons or other incentives to survey respondents if you are absolutely committed to doing predictions based on user similarity.
There are two other ways to cold-start a recommendation engine.
Build a model yourself.
Get your suppliers to fill in key information to a skeleton model. (Also may require $ incentives.)
Lots of potential pitfalls in all of these, which are too common sense to mention.
As you might expect, there is no free lunch here. But think about it this way: recommendation engines are not a business plan. They merely enhance the business plan.
There are three things needed to address the Cold-Start Problem:
The data must have been profiled such that you have many different features (with product data the term used for 'feature' is often 'classification facets'). If you don't properly profile data as it comes in the door, your recommendation engine will stay 'cold' as it has nothing with which to classify recommendations.
MOST IMPORTANT: You need a user-feedback loop with which users can review the recommendations the personalization engine's suggestions. For example, Yes/No button for 'Was This Suggestion Helpful?' should queue a review of participants in one training dataset (i.e. the 'Recommend' training dataset) to another training dataset (i.e. DO NOT Recommend training dataset).
The model used for (Recommend/DO NOT Recommend) suggestions should never be considered to be a one-size-fits-all recommendation. In addition to classifying the product or service to suggest to a customer, how the firm classifies each specific customer matters too. If functioning properly, one should expect that customers with different features will get different suggestions for (Recommend/DO NOT Recommend) in a given situation. That would the 'personalization' part of personalization engines.
An initial draft of requirements specification has been completed and now it is time to take stock of requirements, review the specification. Part of this process is to make sure that there are no sizeable gaps in the specification. Needless to say that the gaps lead to highly inaccurate estimates, inevitable scope creep later in the project and ultimately to a death march.
What are the good, efficient techniques for pinpointing missing and implicit requirements?
This question is about practical techiniques, not general advice, principles or guidelines.
Missing requirements is anything crucial for completeness of the product or service but not thought of or forgotten about,
Implicit requirements are something that users or customers naturally assume is going to be a standard part of the software without having to be explicitly asked for.
I am happy to re-visit accepted answer, as long as someone submits better, more comprehensive solution.
Continued, frequent, frank, and two-way communication with the customer strikes me as the main 'technique' as far as I'm concerned.
It depends.
It depends on whether you're being paid to deliver what you said you'd deliver or to deliver high quality software to the client.
If the former, simply eliminate ambiguity from the specifications and then build what you agreed to. Try to stay away from anything not measurable (like "fast", "cool", "snappy", etc...).
If the latter, what Galwegian said + time or simply cut everything not absolutely drop-dead critical and build that as quickly as you can. Production has a remarkable way of illuminating what you missed in Analysis.
evaluate the lifecycle of the elements of the model with respect to a generic/overall model such as
acquisition --> stewardship --> disposal
do you know where every entity comes from and how you're going to get it into your system?
do you know where every entity, once acquired, will reside, and for how long?
do you know what to do with each entity when it is no longer needed?
for a more fine-grained analysis of the lifecycle of the entities in the spec, make a CRUDE matrix for the major entities in the requirements; this is a matrix with the operations/applications as the rows and the entities as the columns. In each cell, put a C if the application Creates the entity, R for Reads, U for Updates, D for Deletes, or E for "Edits"; 'E' encompasses C,R,U, and D (most 'master table maintenance' apps will be Es). Then check each column for C,R,U, and D (or E); if one is missing (except E), figure out if it is needed. The rows and columns of the matrix can be rearranged (manually or using affinity analysis) to form cohesive groups of entities and applications which generally correspond to subsystems; this may assist with physical system distribution later.
It is also useful to add a "User" entity column to the CRUDE matrix and specify for each application (or feature or functional area or whatever you want to call the processing/behavioral aspects of the requirements) whether it takes Input from the user, produces Output for the user, or Interacts with the user (I use I, O, and N for this, and always make the User the first column). This helps identify where user-interfaces for data-entry and reports will be required.
the goal is to check the completeness of the specification; the techniques above are useful to check to see if the life-cycle of the entities are 'closed' with respect to the entities and applications identified
Here's how you find the missing requirements.
Break the requirements down into tiny little increments. Really small. Something that can be built in two weeks or less. You'll find a lot of gaps.
Prioritize those into what would be best to have first, what's next down to what doesn't really matter very much. You'll find that some of the gap-fillers didn't matter. You'll also find that some of the original "requirements" are merely desirable.
Debate the differences of opinion as to what's most important to the end users and why. Two users will have three opinions. You'll find that some users have no clue, and none of their "requirements" are required. You'll find that some people have no spine, and things they aren't brave enough to say out loud are "required".
Get a consensus on the top two or three only. Don't argue out every nuance. It isn't possible to envision software. It isn't possible for anyone to envision what software will be like and how they will use it. Most people's "requirements" are descriptions of how the struggle to work around the inadequate business processes they're stuck with today.
Build the highest-priority, most important part first. Give it to users.
GOTO 1 and repeat the process.
"Wait," you say, "What about the overall budget?" What about it? You can never know the overall budget. Do the following.
Look at each increment defined in step 1. Provide a price-per-increment. In priority order. That way someone can pick as much or as little as they want. There's no large, scary "Big Budgetary Estimate With A Lot Of Zeroes". It's all negotiable.
I have been using a modeling methodology called Behavior Engineering (bE) that uses the original specification text to create the resulting model when you have the model it is easier to identify missing or incomplete sections of the requirements.
I have used the methodolgy on about six projects so far ranging from less than a houndred requirements to over 1300 requirements. If you want to know more I would suggest going to www.behaviorengineering.org there some really good papers regarding the methodology.
The company I work for has created a tool to perform the modeling. The work rate to actually create the model is about 5 requirements for a novice and an expert about 13 requirements an hour. The cool thing about the methodolgy is you don't need to know really anything about the domain the specification is written for. Using just the user text such as nouns and verbs the modeller will find gaps in the model in a very short period of time.
I hope this helps
Michael Larsen
How about building a prototype?
While reading tons of literature about software requirements, I found these two interesting books:
Problem Frames: Analysing & Structuring Software Development Problems by Michael Jackson (not a singer! :-).
Practical Software Requirements: A Manual of Content and Style by Bendjamen Kovitz.
These two authors really stand out from the crowd because, in my humble opinion, they are making a really good attempt to turn development of requirements into a very systematic process - more like engineering than art or black magic. In particular, Michael Jackson's definition of what requirements really are - I think it is the cleanest and most precise that I've ever seen.
I wouldn't do a good service to these authors trying to describe their aproach in a short posting here. So I am not going to do that. But I will try to explain, why their approach seems to be extremely relevant to your question: it allows you to boil down most (not all, but most!) of you requirements development work to processing a bunch of check-lists* telling you what requirements you have to define to cover all important aspects of the entire customer's problem. In other words, this approach is supposed to minimize the risk of missing important requirements (including those that often remain implicit).
I know it may sound like magic, but it isn't. It still takes a substantial mental effort to come to those "magic" check-lists: you have to articulate the customer's problem first, then analyze it thoroughly, and finally dissect it into so-called "problem frames" (which come with those magic check-lists only when they closely match a few typical problem frames defined by authors). Like I said, this approach does not promise to make everything simple. But it definitely promises to make requirements development process as systematic as possible.
If requirements development in your current project is already quite far from the very beginning, it may not be feasible to try to apply the Problem Frames Approach at this point (although it greatly depends on how your current requirements are organized). Still, I highly recommend to read those two books - they contain a lot of wisdom that you may still be able to apply to the current project.
My last important notes about these books:
As far as I understand, Mr. Jackson is the original author of the idea of "problem frames". His book is quite academic and theoretical, but it is very, very readable and even entertaining.
Mr. Kovitz' book tries to demonstrate how Mr. Jackson ideas can be applied in real practice. It also contains tons of useful information on writing and organizing the actual requirements and requirements documents.
You can probably start from the Kovitz' book (and refer to Mr. Jackson's book only if you really need to dig deeper on the theoretical side). But I am sure that, at the end of the day, you should read both books, and you won't regret that. :-)
HTH...
I agree with Galwegian. The technique described is far more efficient than the "wait for customer to yell at us" approach.