With reference to the Web page below, using the Retrieve & Rank service of IBM Bluemix, we are creating a bot that can respond to inquiries.
Question:
After learning the ranker once, based on the user's response to the inquiry, how can we construct a mechanism to continuously learn and improve response accuracy?
Assumption:
Because there was no API of R&R service to continuously learn from the inquiry response result of the user, tuning the GroundTruth file,
I suppose that it is necessary to periodically perform such a process as training the ranker again.
Tuning contents of assumed GT file:
If there is a new question, add a set of questions and answers
Increase or decrease the relevance score of responses if there is something that could not be answered well by existing question
(If bot answered incorrectly, lower the score, if there is a useful answer, raise the score)
In order to continuously learn, you will want to do the following:
capture new examples i.e. each user input and corresponding result
review those examples and create new ranker examples, adjust relevance scores, etc
add those new examples to the ranker
retrain the ranker using your new and existing examples
NOTE: Be sure to validate that new updates to the ranker data improves the overall system performance. k-fold validation is a great way to measure this.
All in all, learning is a continuous process that should be repeated indefinitely or until system performance is deemed suffcient.
Related
is there any possibility to get exact time spent on a certain level in a game via firebase analytics? Thank you so much đ
I tried to use logEvents.
The best way to do so would be measuring the time on the level within your codebase, then have a very dedicated event for level completion, in which you would pass the time spent on the level.
Let's get to details. I will use Kotlin as an example, but it should be obvious what I'm doing here and you can see more language examples here.
firebaseAnalytics.setUserProperty("user_id", userId)
firebaseAnalytics.logEvent("level_completed") {
param("name", levelName)
param("difficulty", difficulty)
param("subscription_status", subscriptionStatus)
param("minutes", minutesSpentOnLevel)
param("score", score)
}
Now see how I have a bunch of parameters with the event? These parameters are important since they will allow you to conduct a more thorough and robust analysis later on, answer more questions. Like, Hey, what is the most difficult level? Do people still have troubles on it when the game difficulty is lower? How many times has this level been rage-quit or lost (for that you'd likely need a level_started event). What about our paid players, are they having similar troubles on this level as well? How many people have ragequit the game on this level and never played again? That would likely be easier answer with sql at this point, taking the latest value of the level name for the level_started, grouped by the user_id. Or, you could also have levelName as a UserProperty as well as the EventProperty, then it would be somewhat trivial to answer in the default analytics interface.
Note that you're limited in the number of event parameters you can send per event. The total number of unique parameter names is limited too. As well as the number of unique event names you're allowed to have. In our case, the event name would be level_completed. See the limits here.
Because of those limitations, it's important to name your event properties in somewhat generic way so that you would be able to efficiently reuse them elsewhere. For this reason, I named minutes and not something like minutes_spent_on_the_level. You could then reuse this property to send the minutes the player spent actively playing, minutes the player spent idling, minutes the player spent on any info page, minutes they spent choosing their upgrades, etc. Same idea about having name property rather than level_name. Could as well be id.
You need to carefully and thoughtfully stuff your event with event properties. I normally have a wrapper around the firebase sdk, in which I would enrich events with dimensions that I always want to be there, like the user_id or subscription_status to not have to add them manually every time I send an event. I also usually have some more adequate logging there Firebase Analytics default logging is completely awful. I also have some sanitizing there, lowercasing all values unless I'm passing something case-sensitive like base64 values, making sure I don't have double spaces (so replacing \s+ with " " (space)), maybe also adding the user's local timestamp as another parameter. The latter is very helpful to indicate time-cheating users, especially if your game is an idler.
Good. We're halfway there :) Bear with me.
Now You need to go to firebase and register your eps (event parameters) into cds (custom dimensions and metrics). If you don't register your eps, they won't be counted towards the global cd limit count (it's about 50 custom dimensions and 50 custom metrics). You register the cds in the Custom Definitions section of FB.
Now you need to know whether this is a dimension or a metric, as well as the scope of your dimension. It's much easier than it sounds. The rule of thumb is: if you want to be able to run mathematical aggregation functions on your dimension, then it's a metric. Otherwise - it's a dimension. So:
firebaseAnalytics.setUserProperty("user_id", userId) <-- dimension
param("name", levelName) <-- dimension
param("difficulty", difficulty) <-- dimension (or can be a metric, depends)
param("subscription_status", subscriptionStatus) <-- dimension (can be a metric too, but even less likely)
param("minutes", minutesSpentOnLevel) <-- metric
param("score", score) <-- metric
Now another important thing to understand is the scope. Because Firebase and GA4 are still, essentially just in Beta being actively worked on, you only have user or hit scope for the dimensions and only hit for the metrics. The scope basically just indicates how the value persists. In my example, we only need the user_id as a user-scoped cd. Because user_id is the user-level dimension, it is set separately form the logEvent function. Although I suspect you can do it there too. Haven't tried tho.
Now, we're almost there.
Finally, you don't want to use Firebase to look at your data. It's horrible at data presentation. It's good at debugging though. Cuz that's what it was intended for initially. Because of how horrible it is, it's always advised to link it to GA4. Now GA4 will allow you to look at the Firebase values much more efficiently. Note that you will likely need to re-register your custom dimensions from Firebase in GA4. Because GA4 is capable of getting multiple data streams, of which firebase would be just one data source. But GA4's CDs limits are very close to Firebase's. Ok, let's be frank. GA4's data model is almost exactly copied from that of Firebase's. But GA4 has a much better analytics capabilities.
Good, you've moved to GA4. Now, GA4 is a very raw not-officially-beta product as well as Firebase Analytics. Because of that, it's advised to first change your data retention to 12 months and only use the explorer for analysis, pretty much ignoring the pre-generated reports. They are just not very reliable at this point.
Finally, you may find it easier to just use SQL to get your analysis done. For that, you can easily copy your data from GA4 to a sandbox instance of BQ. It's very easy to do.This is the best, most reliable known method of using GA4 at this moment. I mean, advanced analysts do the export into BQ, then ETL the data from BQ into a proper storage like Snowflake or even s3, or Aurora, or whatever you prefer and then on top of that, use a proper BI tool like Looker, PowerBI, Tableau, etc. A lot of people just stay in BQ though, it's fine. Lots of BI tools have BQ connectors, it's just BQ gets expensive quickly if you do a lot of analysis.
Whew, I hope you'll enjoy analyzing your game's data. Data-driven decisions rock in games. Well... They rock everywhere, to be honest.
I tried to went through numerous articles trying to understand what should be my first step to incorporate associative analysis (may be Market Basket analysis) into my system. They all go deep into implementation of algorithm but no one talked about how to store data in the first place.
I will really appreciate if someone can give me some start pointers or article links that I can begin with.
The first thing I want to implement is to track user clicks and provide suggestions based on tracked data.
E.g. User clicked on link A and subsequently on link B and link C. I can track this activity with some metadata associated (user, user organization, user role etc.)
I do not want it to be limited only to links. In future, I want to add number of similar usecases into the system and want to make it smart. E.g. If user set specific values for fields A and B, most likely he/she will set value <bla> for field C.
My system may generate several thousand such data points in a day (E.g. user clicks, field selection etc.).
Below are my questions:
How should I store my data? Go SQL or No SQL (I briefly looked into Mongo DB and it looked promising)
What tool should I use to perform the associative analysis? Are there any open source tools I can use?
It depend. Does your data suitable for NoSql databases? To answer this question it's better to read CAP Theorem and it's case studies: https://en.wikipedia.org/wiki/CAP_theorem or http://robertgreiner.com/2014/06/cap-theorem-explained/
. Some time you want Consistency(depending to your data) and Availability => so that it's better to use Relational Databases like Mysql(Try to read case studies and analyse your data to pick the best tools)
There is large number of open source libraries, but in my opinion it's better to first read some concepts and algorithms. Try searching for Apriori,ECLAT, FP-GROWTH Algorithms and get concepts of them. then you can pick a tool or write the code your self. Some usefull tools(depending to your programming language):
Python: https://github.com/asaini/Apriori, https://github.com/enaeseth/python-fp-growth, https://github.com/enaeseth/python-fp-growth/blob/master/fp_growth.py
PHP: https://github.com/sigidhanafi/fp-growth-php
JAVA: https://github.com/goodinges/FP-Growth-Java, http://www.philippe-fournier-viger.com/spmf/
Also you can use Spark: https://spark.apache.org/docs/1.1.1/mllib-guide.html
I'm currently working on my final year research project, which is an application which analyzes travel reviews found online, and give out a sentiment score for particular tourist attractions as a result, by conducting aspect level sentiment analysis.
I have a newly scraped dataset from a famous travel website which does not allow to use their API for research/academic purposes. (bummer)
My supervisor said that I might need to get this dataset annotated before using it for the aforementioned purpose. I am kind of confused as to what data annotation means in this context. Could someone please explain what exactly is happening when a dataset is annotated and how it helps in getting sentiment analysis done?
I was told that I might have to get two/three human annotators and get the data annotated to make it less biased. I'm on a tight schedule and I was wondering if there are any tools that can get it done for me? If so, what will be the impact of using such tools over human annotators? I would also like suggestions for such tools that you would recommend.
I would really appreciate a detailed explanation to my questions, as I am stuck with my project progressing to the next step because of this.
Thank you in advance.
To a first approximation, machine learning algorithms (e.g., a sentiment analysis algorithm) is learning to perform a task that humans currently perform by collecting many examples of the human performing the task, and then imitating them. When your supervisor talks about "annotation," they're talking about collecting these examples of a human doing the sentiment annotation task: annotating a sentence for sentiment. That is, collecting pairs of sentences and their sentiment as judged by humans. Without this, there's nothing for the program to learn from, and you're stuck hoping the program can give you something from nothing -- which it never will.
That said, there are tools for collecting this sort of data, or at least helping. Amazon Mechanical Turk and other crowdsourcing platforms are good resources for this sort of data collection. You can also take a look at something like: http://www.crowdflower.com/type-sentiment-analysis.
Closed. This question does not meet Stack Overflow guidelines. It is not currently accepting answers.
We donât allow questions seeking recommendations for books, tools, software libraries, and more. You can edit the question so it can be answered with facts and citations.
Closed 2 years ago.
Improve this question
I am considering a project in which a publication's content is augmented by relevant, publicly available tweets from people in the area. But how could I programmatically find the relevant Tweets? I know that generating a structure representing the meaning of natural language is pretty much the holy grail of NLP, but perhaps there's some tool I can use to at least narrow it down a bit?
Alternatively, I could just use hashtags. But that requires more work on behalf of the users. I'm not super familiar with Twitter - do most people use hashtags (even for smaller scale issues), or would relying on them cut off a large segment of data?
I'd also be interested in grabbing Facebook statuses (with permission from the poster, of course), and hashtag use is pretty rare on Facebook.
I could use simple keyword search to crudely narrow the field, but that's more likely to require human intervention to determine which tweets should actually be posted alongside the content.
Ideas? Has this been done before?
There are two straightforward ways to go about finding tweets relevant to your content.
The first would be to treat this as a supervised document classification task, whereby you would train a classifier to annotate tweets with a certain predetermined set of topic labels. You could then use the labels to select tweets that are appropriate for whatever content you'll be augmenting. If you don't like using a predetermined set of topics, another approach would be to simply score tweets according to their semantic overlap with your content. You could then display the top n tweets with the most semantic overlap.
Supervised Document Classification
Using supervised document classification would require that you have a training set of tweets labeled with the set of topics you'll be using. e.g.,
tweet: NBA finals rocked label: sports
tweet: Googlers now allowed to use Ruby! label: programming
tweet: eating lunch label: other
If you want to collect training data without having to manually label the tweets with topics, you could use hashtags to assign topic labels to the tweets. The hashtags could be identical with the topic labels, or you could write rules to map tweets with certain hashtags to the desired label. For example, tweets tagged either #NFL or #NBA could all be assigned a label of sports.
Once you have the tweets labeled by topic, you can use any number of existing software packages to train a classifier that assigns labels to new tweets. A few available packages include:
NLTK (Python) - see Chapter 6 in the NLTK book on Learning to Classify Text
Classifier4J (Java)
nBayes (C#)
Semantic Overlap
Finding tweets using their semantic overlap with your content avoids the need for a labeled training set. The simplest way to estimate the semantic overlap between your content and the tweets that you're scoring is to use a vector space model. To do this, represent your document and each tweet as a vector with each dimension in the vector corresponding to a word. The value assigned to each vector position then represents how important that word is to the meaning of document. One way to estimate this would be to simply use the number of times the word occurs in the document. However, you'll likely get better results by using something like TF/IDF, which up-weights rare terms and down-weights more common ones.
Once you've represented your content and the tweets as vectors, you can score the tweets by their semantic similarity to your content by taking the cosine similarity of the vector for your content and the vector for each tweet.
There's no need to code any of this yourself. You can just use a package like Classifier4J, which includes a VectorClassifier class that scores document similarity using a vector space model.
Better Semantic Overlap
One problem you might run into with vector space models that use one term per dimension is that they don't do a good job of handling different words that mean roughly the same thing. For example, such a model would say that there is no similarity between The small automobile and A little car.
There are more sophisticated modeling frameworks such as latent semantic analysis (LSA) and latent dirichlet allocation (LDA) that can be used to construct more abstract representations of the documents being compared to each other. Such models can be thought of as scoring documents not based on simple word overlap, but rather in terms of overlap in the underlying meaning of the words.
In terms of software, the package Semantic Vectors provides a scalable LSA-like framework for document similarity. For LDA, you could use David Blei's implementation or the Stanford Topic Modeling Toolbox.
Great question. I think for twitter your best bet is to use hashtags because otherwise you need to create algorithms or find existing algorithms that do language analysis and improve over time based on user input/feedback.
For facebook you can kind of do what bing implemented a while back. As I covered in this article here:
http://www.socialtimes.com/2010/06/bing-adds-facebook-and-twitter-features-steps-up-social-services/
I wrote: For example, a search for âNBA Finalsâ will return fan-page content from Facebook, including posts from a local TV station. So if you're trying to augmented NBA related content, you could do a similar search as Bing provides - searching publically available fan-page content the way spiders index them for search engines. I'm not a developer so i'm not sure of the intricacies but I know it can be done.
Also you can display popular shared links from users who are publishing to âeveryoneâ will be aggregated for all non-fan page content. I'm not sure if this is limited to being published to 'everyone' and/or being 'popular' although I would assume so - but you can double check that.
Hope this helps
The problem with NLP is not the algorithm (although that is an issue) the problem is the resources. There are some open source shallow parsing tools (that's all you would need to get intent) that you could use but parsing thousands or millions of tweets would cost a fortune in computer time.
On the other hand like you said not all tweets have hashtags and there is no promise they will be relevant.
Maybe you can use a mixture of keyword search to filter out a few possibilities (those with the highest keyword density) and then use a deeper data analysis to pick the top 1 or 2. This would keep computer resources at a minimum and you should be able to get relevant tweets.
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.