Decision By Precision?

I2E, the flagship product at Linguamatics, is a text mining engine and so sits in the broad space of search tools such as grep, Ctrl-F, and even Google. In that world, evaluating "how good" or "how relevant", or the "correctness" of a set of search results is interesting for a number of reasons, including:

it may be hard to define what those terms mean, in general cases.
it may be possible to calculate some kind of metric on well-understood, small, data sets but less so at scale.
it may be possible to calculate some kind of metric for simple searches, but less so for complex ones.
on different occasions the person searching may have different intent and needs from the same search.

But today we'll concentrate on two standard metrics that can be easily defined and which have agreed definitions: precision (roughly "how useful the search results are") and recall (roughly "how complete the results are").

Imagine we want to test our search engine. We have a set of documents and we will search them for the single word "testing". The image below, from Wikipedia, shows how we could calculate the metrics.

There's a lot of information in there, let's unpack some of it:

The square represents the documents.
The solid grey circles are occurrences of the word "testing".
The clear grey circles are occurrences of other words.
The central black circle is the set of results from the search.
The term positive means that a word is in the results.
The term negative means that a word is not in the results.
The term true means that a word is classified correctly.
The term false means that a word is classified incorrectly.

Let's overlay some numbers to make it clear. We inspect the documents (outside of the SUT) and find that there are 1000 words of which 100 are occurrences of the word "testing" (these are the solid grey circles.)

We run our search using the SUT and get back 50 results (the central black circle). We inspect those results and find that 35 are the word "testing" (the true positives) and 15 are something else (the false positives - asserted to be correct, but in fact incorrect).

The pictographs at the bottom of the image give us the formulae we need: precision comes only from the set of results we can see, and in this case is 35/50 or 70%. Recall requires knowledge of the whole set of documents, and for us is 35/100 or 35%.

A striking difference but which is better? Can one be better? These things are metrics, so can they be gamed?

Well, if the search simply returned every word in the documents its recall would be 100/100, or 100%, but precision would be very low at 100/1000, or 10%, because precision takes the negative content in the search results into account.

So can you get 100% precision? You certainly can: have the search return only those results with an extremely high confidence of being correct. Imagine only one result is returned, and it's a good one, then precision is 1/1 or 100%. Sadly, recall in this case is 1/100 or 1%.

Which is very interesting, really, but what does it have to do with testing?

Good question; it's background for a rough and ready analogy that squirted out of a conversation at work this week, illustrating the appealing trap of simple confirmatory testing. Imagine that you run your system under test with nominated input, inspect what comes out, and check that against some acceptance criteria. Everything in the output meets the criteria. Brilliant! Job done? Or precision 100%?

Images: https://flic.kr/p/sqtWUT, Wikipedia

Meet Me Halfway?

The Association for Software Testing is crowd-sourcing a book, Navigating the World as a Context-Driven Tester , which aims to provide responses to common questions and statements about testing from a context-driven perspective . It's being edited by Lee Hawkins who is posing questions on Twitter , LinkedIn , Mastodon , Slack , and the AST mailing list and then collating the replies, focusing on practice over theory. I've decided to contribute by answering briefly, and without a lot of editing or crafting, by imagining that I'm speaking to someone in software development who's acting in good faith, cares about their work and mine, but doesn't have much visibility of what testing can be. Perhaps you'd like to join me? --00-- "Stop answering my questions with questions." Sure, I can do that. In return, please stop asking me questions so open to interpretation that any answ...

How do I Test AI?

Recently a few people have asked me how I test AI. I'm happy to share my experiences, but I frame the question more broadly, perhaps something like this: what kinds of things do I consider when testing systems with artificial intelligence components . I freestyled liberally the first time I answered but when the question came up again I thought I'd write a few bullets to help me remember key things. This post is the latest iteration of that list. Caveats: I'm not an expert; what you see below is a reminder of things to pick up on during conversations so it's quite minimal; it's also messy; it's absolutely not a guide or a set of best practices; each point should be applied in context; the categories are very rough; it's certainly not complete. Also note that I work with teams who really know what they're doing on the domain, tech, and medical safety fronts and some of the things listed here are things they'd typically do some or all of. Testing ...

The Best Programmer Dan Knows

I was pairing with my friend Vernon at work last week, on a tool I've been developing. He was smiling broadly as I talked him through what I'd done because we've been here before. The tool facilitates a task that's time-consuming, inefficient, error-prone, tiresome, and important to get right. Vern knows that those kinds of factors trigger me to change or build something, and that's why he was struggling not to laugh out loud. He held himself together and asked a bunch of sensible questions about the need, the desired outcome, and the approach I'd taken. Then he mentioned a talk by Daniel Terhorst-North, called The Best Programmer I Know, and said that much of it paralleled what he sees me doing. It was my turn to laugh then, because I am not a good programmer, and I thought he knew that already. What I do accept, though, is that I am focussed on the value that programs can give, and getting some of that value as early as possible. He sent me a link to the ta...

Notes on Testing Notes

Ben Dowen pinged me and others on Twitter last week , asking for "a nice concise resource to link to for a blog post - about taking good Testing notes." I didn't have one so I thought I'd write a few words on how I'm doing it at the moment for my work at Ada Health, alongside Ben. You may have read previously that I use a script to upload Markdown-based text files to Confluence . Here's the template that I start from: # Date + Title # Mission # Summary WIP! # Notes Then I fill out what I plan to do. The Mission can be as high or low level as I want it to be. Sometimes, if deeper context might be valuable I'll add a Background subsection to it. I don't fill in the Summary section until the end. It's a high-level overview of what I did, what I found, risks identified, value provided, and so on. Between the Mission and Summary I hope that a reader can see what I initially intended and what actually...

On Herding Cats

Last night I was at the Cambridge Tester meetup for a workshop on leadership. It was a two-parter with Drew Pontikis facilitating conversation about workplace scenarios followed by an AMA with a group of experienced managers. I can't come to work this week, my cat died. Drew opened by asking us what our first thoughts would be as managers on seeing that sentence. Naturally, sadness and sympathy, followed by a week ? for a cat ? and I only got a day for my gran! Then practicalities such as maybe there's company policy that covers that , and then the acknowledgement that it's contextual: perhaps this was a long-time emotional support animal . Having established that management decisions are a mixture of emotion, logic, and contingency Drew noted that most of us don't get training in management or leadership then split us into small groups and confronted us with three situations to talk through: Setting personal development goals for others. Dropping a clange...

Reasonable Doubt

In Your job is to deliver code you have proven to work Simon Willison writes: As software engineers we ... need to deliver code that works — and we need to include proof that it works as well. He is coming at this from the perspective of LLM-assisted coding, but most of what he says applies in general. I think this is a reasonable consise summary of his requirements for developers: Manual happy paths: get the system into an initial state, exercise the code, check that it has the desired effect on the state. Manual edge cases: no advice given, just a note that skill here is a sign of a senior engineer. Automated tests: should demonstrate the change like Manual happy paths but also fail if the change is reverted. He notes that, even though LLM tooling can write automated tests, it's humans who are accountable for the code and it's on us to "include evidence that it works as it should." Coincidentally, just the week before I read his post I told one of my...

Vanilla Flavour Testing

I have been pairing with a new developer colleague recently. In our last session he asked me "is this normal testing?" saying that he'd never seen anything like it anywhere else that he'd worked. We finished the task we were on and then chatted about his question for a few minutes. This is a short summary of what I said. I would describe myself as context-driven . I don't take the same approach to testing every time, except in a meta way. I try to understand the important questions, who they are important to, and what the constraints on the work are. With that knowledge I look for productive, pragmatic, ways to explore whatever we're looking at to uncover valuable information or find a way to move on. I write test notes as I work in a format that I have found to be useful to me, colleagues, and stakeholders. For me, the notes should clearly state the mission and give a tl;dr summary of the findings and I like them to be public while I'm working not just w...

Why Test, Test What, Then How?

There was a time when testing was all about the mnemonics . Well, we had no AI back then so thinking about our human craft and how to share what we had learned using our human intelligence with other engaged humans for later recall in their human heads seemed important. But this post isn't about dumping on AI. It isn't about mnemonics either even if WTTWTH does look like something that'd fit well into that ancient world. No, this post is just a snappier version of what I said to my team this week when I was walking through some testing I'd done the day before. It was concerned with a change to a particular turn in the dialog system we're building where multiple variables are in play, some populated by an external call to an LLM service. I wanted to make the point that the bulk of the testing work was done in the research I did and the spreadsheet I made, not in the interaction with our product. That spreadsheet was the result of me looking in our service's cod...

Open Your Mind

Jerome Groopman, in How Doctors Think , reviews ways in which doctors can make poor choices, identifies potential causes, and suggests some practices, for both doctor and patient, that can help to prevent them. I find this interesting for a couple of reasons: first, I work in the health space, although not in a therapeutic area and, second, I like to reflect on my own thought processes. I'll take three broad themes from Groopman's analysis: the business of healthcare and how that impacts a physician's ability to practice; the doctor-patient relationship and how that impacts the experience of both sides; and the cognitive failings that impact a correct and timely diagnosis for any given patient. Naturally, these overlap. The book is written from the perspective of the notoriously commercialised American medical system and is around 20 years old, so doubtless some of the details are different outside of the US and have changed since pu...

Better Isolate Than Never

How do you test a product with distributed logic and non-deterministic components? Well, it depends on the product and when the question is asked. This post is one answer, for now. A service I work on exposes an API for internal and external clients and calls multiple other services. It orchestrates a reasonably complex journey for a user, with their inputs and the responses of the services determining each step. Our customers can configure the service to provide a variant of the standard journey for their users. This gives them choices on the available steps, the order of the steps, and, importantly, some of the logic for taking particular steps. This configuration lives in a different repository to the service, not owned by us, and it can be changed and deployed independently from the service. As you can imagine, we test our service in multiple ways at different granularities using the usual tooling, including unit tests fo...

Hiccupps

Search This Blog