At Cresta, we are democratizing expertise for sales and support teams by making every agent an expert. To distill such expertise into software, we ask top agents to demonstrate, and in turn, help us label best practices. Machine Learning models are then trained for customers to maximize their KPIs. Our models continuously learn what top agents do differently and scale those behaviors across entire teams.
Apart from providing goal-directed suggestions during ongoing live chats which we talked about in our recent Action Directed GPT-2 blogpost, another unique feature that Cresta offers is real-time coaching assist. As shown below, Cresta provides personalized coaching at key moments in a live chat, to inculcate the required behaviors for every agent to perform like a top agent.
The Real-time Coaching and Agent Assist features mentioned above are powered by our Natural Language Understanding (NLU) pipeline, which is responsible for producing models that help us understand and track the state of the conversation as a chat progresses between an agent and a visitor. The 2 most common tasks which our NLU pipeline solves for, are:
In 2019, as our customer base started to rapidly grow, one of the biggest challenges we faced was the time and effort required to label data required by our NLU Classification pipeline. To scale as a software company, we strive to maximize our speed of developing and iterating on the required models. In this blog post, we share how our classification pipeline evolved over time and how we reduced our labeling cost and efforts by over 10x, while continuously pushing our accuracy benchmarks forward.
As was the case for most NLP pipelines across the world in 2019, the first big jump in efficiency came with the introduction of Deep Transfer Learning. Transfer learning, in the form of pre-trained language models, has revolutionized the field of NLP, leading to state-of-the-art results on a wide range of tasks. The idea is to first pre-train a model on a large unlabeled dataset using a language modeling objective, and then fine-tune it on a smaller labeled dataset using a supervised task of choice.
Many practical applications of NLP occur in scenarios where there is a scarcity of labeled data. This is where fine-tuning large pre-trained language models has changed the game completely. These models have shown to be extremely sample-efficient, capable of achieving good performance on many tasks with only a few hundred labeled samples.
At the start of 2019, with multi-task learning in NLP increasingly showing great empirical results, we deployed a multi-headed BERT for all the different tasks in our classification pipeline. As shown in the image below, this enabled us to train all of them together with a shared BERT encoder, maximizing each head’s learning from all the available labeled data. As a result, the final outcome was us reducing the number of labeled samples required per customer while pushing our accuracy benchmarks to previously uncharted regions.
Buoyed by the success of the multi-head architecture, we turned our attention to a problem which was proving to be a costly step in our labeling process: handling a growing Label Taxonomy.
For a model to track the state of a conversation using the classifiers described above, not all messages necessarily belong to a class of interest. As depicted in the image below, this meant that the multi-class classifiers which were deployed had a “None” class, to account for any messages that didn’t fall under the existing set of classes of interest. Instead of starting with a minimal set and then iteratively making data-driven additions to the taxonomy for each customer, this multi-class classification problem formulation forced us to spend considerable time upfront determining the granularity and details of the required taxonomy, and carefully defining what constitutes the “None” class – else otherwise, any future iterations on the taxonomy meant re-visiting all the labeled samples under a large “None” class and updating labels as required.
In short, our workflow was highly resistant to any taxonomy changes, which invariably happened as we entered new verticals and iterated with new customers, causing a lot of re-labeling and label quality issues.
To address the above challenge of iterating on a growing label taxonomy, we converted the multi-class classification problem to a one-vs-all classification problem using a multi-head architecture. As shown in the image below, each head acted as a binary classifier (True/False) for one of the classes, and a False from all the heads implied the “None” class.
The above architecture gave us the flexibility of adding more classes as we iterated on the taxonomy required to produce the experience desired by our customers, without having having to re-label our existing dataset each time. This architecture could be used both for a single-task or in a multi-task setting by simply prepending the class name with the task name to create a unique identifier for each head.
Data labeling interfaces and best practices, in general, have been an under-researched area – as was touched upon by François Chollet’s recent tweet, which sparked a debate amongst the research community. Our experience while trying to scale Machine Learning for business use-cases, pushed us to consider data curation and labeling as any other research problem we were looking to solve.
Labeling cost has 2 dimensions – the number of labeled samples required and the average time required to “correctly” label a sample. We realized that the effort and cost required to reach a high quality labeled dataset was often turning out to be a costly step requiring multiple quality assurance iterations. With a much more flexible one-vs-all architecture, instead of just looking for ways to reduce the number of labeled samples required by our models, we started iterating on optimizing our labeling interface with the goal of reducing the difficulty of labeling a given sample.
Humans usually have a small attention span, and labeling often can be a very tedious and mundane task. We A/B tested a new labeling interface where labelers would be making a single binary decision at a time, True/False for a pair of (sample, class), determining whether the sample belongs to that class or not.
A labeler could pick a class they wanted to focus on and the interface would present a sample to be labeled in a binary fashion, accompanied by clear labeling guidelines and examples, as shown in the image above. This interface allowed labelers to think about one class at a time, resulting in a lower cognitive load for them, while also allowing us to scale and distribute the labeling tasks more efficiently among the labelers. Our results showed that this interface resulted in ~2x faster labeling, with fewer mistakes made by the labelers.
Integrating the Binary Labeling Interface with our one-vs-all architecture meant we had to solve 1 problem: there was no guarantee that for a given sample, all the classes would be labeled. More explicitly, given the large amount of unlabeled data we usually work with, the design choice of labeling one class per sample meant that it was highly likely that for a given labeled sample in our training set, we would not have a supervision signal for all the heads. To address this, we implemented Loss Masking, where for a given sample we masked the loss for all the heads we didn’t have a label for. As demonstrated in the image below, for each sample, the loss is only applied to heads for which we have a label in the training batch.
Next, we turned our attention towards pushing the boundary around how sample-efficient Deep Transfer Learning could be, by introducing Active Learning in the pipeline. Our goal was to explore what can be achieved both in terms of accuracy and the associated labeling cost when these large pre-trained language models are used in conjunction with Active Learning techniques.
Similar to how humans learn, giving a model the power to interactively query a human to obtain labels at certain data points – i.e. introducing human guidance at various intervals – can dramatically improve the learning process. This is the key idea behind Active Learning: that a machine learning algorithm can achieve greater accuracy with fewer training labels if it is allowed to choose the data from which it learns.
Active Learning is an iterative process, which can be described by the following steps
Active Learning by definition involves a periodic human intervention in the process – which meant that for the above workflow to function effectively, we needed a single interface where our data team could seamlessly label new data points and immediately study its effect on the model to minimize the time delay between iterations.
To achieve the above, we developed a single tool with the following features:
As shown in the above image, our goal was to automate as many steps as possible in the iterative workflow, maximizing the speed of iteration and minimizing the manual human effort required. The large pre-trained language models powering our classifiers are known to have a vast amount of world knowledge trapped inside them, allowing them to perform respectably on many tasks without any explicit supervision. Establishing the above automated workflow in a single unified tool allowed us to really utilize this existing knowledge efficiently by only labeling nuanced samples specific to the use-cases we are solving for, i.e. where human supervision is indeed required.
Over the course of the last couple of years, we have been able to reduce the number of labeled samples required for training a classifier from a few thousand to a few hundred per customer. The goal which our Machine Learning team has been striving to achieve is to bring that number down to double-digits, without losing the top of the class accuracy our models showcase in production. This has been our focus for the last few months.
The next big challenge which we have been trying to solve is “Bootstrapping for Active Learning” – finding that minimal representative subset of data to label, for the iterative Active Learning pipeline to really start shining. Stay tuned for more updates on this.
If you are interested in joining us in solving such problems, take a look at our careers page for open positions.
Shubham Gupta for continuous contributions to the Active Learning workflow automation.
Tim Shi for overseeing the various projects described in this blog post.
Jessica Zhao, Lars Mennen, Motoki Wu, Saurabh Misra, Shubham Gupta and Tim Shi for edits and reviews on this blog post.