Category: Content Tagging

What does it mean to be “best-in-class”?

We often refer to our models as “industry-leading” or “best-in-class,” but what does this actually mean in practice? How are we better than our competitors, and by how much? It is easy to throw these terms around, but we mean it — and we have the evidence to back it up. In this blog post, we’ll be walking through some of the benchmarks that we have run against similar products to show how our models outperform the competition.

Visual Moderation

First, let’s take a look at one of our oldest and most popular models: visual moderation. To compare our model to its major competitors, we ran a test set of NSFW, suggestive, and clean images through all models.

Visual moderation is a classification task — in other words, the model’s job is to classify each submitted image into one of several categories (in this case, NSFW or Clean). A popular and effective metric to measure performance in classification models is by looking at their precision and recall. Precision is the number of true positives (i.e., correctly identified NSFW images) over the number of predicted positives (images predicted to be NSFW). Recall is the number of true positives (correctly identified NSFW images) over the number of ground-truth positives (actual NSFW images). 

There is a tradeoff between the two. If you predict all images to be NSFW, you will have perfect recall — you caught all the NSFW images! — but horrible precision because you incorrectly classified many clean images as NSFW. The goal is to have both high recall and high precision, no matter what confidence threshold is used.

With our visual moderation models, we’ve achieved this. We plotted the results of our test as a precision/recall curve, showing that even at high recall we maintain high precision and vice versa while our competitors fall behind us.

The above plot is for NSFW content detection. Our precision at 90% recall is nearly perfect at 99.6%, which makes our error rate a whopping 45 times lower than Public Cloud C. Even Public Clouds A and B, which are closer to us in performance, have error rates 12.5 times higher and 22.5 times higher than ours respectively.

We also benchmarked our model for suggestive content detection, or content that is inappropriate but not as explicit as our NSFW category. Hive’s error rate remains far below the other models, resting at 6 times lower than Public Cloud A and 12 times lower than Public Cloud C. Public Cloud B did not offer a similar category and thus could not be compared.

We only ran our test on NSFW/explicit imagery more broadly because our competitors do not have equivalent classes to ours for other visual moderation classes such as drugs, gore, and terrorism. This makes comparisons difficult, though it also in itself speaks to the fact that we offer far more classes than many of our competitors. With more than 90 subclasses, our visual moderation model far exceeds its peers in terms of the granularity of our results — we don’t just have classes for NSFW, but also for nudity, underwear, cleavage, and other smaller categories that offer our customers a more more in-depth understanding of their content.

Text Moderation

We used precision/recall curves to compare our text moderation model as well. For this comparison, we charted our performance across eight different classes. Hive outperforms all peer models on every single one.

Hive’s error rate on sexual content is 4 times lower than its closest competitor, Public Cloud B. Our other two competitors for that class both have error rates 6 times higher. The threat class boasts similar metrics, with Hive’s error rate between 2 and 4 times lower than all its peers.

Hive’s model for hateful content detection is on par with our competitors, remaining slightly ahead on all thresholds. Our model for bullying content does the same, with an error rate 2 times lower than all comparable models.

Hive is one of few companies to offer text moderation for drugs and weapons, and our error rates here are also worth noting — our only competitor has an error rate 4 and 8 times higher than ours for drugs and weapons respectively.

Hive also offers the child exploitation class, one that few others provide. With this class, we achieve an error rate 8 times lower than our only other major competitor.

Audio Moderation

For Audio Moderation, we evaluate our model using word error rate (WER), which is the gold-standard metric for a speech recognition system. Word error rate is the number of errors divided by the total number of words transcribed, and a perfect word error rate is 0. As you can see, we achieve the best or near-best performance across a variety of languages.

We excel across the board, with the lowest word error rate on the majority of the languages offered. On Spanish in particular, our word error rate is more than 4 times lower than Public Cloud B.

For German and Italian we are very close behind Public Cloud C and remain better than all other competitors.

Optical Character Recognition (OCR)

To benchmark our OCR model, we calculated the F-score for our model as well as several of our competitors. F-score is the harmonic mean of a model’s precision and recall, combining both of them into one measurement. A perfect F-score is 1. When comparing general F-scores, Hive excels as shown below.

We also achieve best-in-class or near-best performance when comparing by language, as shown in the graphs below. With some languages, we excel by quite a large margin. For Chinese and Korean in particular, Hive’s F-score is more than twice all of its competitors. We fall slightly behind in Hindi, yet still perform significantly better than Public Cloud A.

Demographics

We evaluated our age prediction model by calculating mean error, or how far off our age predictions were from the truth. Since the test dataset we used is labeled using age ranges and not individual numbers, mean error is defined as the distance in years from the closest end of the correct age range (i.e., guessing 22 for someone in the range 25-30 is an error of 3 years). A perfect mean error is 0.

As you can see from this distribution, Hive has a significantly lower mean error rate in the three lowest age buckets (0-2, 3-9, and 10-19). In the age range 0-2, our mean error rate is 11 times less than Public Cloud A’s. For the range 3-9 and 10-19, that difference becomes 5 times greater and 3 times greater respectively — still quite a large margin. Hive also excels notably at the oldest age bucket (70+), where our mean error rate is nearly 7 times less than Public Cloud A’s.

For a broader analysis, we compared our overall mean error across all age buckets, as well as the accuracy of our gender predictions.

AutoML

One of the newest additions to our product suite, our AutoML platform allows you to train image classification, text classification, and fine-tune large language models with your own custom datasets. To evaluate the effectiveness of this tool, we used the same test set to train models both on our platform and on competitor’s platforms and measured the performance of the resulting model. 

For image classification, we used three different classification tasks to account for the fact that different tasks have different levels of inherent difficulty and thus may yield higher or lower performing models. We also used three different dataset sizes for each classification task in order to measure how well the AutoML platform is able to work with limited amounts of examples.

We compared the resulting models using balanced accuracy, which is the arithmetic mean of a model’s true positive rate and true negative rate. A perfect balanced accuracy is 100%.

As shown in the above tables, Hive achieves best or near-best accuracy across all sets. Our results are quite similar to Public Cloud B’s, pulling ahead on the product dataset. We fell to near-best performance on the smoking dataset, which is the most difficult of the three classification tasks. Even then, we remained within a few percentage points of the winner, Public Cloud B.

For text classification, we trained models for three different categories: sexual content, drugs, and bullying. The results are in the table below. Hive outperforms all competitors on all three categories using all dataset sizes.

Another important consideration when it comes to AutoML is training time. An AutoML tool could build accurate models, but if it takes an entire day to do so it still may not be a great solution. We compared the time it took to train Hive’s text classification tool for the drugs category, and found that our platform was able to train the model 10 times as fast as Private Company A and 32 times as fast as Public Cloud B. And for the smallest dataset size of 100 examples, we trained the model 18 times faster than Private Company A and 268 times faster than Public Cloud B. That’s a pretty significant speedup.

Measuring the performance of fine-tuned LLMs on our foundation model is a bit more complicated. Here we evaluate two different tasks: question answering and closed-domain classification. 

To measure performance on the question answering task, we used a metric called token accuracy. Token accuracy indicates how many tokens are the same between the model’s response and the expected response from the test set. A perfect token accuracy is 100%. As shown below, our token accuracy is higher than our competitors or around the same for all dataset sizes.

This is also true for the classification task, where maintained roughly the same performance as Public Cloud A across the various dataset sizes. Below are the full results of our comparison.

Final Thoughts

As illustrated throughout this in-depth look into the performance of our models, we truly earn the title “best-in-class.” We conduct these benchmarks not just to justify that title, but more so as part of our constant effort to make our models the best that they can be. Reviewing these analyses helps us to identify our strengths, yes, but also our weaknesses and where we can improve.

If you have any questions about any of the benchmarks we’ve discussed here or any other questions about our models, please don’t hesitate to reach out to us at sales@thehive.ai.

Hive was thrilled to have our CTO Dmitriy present at the Workshop on Multimodal Content Moderation during CVPR last week, where we provided an overview of a few important considerations when building machine learning models for classification tasks. What are the effects of data quantity and quality on model performance? Can we use synthetic data in the absence of real data? And after model training is done, how do we spot and address bias in the model’s performance?

Read on to learn some of the research that has made our models truly best-in-class.

The Importance of Quality Data

Data is, of course, a crucial component in machine learning. Without data, models would have no examples to learn from. It is widely accepted in the field that the more data you train a machine learning model with, the better. Similarly, the cleaner that data is, the better. This is fairly intuitive — the basic principle is true for human learners, too. The more examples to learn from, the easier it is to learn. And if those examples aren’t very good? Learning becomes more difficult.

But how important is good, clean data to building a good machine learning model? Good data is not always easy to come by. Is it better to use more data at the expense of having more noise? 

To investigate this, we trained a binary image classifier to detect NSFW content, varying the amount of data between 10 images and 100k images. We also varied the noise by flipping a percentage of the labels on between 0% and 50% of the data. We then plotted the balanced accuracy of the resulting models using the same test set. 

The result? It turns out that data quality is more important than we may think. It was clear that, as expected, accuracy was the best when the data was both as large as possible (100k examples) and as clean as possible (0% noise). From there, however, the table gets more interesting.

As seen above, the model trained with only 10k data and no noise performs better than the model trained with ten times as much data at 100k and 10% noise. The general trend appears to be similar — clean data matters very much, and it can quickly tank performance even when using the maximum amount of data. In other words, less data is sometimes preferable to more data if it is cleaner.

We wondered how this would change with a more detailed classification problem, so we built a new binary image classifier. This time, we trained the model to detect images of smoking, which is detecting signal from a small part of an image. 

The outcome, shown below, echoes the results from the NSFW model — clean data has a great impact on performance even with a very large dataset. But the quantity of data appears to be more important than it was in the NSFW model. While 5000 examples with no noise got around 90% balanced accuracy for the NSFW model, that same amount of noiseless data only got around 77% for the smoking classifier. The increase in performance, while still strongly tied to data quantity, was noticeably slower and only the largest datasets produced well-performing models.

It makes sense that quantity of data would be more important with a more difficult classification task. Data noise also remained a crucial factor for the models trained with more data — the 50k model with 10% noise performed about the same as the 100k model with 10% noise, illustrating once more that more data is not always better if it is still noisy.

Our general takeaways here are that while both data quality and quantity matter quite a bit, clean data is more important beyond a certain quantity threshold. This threshold is where performance increases begin to plateau as the data grows larger, yet noisy data continues to have significant effects on model quality. And as we saw by comparing the NSFW model and the smoking one, this quality threshold also changes depending on the difficulty of the classification task itself.

Training on Synthetic Data: Does it Help or Hurt?

So having lots of clean data is important, what can be done when good data is hard to find or costly to acquire? With the rise of AI image generation over the past few years, more and more companies have been experimenting with generated images to supplement visual datasets. Can this kind of synthetic data be used to train visual classification models that will eventually classify real data?

In order to try this out, we trained five different binary classification models to detect smoking. Three of the models were trained exclusively with real data (10k, 20k, and 40k examples respectively), one was a mix of real and synthetic images (10k real and 30k synthetic), and one was trained entirely on synthetic data (40k). Each datatest had an even split of 50% smoking and 50% nonsmoking examples. To evaluate the models, we used two balanced test sets: one with 4k real images and one with 4k of synthetic images. All synetic images were created using Stable Diffusion.

Looking at the precision and recall curves for the various models, we made an interesting discovery. Unsurprisingly, the largest of the entirely real datasets performed the best (40k). The one trained on 10k real images and 30k synthetic images performed significantly better than the one trained only on 10k real images.

These results suggest that while large amounts of real data are best, a mixture of synthetic and real data could in fact boost model performance when little data is available.

Keeping an Eye Out For Bias

After model training is finished, extensive testing must be done in order to make sure there aren’t any biases in the model results. Biases can come in the form of biases that exist in the real world and are thus often ingrained in real-world data, such as racial bias or gender bias, but can also come in the form of biases that occur in the data by coincidence. 

A great example of how unpredictable certain biases can be came recently during a model training for NSFW detection, where the model started flagging many pictures of computer keyboards as false positives. Upon closer investigation, this occurred because many of the NSFW pictures in our training data were photos of computers whose screens were displaying explicit content. Since the computer screens were the focus of these images, keyboards were also often included, leading to the false association that keyboards are an indicator of NSFW imagery.

In order to correct this bias, we added more non-NSFW keyboard examples to the training data. Improving this bias in this way not only helps the model by addressing the bias itself, but also boosts general model performance. Of course, addressing bias is even more critical when dealing with data that carries current or historical biases against minority groups, thereby perpetuating them by ingraining them into future technology. The importance of detecting and correcting these biases cannot be overstated, since leaving them unaddressed carries a significant amount of risk beyond simply calling a keyboard NSFW.

Regardless of the type of bias, it’s important to note that biases aren’t always readily apparent. The original model prior to addressing the bias had a balanced accuracy of 80%, which is high enough that the bias may not have been immediately noticeable since errors weren’t extremely frequent. The takeaway here is thus not just that bias correction matters, but that looking into potential biases is necessary even when you might not think they’re there.

Takeaways

Visual classification models are in many ways the heart of Hive — they were our main launching point into the space of content moderation and AI-powered APIs more broadly. We’re continuously searching for ways to keep improving these models as the research surrounding them grows and evolves. Conclusions like those discussed here — the importance of clean data, particularly when you have lots of it, the possible use of synthetic data when real data is lacking, and the need to find and correct all biases (don’t forget about the unexpected ones!) — greatly inform the way we build and maintain our products.

We’re excited to announce Hive’s new AutoML tool that provides customers with everything they need to train, evaluate, and deploy customized machine learning models. 

Our pre-trained models solve a wide range of use cases, but we will always be bounded by the number of models we can build. Now customers who find that their unique needs and moderation guidelines don’t quite match with any of our existing solutions can create their own, custom-built for their platform and easily accessible via API.

AutoML can be used to augment our current offerings or to create new models entirely. Want to flag a particular subject that doesn’t exist as a head in our Text Moderation API, or a certain symbol or action that isn’t part of our Visual Moderation? With AutoML, you can quickly build solutions for these problems that are already integrated with your Hive workflow.

Let’s walk through our AutoML process to illustrate how it works. In this example, we’ll build a text classification model that can determine whether or not a given news headline is satirical. 

First, we need to get our data in the proper format. For text classification models, all dataset files must be in CSV format. One column should contain the text data (titled text_data) and all other columns represent model heads (classification categories). The values within each row of any given column represent the classes (possible classifications) within that head. An example of this formatting for our satire model is shown below:

The first page you’ll see on Hive’s AutoML platform is a dashboard with all of your organization’s training projects. In the image below, you’ll see how the training and deployment status of old projects are displayed. To create our satire classifier, we’re going to make a new project by hitting the “Create New Project” button in the top right corner.

We’ll then be prompted to provide a name and description for the project as well as training data in the form of a CSV file. For test data, you can either upload a separate CSV file or choose to randomly split your training data into two files, one to be used for training and the other for testing. If you decide to split your data, you will be able to choose the percentage that you would like to split off.

After all of that is entered, we are ready to train! Beginning model training is as easy as hitting a single button. While your model trains, you can easily view its training status on the Training Projects page.

Once training is completed, your project page will show an analysis of the model’s performance. The boxes at the top allow you to decide if you want to look at this analysis for a particular class or overall. If you’re building a multi-headed model, you can choose which head you’d like to evaluate as well. We provide precision, recall, and balanced accuracy for all confidence thresholds as well as a PR curve. We also display a confusion matrix to show how many predictions were correct and incorrect per class.

Once you’re satisfied with your model’s performance, select the “Create Deployment” to launch the model. Similarly to model training, deployment will take a few moments. After model deployment is complete, you can view the deployment in your Hive customer dashboard, where you can access your API key, view current tasks, as well as access other information as you would with our pre-trained models.

We’re very excited to be adding AutoML to our offerings. The platform currently supports both text and image classification, and we’re working to add support for large language models next. If you’d like to learn more about our AutoML platform and other solutions we’re building, please feel free to reach out to sales@thehive.ai or contact us here.

Recently, image-based content featuring embedded text – such as memes, captioned images and GIFs, and screenshots of text – have exploded in popularity across many social platforms. These types of content can present unique challenges for automated moderation tools. Not only does embedded text need to be detected and ordered accurately, it also must be analyzed with contextual awareness and attention to semantic nuance. 

Emojis have historically been another obstacle for automated moderation. Thanks to native support across many devices and platforms, these characters have evolved into a new online lexicon for accentuating or replacing text. Many emojis have also developed connotations that are well-understood by humans but not directly related to the image itself, which can make it difficult for automated solutions to identify harmful or inappropriate text content.

To help platforms tackle these challenges, Hive offers optical character recognition (OCR)-based moderation as part of our content moderation suite. Our OCR models are optimized for the types of digitally-generated content that commonly appears on social platforms, enabling robust AI moderation on content forms that are widespread yet overlooked by other solutions. Our OCR moderation API combines competitive text detection and transcription capabilities with our best-in-class text moderation model (including emoji support) into a single response, making it easy for platforms to take real-time enforcement actions across these popular content formats. 

OCR Model for Text Recognition

Effective OCR moderation starts with training for accurate text detection and extraction. Hive’s OCR model is trained on a large, proprietary set of examples that optimizes for how text commonly appears within user-generated digital content. Hive has the largest distributed workforce for data labeling in the world, and we leaned on this capability to provide tens of millions of human annotations on these examples to build our model’s understanding. 

We recently conducted a head-to-head comparison of our OCR model against top public cloud solutions using a custom evaluation dataset sourced from social platforms. We were particularly interested in test examples that featured digitally-generated text – such as memes and captioned images – to capture how content commonly appears on social platforms and selected evaluation data accordingly. 

In this evaluation, we looked at end-to-end text recognition, which includes both text detection and text transcription. Here, Hive’s OCR model outperformed or was competitive with other models on both exact transcription and transcription allowing character-level errors. At 90% recall, Hive’s OCR model achieved a precision of 98%, while public cloud models ranged from ~88% to 97%, implying a similar or lower end-to-end error rate.

OCR Moderation: Language Support

We recognize that many platforms’ moderation needs extend beyond English-speaking users. Hive’s OCR model supports text recognition and transcription for many widely spoken languages with comparable performance, many of which are also supported by our text moderation solutions. Here’s an overview of our current language support:

Language	OCR Support?	Text Moderation Support?
English	Yes	Yes (Model)
Spanish	Yes	Yes (Model)
French	Yes	Yes (Model)
German	Yes	Yes (Model)
Mandarin	Yes	Yes (Pattern Match)
Russian	Yes	Yes (Pattern Match)
Portuguese	Yes	Yes (Model)
Arabic	Yes	Yes (Model)
Korean	Yes	Yes (Pattern Match)
Japanese	Yes	Yes (Pattern Match)
Hindi	Yes	Yes (Model)
Italian	Yes	Yes (Pattern Match)

Moderation of Detected Text

Hive’s OCR moderation solution goes beyond producing a transcript – we then apply our best-in-class text moderation model to understand the meaning of that speech in context (including any detected emojis). Our backend will automatically feed text detected in an image as an input to our text moderation model, making our model classifications on image-based text accessible with a single API call. Our text model is generally robust to misspellings and character substitutions, enabling high classification accuracies on text extracted via OCR even if errors occur in transcription. 

Hive’s text moderation model can classify extracted text across several sensitive or inappropriate categories, including sexuality, threats or descriptions of violence, bullying, and racism. 

Another critical use-case is moderating spam and doxxing: OCR moderation will quickly and accurately flag images containing emails, phone numbers, addresses and other personal identifiable information.  Finally, our text moderation model can also identify promotions such as soliciting services, asking for shares and follows, soliciting donations, or links to external sites. This gives platforms new tools to curate user experience and remove junk content. 

We understand that verbal communication is rarely black and white – context and linguistic nuance can have profound effects on how meaning and intent of words are perceived. To help navigate these gray areas, our text model responses supplement classifications with a score from benign (score = 0) to severe (score = 3), which can be used to adapt any necessary moderation actions to platforms’ individual needs and sensitivities. You can read more about our text models in previous blog posts or in our documentation.

Our currently supported moderation classes in each language are as follows:

Language	Classes
English	Sexual, Hate, Violence, Bullying
Spanish	Sexual, Hate
Portuguese	Sexual, Hate
French	Sexual
German	Sexual
Hindi	Sexual
Arabic	Sexual

Emoji Classification for Text Moderation

Emoji recognition is a unique feature of Hive’s OCR moderation model that opens up new possibilities for identifying harmful or harassing text-based content. Emojis can be particularly useful in moderation contexts because they can subtly (or not-so-subtly) alter how accompanying text is interpreted by the reader. Text that is otherwise innocuous can easily become inappropriate when accompanied by a particular emoji and vice-versa.

Hive OCR is able to detect and classify any emojis supported by Apple, Samsung, or Google devices. Our OCR model currently achieves a weighted accuracy of over 97% when classifying emojis. This enables our text moderation model to account for contextual meaning and connotations of emojis used in input text. 

To get a sense of our model’s understanding, let’s take a look at some examples of how use of emojis (or inclusion of text around emojis) changes our model predictions to align with human understanding. Each of these examples is from a real classification task submitted to our latest model release.

Here’s a basic example of how adding an emoji changes our model response from classifying as clean to classifying as sensitive.  Our models understand not only the verbal concept represented by the emoji, but what the emoji means semantically based on where it is located in the text. In this case, the bullying connotation of the “garbage” or “trash” emoji would be completely missed by an analysis of the text alone. 

Our model is similarly sensitive to changes in semantic meaning caused by substitutions of emojis for text.

In this case, our model catches the sexual connotation added by the eggplant emoji in place of the word “eggplant.” Again, the text alone without an emoji – “lemme see that !” – is completely clean.

In addition to understanding how emojis can alter the meaning of text, our model is also sensitive to how text can change implications of emojis themselves.

Here, adding the phrase “hey hotty” transforms an emoji usually used innocuously into a message with suggestive intent, and our model prediction changes accordingly.  

Finally, Hive’s OCR and text moderation models are trained to differentiate between each skin tone option for emojis in the “People” category and understand their implications in the context of accompanying text. We are currently exploring how the ability to differentiate between light and darker skin tones can enable new tools to identify hateful, racist, or exclusionary text content.

OCR Moderation: Final Thoughts

User preferences for online communication are constantly evolving in both medium and content, which can make it challenging for platforms to keep up with abusive users. Hive prides itself on identifying blindspots in existing moderation tools and developing robust AI solutions using high-quality training data tailored to these use-cases. We hope that this post has showcased what’s possible with our OCR moderation capabilities and given some insight into our future directions. 

Feel free to contact sales@thehive.ai if you are interested in adding OCR capabilities to your moderation suite, and please stay tuned as we announce new features and updates!