Mitigating offensive search suggestions with deep learning 06-12

Image credit : Shyam's Imagination Library

The humble search bar is the window through which most Internet users experience the web. Deep
learning is set to enhance the capabilities of this simple tool such that search engines can now anticipate what the user is looking for whilst moderating offensive suggestions before the query is complete.

A lack of contextual and deeper understanding of the intent of search queries often leads to inappropriate or offensive suggestions. Striving for a safer and saner web, the Microsoft team began its research with deep learning techniques to help detect and automatically prune search suggestions.

Our paper titled “Convolutional Bi-Directional LSTM for Detecting Inappropriate Query Suggestions in Web Search” received the “Best Paper Award” at the recent Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD) 2017. It was picked amongst a record-breaking 458 submissions at the leading international conference on knowledge discovery and data mining. Winning the award has been a humbling experience for the team. It upholds our efforts to establish deep machine learning as a powerful weapon against online vitriol.

Here’s a brief overview of what we found:

The Challenge

Human beings intuitively detect offensive language and derogatory speech online. Algorithms, however, struggle to clearly distinguish toxic language from benign comments. Systems often misjudge the level of toxicity or hate in certain phrases that require a deeper understanding of cultural norms, slang and context.

The problem is usually rendered difficult due to unique challenges posed by search queries such as lack of sufficient context, natural language ambiguity and presence of spelling mistakes and variations. For example, Marvin Gaye’s classic hit ‘If I Should Die Tonight’ could be deemed offensive simply because of the inclusion of the phrase ‘should die tonight’. Similarly, algorithms frequently misclassify offensive words and phrases as ‘clean’ simply because they were misspelled or used euphemisms. For example, the phrase ‘shake and bake’ is both a registered trademark for a popular food brand and street code for preparing illegal drugs on the move.

Safeguarding the Online Experience

The impact of inappropriate language and offensive speech online cannot be overstated. Internet access is ubiquitous and users cover diverse age groups and cultural backgrounds. Making search queries, online communication, and instant messaging safe for children, minorities, and sensitive communities is essential to preserve the integrity of the digital world.

Inappropriate suggestions on search queries or offensive comments on news articles could cause significant harm to vulnerable groups such as children and marginalized communities. Unsuitable suggestions could tarnish the reputation of corporations, inadvertently help someone cause harm to themselves with risky information, or lead to legal complications with authorities and regulators. Problems such as intimidation, threats, cyber bullying, trolling, explicit and suggestive content, and racist overtones need to be curtailed to help keep the Internet open and safely accessible to everyone.

The Solution

Conventional solutions to this problem have typically involved using –

1. A manually curated list of patterns involving such offensive words, phrases and slangs or
2. Classical Machine Learning (ML) techniques which use various hand-crafted features (typically words etc.) for learning the intent classifier or
3. Standard off-the-shelf deep learning model architectures such as CNN, LSTMs or Bi-directional LSTMs (BLSTMs).

In our current work, we propose a novel deep learning architecture called, “Convolutional Bi-Directional LSTM (C-BiLSTM)” – which combines the strengths of Convolutional Neural Networks (CNNs) with Bi-directional LSTMs (BLSTMs). Given a query, C-BiLSTM uses a convolutional layer for extracting feature representations for each query word which is then fed as input to the BLSTM layer which captures the various sequential patterns in the entire query and outputs a richer representation encoding them. The query representation thus learnt passes through a deep, fully connected network which predicts the target class – whether it is offensive or clean. C-BiLSTM doesn’t rely on hand-crafted features, is trained end-to-end as a single model, and effectively captures both local features as well as their global semantics.

Applying the technique on 79041 unique real-world search queries along with their class labels (inappropriate/clean), revealed that this novel approach was significantly more effective than conventional models based on patterns, classical ML techniques using hand-crafted features. C-BiLSTM also outperformed standard off-the-shelf deep learning models such as CNN, LSTM and BLSTM when applied to the same dataset. Our final C-BiLSTM model achieves a precision of 0.9246, recall of 0.8251 and an overall F1 score of 0.8720.

Although the focus of the paper was detecting offensive terms in Query Auto Completion (QAC) in search engines, the technique can be applied to other online platforms as well. Comments on news articles can be cleaned up and inappropriate conversations can be flagged up for abuse. Trolling can be detected and a safe search experience can be enabled for children online.  This technique could also help make chatbots and autonomous virtual assistants more contextually aware, culturally sensitive, and dignified in their responses. More details about this technique and implementation could be found in the actual paper.

Final Thoughts

The deep learning technique detailed in this study could be a precursor to better tools that can fight online vitriol. When APIs based on this system are applied to social media platforms, email services, chat rooms, discussion forums, and search engines, the results will be parsed through multiple filters to ensure users are not exposed to offensive content.

Curtailing offensive language can transform the Internet, making it safely accessible to a wider audience. Making the web an open, safe and secure place for ideas and innovations is a core part of our mission to empower everyone, everywhere through the power of technology.

Tags best paper award C-BiLSTM deep learning machine learning PAKDD 2017 award query auto completion (QAC)  

View at the original source

ARM’s new processors are designed to power the machine-learning machines 06-06

ML plus AI, AR, and VR — ARM is taking on the full set of trendy initialisms

On the eve of Computex, Taiwan’s big showpiece event where PC makers roll out the latest and best implementations of Intel CPUs, mobile rival ARM is announcing its own big news with the unveiling of a new generation of ARM CPUs and GPUs. Official today, the ARM Cortex-A75 is the new flagship-tier mobile processor design, with a claimed 22 percent improvement in performance over the incumbent A73. It’s joined by the new Cortex-A55, which has the highest power efficiency of any mid-range CPU ARM’s ever designed, and the Mali-G72 graphics processor, which also comes with a 25 percent improvement in efficiency relative to its predecessor G71.

The efficiency improvements are evolutionary and predictable, but the revolutionary aspects of this new lineup relate to artificial intelligence: this is the first set of processing components designed specifically to tackle the challenges of onboard AI and machine learning. Plus, last year’s updates to improve performance in the power-hugry tasks of augmented and virtual reality are being extended and elaborated.

Before we dive into the detail of this year’s changes, it’s worth recapping what ARM does and why it’s important. This English company, now owned by Japan’s SoftBank, is responsible for designing the processor architecture of practically every mobile device — you’ll have heard of Qualcomm’s Snapdragon, Samsung’s Exynos, and Apple’s A-series of mobile chips, all of which are built using ARM’s instruction sets and based on ARM’s design blueprints. When we talk about the oncoming wave of mobile AI, mobile VR, and smartphones that can perform machine-learning tasks without sending them off to processor farms up in the cloud, developing the capabilities for those tasks starts with ARM. 

The new Cortex-A75 and A55 are the first Dynamiq CPUs from ARM. Dynamiq is the branding chosen to describe a much more flexible set of design options for silicon vendors like Qualcomm. Where previously ARM allowed for designs that paired a cluster of so-called big CPUs (from its A7x class) and a matched number of little CPUs (from the A5x series), the new design makes it possible to spec a single, mixed-up cluster composed of both big and little CPUs, to a maximum of eight. Thus, chip makers can now have, for example, seven little A55 cores and just one big A75 one: for a favorable mix of long battery life, cost efficiency, and a high ceiling of single-threaded performance when it’s called for.

"50x improvement in AI performance over the next three to five years"

A RM marketing chief John Ronco says he anticipates a "50x improvement in AI performance over the next three to five years thanks to better architecture, micro-architecture, and software optimizations." ARM’s Dynamiq changes include a redesigned memory subsystem and tweaks to how CPU caches work — which has led to a doubling of memory streaming performance on the A55 relative to the A53 preceding it. Given that the A53 has shipped on 1.7 billion devices over the past three years, it’s truly the A55 that will make the biggest difference in achieving Ronco’s ambitious forecast. In most applications, the new mid-range core will be 10 to 30 percent better than previously, offering up to 15 percent better power efficiency and 18 percent better single-thread performance.

But it’s the fact that the new chip designs will be 10 times more configurable, with up to 3,000 different configurations, that will allow chipmakers far greater flexibility to make the most of them by tailoring them to specific tasks.

Interestingly, ARM won’t just be powering machine learning with its new chips, it’ll benefit from ML too. The new designs benefit from an improved branch predictor that uses neural network algorithms to improve data prefetching and overall performance.

The Cortex-A75 makes double-figure performance improvements across the board, with ARM claiming it’s on average 22 percent better than the A73, with 16 percent higher memory throughput, and a 34 percent improvement in its Geekbench score. Single-threaded performance, according to ARM’s Ronco, is up by 20 percent, purely by improving the instructions-per-clock efficiency. The A75 chip is roughly 2.5x the size of the A55, and its intended uses are for infrastructure, automotive, and rich mobile applications. Yes, that means VR, AR, and high-fidelity games, the latter of which ARM’s research has shown have been rapidly increasing in popularity.

A major architectural change with the A75 is the opening of a larger power envelope for chips using this core, scaling up to 2W of power consumption, and thus offering up to 30 percent of extra performance on larger-screen devices. This is entirely targeted at the upcoming Windows on ARM reboot, expected later this year. It’s worth noting that in ARM’s world a "large" screen basically amounts to a laptop — and the company set up a dedicated Large Screen Compute division a year and a half ago to more aggressively target the clamshell devices that Intel has been dominant in. 

As to the new Mali GPU, it has 32 shader cores, 25 percent higher energy efficiency, and a 20 percent better performance density (aka performance per mm² of space). The Mali-G72 is at the heart of ARM’s push toward improving machine learning efficiency, and ARM claims it’s showing itself to be 17 percent better than the G71 in ML benchmarks. The design optimizations from the company are tailored to accelerate inference engines rather than training engines — that’s to say ARM chips will be best at using accumulated ML capabilities rather than developing them, which makes perfect sense for mobile applications. Training AI will be a task better left to Nvidia and AMD graphics cards or Google’s custom TensorFlow TPUs.

The Cortex-A75 and A55 designs were released to ARM’s partners at the end of 2016, so by this point, they’ve all had a few months to decide what to do with them. ARM says a "realistic time window" for new mobile devices powered by its latest designs would be the first quarter of 2018 — though the company is also conscious of a new phenomenon it describes as "China speed," where Chinese phone vendors will put its designs into products almost immediately. The Huawei Mate 9, for example, was released just eight months after ARM distributed the Mali-G71 to partners. This faster Chinese cadence could lead to some A75- and A55-based designs this year, but then bulk of them are likely to arrive with the usual smartphone refresh cycle early next year.



View at the original source