Is the future of research voice controlled? It might be, because when I originally had the idea for this post my first instinct was to grab my phone and dictate my half-formed ideas into a note, rather than typing it out. Writing things down often makes them seem wrong and not at all what we are trying to say in our heads. (Maybe it’s not so new, since as you may remember Socrates had a similar instinct.) The idea came out of a few different talks at the national Code4Lib conference held in Los Angeles in March of 2017 and a talk given by Chris Bourg. Among these presentations the themes of machine learning, artificial intelligence, natural language processing, voice search, and virtual assistants intersect to give us a vision for what is coming. The future might look like a system that can parse imprecise human language and turn it into an appropriately structured search query in a database or variety of databases, bearing in mind other variables, and return the correct results. Pieces of this exist already, of course, but I suspect over the next few years we will be building or adapting tools to perform these functions. As we do this, we should think about how we can incorporate our values and skills as librarians into these tools along the way.
Natural Language Processing
I will not attempt to summarize natural language processing (NLP) here, except to say that speaking to a computer requires that the computer be able to understand what we are saying. Human—or natural—language is messy, full of nuance and context that requires years for people to master, and even then often leads to misunderstandings that can range from funny to deadly. Using a machine to understand and parse natural language requires complex techniques, but luckily there are a lot of tools that can make the job easier. For more details, you should review the NLP talks by Corey Harper and Nathan Lomeli at Code4Lib. Both these talks showed that there is a great deal of complexity involved in NLP, and that its usefulness is still relatively confined. Nathan Lomeli puts it like this. NLP can “cut strings, count beans, classify things, and correlate everything”. 1 Given a corpus, you can use NLP tools to figure out what certain words might be, how many of those words there are, and how they might connect to each other.
Processing language to understand a textual corpus has a long history but is now relatively easy for anyone to do with the tools out there. The easiest is Voyant Tools, which is a project by Sinclair, Stéfan Sinclair and Geoffrey Rockwell. It is a portal to a variety of tools for NLP. You can feed it a corpus and get back all kind of counts and correlations. For example, Franny Gaede and I used VoyantTools to analyze social justice research websites to develop a social justice term corpus for a research project. While a certain level of human review is required for any such project, it’s possible to see that this technology can replace a lot of human-created language. This is already happening, in fact. A tool called Wordsmith can create convincing articles about finance, sports, and technology, or really any field with a standard set of inputs and outputs in writing. If computers are writing stories, they can also find stories.
Talking to the Voice in the Machine
Finding those stories, and in turn, finding the data with which to tell more stories, is where machine learning and artificial intelligence enter. In libraries we have a lot of words, and while we have various projects that are parsing those words and doing things with them, we have only begun to see where this can go. There are two sides to this. Chris Bourg’s talk at Harvard Library Leadership in a Digital Age, asks the question “What happens to libraries and librarians when machines can read all the books?” One suggestion she makes is that:
we would be wise to start thinking now about machines and algorithms as a new kind of patron — a patron that doesn’t replace human patrons, but has some different needs and might require a different set of skills and a different way of thinking about how our resources could be used. 2
One way in which we can start to address the needs of machines as patrons is by creating searches that work with them, which is for now ultimately to serve the needs of humans, but in the future could be for their own artificial intelligence purposes. Most people are familiar with virtual assistants that have popped up on all platforms over the past few years. As an iOS and a Windows user, I am now constantly invited to speak to Siri or Cortana to search for answers to my questions or fix something in my schedule. While I’m perfectly happy to ask Siri to remind me to bring my laptop to work at 7:45 AM or to wake me up in 20 minutes, I find mixed results when I try to ask a more complex question. 3 Sometimes when I ask the temperature on the surface of Jupiter I get the answer, other times I get today’s weather in a town called Jupiter. This is not too surprising, as asking “What is the temperature of Jupiter?” could mean a number of things. It’s on the human to specify to the computer to which domain of knowledge you are referring, which requires knowing exactly how to ask the question. Computers cannot yet do a reference interview, since they cannot pick up on the subtle hidden meanings or helping with the struggle for the right words that librarians do so well. But they can help with certain types of research tasks quite well, if you know how to ask the question. Eric Frierson (PPT) gave a demonstration of his project working on voice powered search in EBSCO using Alexa. In the presentation he demonstrates the Alexa “skills” he set up for people to ask Alexa for help. They are “do you have”, “the book”, “information about”, “an overview of”, “what I should read after”, or “books like”. There is a demonstration of what this looks like on YouTube. The results are useful when you say the correct thing in the correct order, and for an active user it would be fairly quick to learn what to say, just as we learn how best to type in a search query in various services.
Why ask a question of a computer rather than type in a question to a computer? For the reason I started this piece with, certainly–voice is there, and it’s often easier to say what you mean than write it. This can be taken pragmatically as well. If you find typing difficult, being able to speak makes life easier. When I was home with a newborn baby I really appreciated being able to dictate and ask Siri about the weather forecast and what time the doctor’s appointment was. Herein lies one of the many potential pitfalls of voice: who is listening to what you are saying? One recent news story puts this in perspective, as Amazon agreed to turn over data from Alexa to police in a murder investigation after the suspect gave the ok. They refused to do at first, but it is an open question as to the legal nature of the conversation with a virtual assistant. Nor is it entirely clear when you speak to a device where the data is being processed. So before we all rush out and write voice search tools for all our systems, it is useful to think about where that data lives what the purpose of it is.
If we would protect a user’s search query by ensuring that our catalogs are encrypted (and let’s be honest, we aren’t there yet), how do we do the same for virtual search assistants in the library catalog? For Alexa, that’s built into creating an Alexa skill, since a basic requirement for the web service used is that it meet Amazon’s security requirements. But if this data is subject to subpoena, we would have to think about it in the same way we would any other search data on a third party system. And we also have to recognize that these tools are created by these companies for commercial purposes, and part of that is to gather data about people and sell things to them based on that data. Machine learning could eventually build on that to learn a lot more about people than they think, which the Amazon Echo Look recently brought up as a subject of debate. There are likely to be other services popping up in addition to those offered by Amazon, Google, Apple, and Microsoft. Before long, we might expect our vendors to be offering voice search in their interfaces, and we need to be aware of the transmission of that data and where it is being processed. A recent alliance formed called The Voice Privacy Alliance, which is developing some standards for this.
The invisibility of the result processing has another dark side. The biases inherent in the algorithms become even more hidden, as the first result becomes the “right” one. If Siri tells me the weather in Jupiter, that’s a minor inconvenience, but if Siri tells me that “Black girls” are something hypersexualized, as Safiya Noble has found that Google does, do I (or let’s say, a kid) necessarily know something has gone wrong? 4 Without human intervention and understanding, machines can perpetuate the worst side of humanity.
This comes back to Chris Bourg’s question. What happens to librarians when machines can read all the books, and have a conversation with patrons about those books? Luckily for us, it is unlikely that artificial intelligence will ever be truly self-aware with desires, metacognition, love, and need for growth and adventure. Those qualities will continue to make librarians useful to creating vibrant and unique collections and communities. But we will need to fit that in a world where we are having conversations with our computers about those collections and communities.
- Lomeli, Nathan. “Natural Language Processing: Parsing Through The Hype”. Code4Lib. Los Angeles, CA. March 7, 2017. ↩
- “What Happens to Libraries and Librarians When Machines Can Read All the Books?” Feral Librarian, March 17, 2017. https://chrisbourg.wordpress.com/2017/03/16/what-happens-to-libraries-and-librarians-when-machines-can-read-all-the-books/. ↩
- As a side issue, I don’t have a private office and I feel weird speaking to my computer when there are people around. ↩
- Noble, Safiya Umoja. “Google Search: Hyper-Visibility as a Means of Rendering Black Women and Girls Invisible – InVisible Culture.” InVisible Culture: An Electronic Journal for Visual Culture, no. 19 (2013). http://ivc.lib.rochester.edu/google-search-hyper-visibility-as-a-means-of-rendering-black-women-and-girls-invisible/. ↩
A decade ago, Stephen Colbert introduced the concept of “truthiness”, or a fact that was so because it felt right “from the gut.” When we search for information online, we are always up against the risk that the creator of a page is someone who, like Stephen Colbert’s character doesn’t trust books, because “they’re all fact, no heart.”1 Since sites with questionable or outright false facts that “feel right” often end up at the top of Google search results, librarians teach students how to evaluate online sources for accuracy, relevancy, and so on rather than just trusting the top result. But what if there were a way to ensure that truthiness was removed, and only sites with true information appeared at the top of the results?
This idea is what underlies a new Google algorithm called Knowledge-Based Trust(KBT)2. Google’s original founding principles and the PageRank algorithm were based on academic citation practices–loosely summarized, pages linked to by a number of other pages are more likely to be useful than those with fewer links. The content of the page, while it needs to match the search query, is less crucial to its ranking than outside factors, which is otherwise known as an exogenous model. The KBT, by contrast, is an endogenous model relying on the actual content of the page. Ranking is based on the probability that the page is accurate, and therefore more trustworthy. This is designed to address the problem of sites with high PageRank scores that aren’t accurate, either because their truthiness quotient is high, or because they have gamed the system by scraping content and applying misleading SEO. On the other side, pages with great information that aren’t very popular may be buried.
“Wait a second,” you are now asking yourself, “Google now determines what is true?” The answer is: sort of, but of course it’s not as simple as that. Let’s look at the paper in detail, and then come back to the philosophical questions.
Digging Into the KBT
First, this paper is technical, but the basic information is fairly straightforward. This model is based on extracting facts from a web source, evaluating whether those facts are true or not, and then whether a source is accurate or not. This leads to a determination that the facts are correct in an iterative process. Of course, verifying that determination is essential to ensuring that all the algorithms are working correctly, and this paper describes ways of checking the extracted facts for accuracy.
The extractors are described more fully in an earlier version of this work, Knowledge Vault (KV), which was designed to fill in large-scale knowledge bases such as Freebase by extracting facts from a web source using techniques like Natural Language Processing of text files followed by machine learning, HTML DOM trees, HTML tables, and human processed pages with schema.org metadata. The extractors themselves can perform poorly in creating these triples, however, and this is more common than the facts being wrong, and so sites may be unfairly flagged as inaccurate. The KBT project aims to introduce an algorithm for determining what type of error is present, as well as how to judge sites with many or few facts accurately, and lastly to test their assumptions using real world data against known facts.
The specific example given in the paper is the birthplace of President Barack Obama. The extractor would determine a predicate, subject, object triple from a web source and match these strings to Freebase (for example). This can lead to a number of errors–there is a huge problem in computationally determining the truth even when the semantics are straightforward (which we all know it rarely is). For this example, it’s possible to check data from the web against the known value in Freebase, and so if that extractor works set an option to 1 (for yes) and 0 (for no). Then this can be charted in a two-dimensional or three-dimensional matrix that helps show the probability of a given extractor working, as well as whether the value pulled by the extractor was true or not.
They go on to examine two models for computing the data, single-layer and multi-layer. The single-layer model, which looks at each web source and its facts separately, is easier to work with using standard techniques but is limited because it doesn’t take into account extraction errors. The multi-layer model is more complex to analyze, but takes the extraction errors into account along with the truth errors. I am not qualified to comment on the algorithm math in detail, but essentially it computes probability of accuracy for each variable in turn, ultimately arriving at an equation that estimates how accurate a source is, weighted by the likelihood that source contains those facts. There are additional considerations for precision and recall, as well as confidence levels returned by extractors.
Lastly, they consider how to split up large sources to avoid computational bottlenecks, as well as to merge sources with few facts in order to not penalize them but not accidentally combine unrelated sources. Their experimental results determined that generally PageRank and KBT are orthogonal, but with a few outliers. In some cases, the site has a low PageRank but a high KBT. They manually verified the top three predicates with high extraction accuracy scores for web sources with a high KBT to check what was happening. 85% of these sources were trustworthy without extraction errors and with predicates related to the topic of the page, but only 23% of these sources had PageRank scores over 0.5. In other cases, sources had a low KBT but high PageRank, which included sites such as celebrity gossip sites and forums such as Yahoo Answers. Yes, indeed, Google computer scientists finally have definitive proof that Yahoo Answers tends to be inaccurate.
The conclusion of the article with future improvements reads like the learning outcomes for any basic information literacy workshop. First, the algorithm would need to be able to tell the main topic of the website and filter out unrelated facts, to understand which triples are trivial, to have better comprehension of what is a fact, and to correctly remove sites with data scraped from other sources. That said, for what it does, this is a much more sophisticated model than anything else out there, and at least proves that there is a possibility to computationally determine the accuracy of a web source.
What is Truth, Anyway?
Despite the promise of this model there are clearly many potential problems, of which I’ll mention just a few. The source for this exercise, Freebase, is currently in read-only mode as its data migrates to Wikidata. Google is apparently dropping Freebase to focus on their Open Knowledge Graph, which is partially Freebase/Wikidata content and partially schema.org data 3. One interesting wrinkle is that much of Freebase content cites Wikipedia as a source, which means there are currently recursive citations that must be properly cited before they will be accepted as facts. We already know that Wikipedia suffers from a lack of diversity in contributors and topic coverage, so a focus on content from Wikipedia has the danger of reducing the sources of information from which the KBT could check triples.
That said, most of human knowledge and understanding is difficult to fit into triples. While surely no one would search Google for “What is love?” or similar and expect to get a factual answer, there are plenty of less extreme examples that are unclear. For instance, how does this account for controversial topics? I.e. “anthropogenic global warming is real” vs. “global warming is real, but it’s not anthropogenic.” 97% of scientists agree to the former, but what if you are looking for what the 3% are saying?
And we might question whether it’s a good idea to trust an algorithm’s definition of what is true. As Bess Sadler and Chris Bourg remind us, algorithms are not neutral, and may ignore large parts of human experience, particularly from groups underrepresented in computer science and technology. Librarians should have a role in reducing that ignorance by supporting “inclusion, plurality, participation and transparency.” 4 Given the limitations of what is available to the KBT it seems unlikely that this algorithm would markedly reduce this inequity, though I could see how it could be possible if Wikidata could be seeded with more information about diverse groups.
Librarians take note, this algorithm is still under development, and most likely won’t be appearing in our Google results any time in the near future. And even once it does, we need to ensure that we are still paying attention to nuance, edge cases, and our own sense of truthiness–and more importantly, truth–as we evaluate web sources.
- http://thecolbertreport.cc.com/videos/63ite2/the-word—truthiness. ↩
- Dong, X. et al. “Knowledge-Based Trust: Estimating the Trustworthiness of Web Sources”. Proceedings of the VLDB Endowment, 2015. Retrieved from http://arxiv.org/abs/1502.03519 ↩
- https://www.wikidata.org/wiki/Help:FAQ/Freebase ↩
- Sadler, Bess and Chris Bourg, “Feminism and the Future of Library Discovery.” Code4Lib Journal 28, April 2015. ↩