Phrases, natural language processing, and artificial intelligence

Michael Arrington’s long interview with Barney Pell and the Microsoft product manager for search after announcing the Powerset acquisition gives hints as to the Powerset’s capabilities that MSFT search needs most vis-a-vis what Google already has incorporated into their search. A lot of it seems to fall into the “phrase” statistics category instead of “linguistic” understanding category — see below but also see this discussion about Powerset and other players like Hakia in semantic search. Given the performance of today’s mainstream search engines from Google/MSFT/Yahoo, I’m beginning to think that until the day we have artificial intelligence, phrases (bigrams, trigrams, etc) and related statistical techniques will remain the silver bullet for general purpose web search.

The Powerset/MSFT folks have a BHAG (big-hairy-audacious-goal) and it’s hard to fault them for that. The technical challenge that Powerset is running up against is the same sort of wall (i.e. steep learning curve) that shows up in modern computer vision and image recognition applied to things like image-search engines or face-recognition in photo collections. The point here is that the technology works in well-defined problem domains or with sufficient training/hints, and it is also massively parallelizable on today’s low-cost server farms for specific tasks — it either fails miserably or has degraded/unreliable performance when the software/code is not programmed to deal with increased scene or corpus complexity, i.e. combinatorial explosion of inputs that arises with diverse features. This happens when the scene or corpus is noisy and when scene/corpus features are ambiguous or new — exactly the situations where human beings still excel in making sense of the world through judgments, and why humans still write software. For example, see these papers from Stanford computer science class projects (1 and 2) that show how unreliable the face-recognition step can be for auto-tagging Facebook photo collections and in-turn how training can be used to achieve “100% Accuracy in Automatic Face Recognition.”

First of Powerset’s capabilities that MSFT wants is word-sense disambiguation — both inside user queries and in the corpus text — to produce more precise search results. Thhe example he gives is “framed.” Two attached papers listed below from the 1990s talk about two different approaches using statistical collocations (Yarowsky) and latent semantic indexing (Schuetze)… this is something MSFT may not be doing yet today, but Google is doing using perhaps Bayesian word clusters and other techniques.

RN: I think everything that Barney said was right on. I think you see search engines including Live Search and also Google and Yahoo are starting to do more work on this matching not exactly what the user entered but it is usually limited to very simple things. So now all of us do some expansion of abbreviations or expansion of acronyms. If you type “NYC” in a search engine these days, in the last couple years, it understands that it means the same thing as New York. These are very very simple rules based things, and no one understands that bark has one meaning if it about a tree and a different meaning if it is about a dog. Or an example that someone gave the other day was the question of “was so and so framed.” And framed could mean a framed picture or it could mean set-up for criminal activity that did not occur, and so on. And you have to actually understand something of it is a person’s name then it applies to one sense of the word framed if it is not then it doesn’t. So one of the things that Powerset brings that is unique is the ability to apply their search technology to the query to the user’s search in ways that are beyond just the simple pluralization or adding an “-ing” is that Powerset also looks at the document, it looks at the words that are on a web page and this is actually very important. If you look at just the users query, what you have available to you to figure what they are talking about are three words four words five words, maybe even less. That can give you certain hints. If you look at a web page that has hundreds or thousands of words on it you have a lot more information you can use if you understand it linguistically to tell what its about, what kind of quieries it should match and what kind of quieries it shouldn’t match. And Powerset is fairly unique in applying this technology in the index on a fairly large scale already and with Microsoft’s investment and long term commitment we can scale this out even further, an apply it even more of the web, not just the wikipedia content they have thus far.

Second is the potential to support complex but realistic queries in the future — they suggest that this can replace vertical search engines which seems a bit far out.

RN: I think what Peter Norwig is saying has some degree of accuracy and that he is also ignoring some things. So, just for normal queries, queries that are not phrased as questions, there is a lot of linguistic structure. If someone types in a query that is “2 bedroom apartments, under 1000 dollars, within a mile of Portero Hill.” That query is loaded with linguistic content. And that’s a realistic query. That is the type of thing that customers actually want to find on the web. Today there is a sort of helplessness, where customers know that certain queries are too complicated, and they wont even issure them to a search engine. They will go to some deep vertical search engine where they can enter different data into different boxes.

The “2 bedroom apartments, under 1000 dollars, within a mile of Portero Hill” example is interesting, since it requires logic and can’t be solved just using related keywords. Even with access to infinite compute cycles I don’t expect any general purpose search engine applied to the web corpus can do this query reliably today, including Powerset. For verticals like real-estate or travel, we will still need things like Craig’s list or Orbitz/Mobissimo that list classified ads and or build on top of user submissions, metadata, and targeted crawling.

Compared to phrases, linguistic analysis helps provide a better answer to the question of whether a given web page matches the user’s intended query or not. It certainly has the potential to improve recall through more accurate query interpretation and retrieval of web pages from the search engine’s index. By itself linguistic analysis does not improve the precision of search results (by re-ranking) in the way that hyper-link analysis did in Google’s PageRank and other web search engines.

The third use case is query expansion to related keywords using an automatically-generated thesaurus — which I thought Google is already doing.

nce; that is not really an area that is that interesting. But some of these more complex queries really are. For example, shrub vs. tree. If I do a search for decorative shrubs for my yard, and the ideal web page has small decorative trees for my garden, it really should have matched that page and brought it up as a good result. But today Google won’t do it, Yahoo won’t do it, and Live won’t do it. So even in these normal queries there is a lot of value in the linguistics.

I didn’t understand this “decorative shrubs” example. When I try it out on Google I find a result that seems relevant to me. Furthermore using the “~” option in front of any word, Google allows you to search using “related” keywords. If you try “decorative ~shrubs” you get results containing “decorative plants” as well.

Fourth, it seems 5% of queries are linguistic queries today (versus phrase queries), and they expect that fraction to expand as MSFT’s search engine capabilities mature to incorporate linguistic understanding of the query — mostly related to facts or solving problems.

BP: Let me answer the question of does anybody actually search this way. The answer is yes, people do this. It isn’t the most common mode, but we do see that probably 5% of queries are natural language queries. These are not all queries that are phrased in complete sentences, but they are queries where the customer has issued something that has some sort of linguistic structure. Almost any query with a preposition: X and Y, A near B, attribute A of Y, etc. Those things are loaded with linguistic structure…. RN: I have a list of some natural language queries in front of me. Can we just show you some queries that our customers have actually sent to us and are random examples. The first person to see the dark side of the moon. How to get a credit card in Malaysia. Enabling system restore in group policy on domain controller. Timeline of Nvidia. How to measure for draperies. What is the difference between Mrs. and women’s sizes? Does my baby have acid reflux? I could just go on and on and I. These fit in the category that we’ve labeled that match about five percent of queries and they’re really just cases where the customer can’t think of a simpler way to express it.

Finally, the people use case seems to be one of Powerset’s most visible strengths…

BP: We return answers. We actually synthesize, so if you were to say, “What did Tom Cruise star in,” you actually get not just the movies, but the cover art for the different movies. It synthesizes multiple pieces of information to give you a whole different kind of presentation. Or, if you were just to say, “Bill Gates” you’d be given an automatically generated profile of Bill Gates, pulled across many, many articles. It’s no longer just about 10 links, although we can certainly do more relevant job (and will) of the blue links, and a better job of presenting those links.

A friend of mine tried getting answers via Google, and got quick (< 5 minutes for all the searches) and relatively accurate answers without much hassle using phrases:

• “far side moon” (see section on Exploration)
• “get credit card Malaysia”
• “enable system restore group policy”
• “Nvidia history”
• “drapery measurements”
• “misses vs womens”
• “baby acid reflux symptoms”

Those are examples that illustrate why phrases and statistical approaches are not only “good enough” for searching the web corpus, and but they have a very high “signal-to-noise” compared to all other machine search techniques that became available to us in the last 20 years.

The root issue seems to me to be that human symbolic and image comprehension is still well beyond today’s programmable machine capabilities — which has nothing to do with compute power or specific interface/modality (language or vision) — it has much more to do with human intelligence itself that resides inside the cerebral cortex and other parts of the central nervous system. I’d be willing to bet that to achieve their BHAG and make the real quantum leaps from today’s search technology that they advertise, Powerset + MSFT will need the equivalent of artificial intelligence, not just natural language processing (or better image processing in the case of scenes).

Powerset is not the only semantic search engine. Besides Hakia, there is Cognition Technologies — a company founded by a UCLA professor — that also has their own wikipedia search http://WIKIPEDIA.cognition.com. This white paper compares their results to Powerset, query by query, showing how Cognition’s precision is higher and most of the time recall is much lower (more relevant vs unrelated hits) — of course this is a test designed by Cognition, and therefore a display of their own strengths. In one example the show Cognition returning exactly one result and Powerset returning 259. Also see Top-Down and Bottom-Up Semantics by Cognition’s CTO. Cognition’s other white paper compares them query-by-query to Google, Yahoo, Verity, and Autonomy — see section VI. In these tests, only Google has an observable “statistical” boost but doesn’t quite work as well as Cognition — more precision and higher relevant recall. In summary, Cognition achieves these results using the semantic processing and natural language parsing technology described here,

Cognition’s Semantic NLP Understands:

• Word stems – the roots of words;
• Words/Phrases – with individual meanings of ambiguous words and phrases listed out;
• The morphological properties of each word/phrase, e.g., what type of plural does it take, what type of past tense, how does it combine with affixes like “re” and “ation”;
• How to disambiguate word senses – This allows Cognition’s technology to pick the correct word meaning of ambiguous words in context;
• The synonym relations between word meanings;
• The ontological relations between word meanings; one can think of this as a hierarchical grouping of meanings or a gigantic “family tree of English” with mothers, daughters, and cousins;
• The syntactic and semantic properties of words. This is particularly useful with verbs, for example. Cognition encodes the types of objects different verb meanings can occur with.

Also see Q-Go, a European NLP technology company for customer service applications that partners with MSFT,

The intelligent, linguistic technology-based software that Q-go develops makes sure that customers receive a direct, immediate answer to all of their questions from a company’s website. Online, in their own language, and with at least the same comprehensiveness and quality as the answers provided by call centres. Not only is this easier for organisations, it’s also faster and cheaper. Q-go’s headquarters are located in Amsterdam, and the company has four local offices in Barcelona, Madrid, Frankfurt and Zurich.

A post by Don Dodge (MSFT) reveals that the Powerset’s “semantic rules” can be applied to MSFT’s existing index of the web and therefore likely to be as much about word word co-occurence statistics + clusters as it is using linguistic logic to gain an understanding from the corpus. The example is also illustrative of how Powerset breaks down a natural language query, and the post goes onto explain how Powerset may also be useful in vertical search applications…

Powerset is using linguistics and (NLP) to better understand the meaning and context of search queries. But the real power of Powerset is applied to the search index, not the query. The index of billions of web pages is indexed in the traditional way. The big difference is in the post processing of the index. They analyze the indexed pages for “semantics”, context, meaning, similar words, and categories. They add all of this contextual meta data to the search index so that search queries can find better results.

Who is the best ballplayer of all time? Powerset breaks this query down very carefully using linguistic ontologies and all sorts of proprietary rules. For example, they know that “ballplayer” can mean Sports. Sports can be separated into categories that involve a “Ball”. Things like baseball, basketball, soccer, and football. Note that soccer does not include the word ball, yet Powerset knows this is a sport that includes a ball.

Powerset knows that “ballplayer” can mean an individual player of a sport that includes a ball. They know that “best of all time” means history, not time in the clock sense.

Finally, an alternate theory that says Powerset was acquired for contextual ad placement, not search.

For the foreseeable future, phrases and statistical approaches will probably continue to deliver the greatest signal-to-noise-ratio for machine indexing of web content in the absence of a breakthrough in artificial intelligence. The evidence is not 100% conclusive, but during the last several weeks I’ve accumulated research papers to support this hypothesis…

• Part-of-speech (POS) tagging for noun phrases is the standard way to find collocations, an approach discussed here based on statistics may work better: this is based on late-1980s work by Church and Hanks “Word Association Norms, Mutual Information, and Lexicography” and “Using statistics in lexical analysis.”
• For a textbook introduction, see Chapter 5, “Collocations” and Chapter 7, “Word Sense Disambiguation” in Foundations of Statistical Natural Language Processing by Christopher D. Manning, Hinrich Schuetze
  
• By constructing a similarity measure between similar words in n-grams (n<=5) using the Wikipedia corpus, the authors of “Finding Similar RSS News Articles Using Correlation Based Phrase Matching” found that they could discover related articles and the highest accuracy was achieved for bigrams (91%) and trigrams (90%) versus unigrams (86%), 4-grams (70%) and 5-grams (60%).
• The paper “More Effective Web Search Using Bigrams and Trigrams” found that phrases (bigrams and trigrams) found by analysis of the top 100 initial search results in typical2-3 word search queries (1) improved the relevance of subsequent searches that included the bigrams/trigrams (2) did even better after verifying the bigrams/trigrams with the user (using relevance feedback) and (3) improved readability of topic areas beyond the first ten results.
• “A Language Model Approach to Keyphrase Extraction” by Tomokiyo compares metrics to rank phrases in terms of their informativeness, and finds derivatives of a KL-metic win over likelihood ratios – this kind of phrase-finding technology was used at Intelliseek in “BlogPulse: Automated Trend Discovery for Weblogs” and “Deriving marketing intelligence from online discussion”.
• The authors of “The Use of Bigrams to Enhance Text Categorization” found that bigrams improved the quality of feature sets used in Naïve Bayes classification of documents into categories.
• For a given word, assignment of the single sense per phrase containing that word that appears most commonly in a document was shown to be highly effective (90-97% accuracy) for word-sense disambiguation in “Unsupervised Word Sense Disambiguation Rivaling Supervised Methods” by David Yarowsky – also see pages 250-253 on “Foundations of Statistical Natural Language Processing” by Manning/Schuetze.
  
• In “Diverse Topic Phrase Extraction through Latent Semantic Analysis,” … “We propose a novel algorithm for extracting diverse topic phrases in order to provide summary for large corpora. Previous works often ignore the importance of diversity and thus extract phrases crowded on some hot topics while failing to cover other less obvious but important topics. We solve this problem through document re-weighting and phrase diversification by using latent semantic analysis (LSA).”
• New word senses can be discovered based on representing co-occurrence and using SVD in “Automatic Word Sense Discrimination” by Hinrich Schutze.
• Phrase-clustering of clickstream content to detect user interests (Slovakia, 2002). This book chapter (Intelligent Support for Information Retrieval in the WWW Environment, 2002) uses clustering to automatically infer/detect a user’s interests from inside the content of web-pages that the user visited – personalization derived from the clickstream content. Reports an 80% success rate in classification using unigrams, bigrams, and trigrams.

Also some papers on techniques/tools for extracting keyphrases and performance evaluations that I came across…

• Google made n-gram data available for all 5-grams (or smaller) appearing more than 40 times on the web, “All Our N-gram are Belong to You”
• Open-Calais by Reuters can analyze for named entities, facts, and events and their API
• The chapter “Natural Language Tools” in pages 149-171 in “Advanced Perl Programming” by Simon Cozens (O’Reilly) — you get a very “quick & dirty” introduction to a number of natural language processing concepts and ways to implement and play around with them. Although Perl has many natural language processing tools, the Cozens book cuts to the chase, explains which are the easiest tools to use, and shows you how to use them.
• “Coherent Keyphrase Extraction via Web Mining” discusses four variants of the KEA key-phrase extraction algorithms – two baseline and two enhanced – and evaluates their performance on a set of computer science and physics research papers. The enhanced versions use statistical association based on web term/phrase frequencies to “weed out” key-phrases that are unlikely to be associated with the text, and they improve on the baseline by generating a higher percentage of key-phrases that match author-generated key-phrases.
• “Narrative Text Classification for Automatic Key Phrase Extraction in Web Document Corpora” compares three key-phrase extraction algorithms (TF.IDF, KEA, and Keyterm) for summarizing web-pages on websites. They find that Keyterm performs best in tests with users. KEA is a keyphrase extraction package, and another one is Lingpipe.
• Paper from MSFT Research for clustering algorithms to find user-verifable “related” phrases within search results — could in theory be applied to any list of documents, “Learning to Cluster Web Search Results”
  
• Autonomy’s Clustering for Enterprise.