On Sun, Jan 10, 2016 at 10:56 PM, LAU <[email protected]> wrote: > There are few conference available with Jerome Pesenti, Vice President of > Watson Core Technologie, who talks about techniques inside Watson. > > Jerome Pesenti said at a conference (at Paris Tech, a french engineering > school, date unknown ~09/2015) that : > - Watson did not use deep learning in the jeopardy version > - But the system evolves continously, they are replacing many things in > Watson by deep learning. > He said that they are replacing codes in the jeopardy version by deep > learning because it's much more efficient in natural language processing and > others. With deep learning, there will soon be a version of jeopardy for > other languages than English. >
This is from https://developer.ibm.com/watson/wp-content/uploads/sites/19/2013/11/The-Era-of-Cognitive-Systems-An-Inside-Look-at-IBM-Watson-and-How-it-Works1.pdf Many natural language systems have attempted to emphasize precision within the confines of specific well-formed rules. For example, sentiment analysis often looks for a set of specific words and their synonyms within a social media site. These systems then, without further assessment of the context in which those words are being used, tally the number of times those words are co-located with some brand in the same phrase. For example, it takes the phrase, “… stopped by the IBM Donut Store for a coffee this morning, it was great …” and then asserts that the collocation of the brand name and the term “great” are an indication of a positive sentiment. However, consider if the rest of the phrase is, “…, it was great to hear that a new Fictional Coffee Shop is opening soon, so I am not tempted to eat donuts every morning.” Then, the system might miss that the sentiment is not about the IBM Donut Store. We call this concept shallow natural language processing (NLP) because, although it might be fairly precise within its more narrow focus, it is not very accurate. However, it is also important to realize that shallow NLP actually has an important role in many systems. If your intent is to create a statistically relevant assessment of sentiment trends over huge quantities of information, the lack of accuracy for each individual example is likely not an issue. Assuming that there are approximately as many false-positives as there are false-negatives over a sufficiently large sample set, they cancel each other out. And if the pool of canceled tallies remains relatively constant across sample sets over time, the remaining uncanceled data yields statistically relevant trending information. Thus, the additional processing costs that are required for the additional accuracy for any instance might be unwarranted. However, when the individual instances matter, the systems that are designed to be precise without focusing on high levels of accuracy tend to be brittle. That is, they perform well within the narrow parameters of their intended design, but they do not perform well when those parameters change… IBM Watson is a deep NLP system. It achieves accuracy by attempting to assess as much context as possible. It gets that context both within the passage of the question and from the knowledge base (called a corpus) that is available to it for finding responses. Shallow natural language processing can be fairly precise within its more narrow focus, but is not very accurate. We are seeing a shift in construction techniques for natural language processing when accuracy is needed. When preparing for the quiz show, JEOPARDY!, Watson was asked the following question (clue) from the category Lincoln Blogs: “Treasury Secy. Chase just submitted this to me for the third time - guess what pal, this time I'm accepting it.” First, notice the abbreviation, “Secy.”, which had to be taken to mean Secretary. Further notice that Secretary is not meant here to be someone who takes dictation and manages an appointment book. The combined terms Treasury Secretary is significant here as a noun and a role. Therefore, to answer this question, Watson had to find a passage that involved submitting and accepting something between Treasury Secretary Chase and Lincoln (the category of the clue). However, also notice that the category does not say “President Lincoln” necessarily. The correct answer turned out to be “What is a resignation?”. --------------- So this says that Watson is a Deep NLP System. It could mean that it made many deep searches into the corpus of text in order to make better contextual decisions based on fundamental statistics (like word co-occurrence). There may have not been anything like multiple layers of neural nets, but I doubt if the use of neural nets are a requirement for 'Deep Learning'. So then whether Watson-Jeopardy used somethin like Deep Learning (what I called deep learning with a little d and a little l) boils down to the question of whether the NLP rules for deriving the contextual decisions were designed by programmers or whether any of them were derived by machine learning. Presuming that some of the rules they used were derived from computer programs (not explicitly designed by a some guy) then it boils down to the question of whether Watson *itself* used machine learning to derive deep NLP rules (of learning). ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
