Archive for the ‘Semantic Vectors’ Category

From Frequency to Meaning: Vector Space Models of Semantics

Thursday, September 18th, 2014

From Frequency to Meaning: Vector Space Models of Semantics by Peter D. Turney and Patrick Pantel.

Abstract:

Computers understand very little of the meaning of human language. This profoundly limits our ability to give instructions to computers, the ability of computers to explain their actions to us, and the ability of computers to analyse and process text. Vector space models (VSMs) of semantics are beginning to address these limits. This paper surveys the use of VSMs for semantic processing of text. We organize the literature on VSMs according to the structure of the matrix in a VSM. There are currently three broad classes of VSMs, based on term–document, word–context, and pair–pattern matrices, yielding three classes of applications. We survey a broad range of applications in these three categories and we take a detailed look at a specific open source project in each category. Our goal in this survey is to show the breadth of applications of VSMs for semantics, to provide a new perspective on VSMs for those who are already familiar with the area, and to provide pointers into the literature for those who are less familiar with the field.

At forty-eight (48) pages with a thirteen (13) page bibliography, this survey of vector space models (VSMs) of semantics should keep you busy for a while. You will have to fill in VSMs developments since 2010 but mastery of this paper will certain give you the foundation to do so. Impressive work.

I do disagree with the authors when they say:

Computers understand very little of the meaning of human language.

Truth be told, I would say:

Computers have no understanding of the meaning of human language.

What happens with a VSM of semantics is that we as human readers choose a model we think represents semantics we see in a text. Our computers blindly apply that model to text and report the results. We as human readers choose results that we think are closer to the semantics we see in the text, and adjust the model accordingly. Our computers then blindly apply the adjusted model to the text again and so on. At no time does the computer have any “understanding” of the text or of the model that it is applying to the text. Any “understanding” in such a model is from a human reader who adjusted the model based on their perception of the semantics of a text.

I don’t dispute that VSMs have been incredibly useful and like the authors, I think there is much mileage left in their development for text processing. That is not the same thing as imputing “understanding” of human language to devices that in fact have none at all. (full stop)

Enjoy!

I first saw this in a tweet by Christopher Phipps.

PS: You probably recall that VSMs are based on creating a metric space for semantics, which have no preordained metric space. Transitioning from a non-metric space to a metric space isn’t subject to validation, at least in my view.

Recursive Deep Models for Semantic Compositionality…

Tuesday, October 1st, 2013

Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank by Richard Socher, Alex Perelygin, Jean Y. Wu, Jason Chuang, Christopher D. Manning, Andrew Y. Ng and Christopher Potts.

Abstract:

Semantic word spaces have been very useful but cannot express the meaning of longer phrases in a principled way. Further progress towards understanding compositionality in tasks such as sentiment detection requires richer supervised training and evaluation resources and more powerful models of composition. To remedy this, we introduce a Sentiment Treebank. It includes fine grained sentiment labels for 215,154 phrases in the parse trees of 11,855 sentences and presents new challenges for sentiment compositionality. To address them, we introduce the Recursive Neural Tensor Network. When trained on the new treebank, this model outperforms all previous methods on several metrics. It pushes the state of the art in single sentence positive/negative classification from 80% up to 85.4%. The accuracy of predicting fine-grained sentiment labels for all phrases reaches 80.7%, an improvement of 9.7% over bag of features baselines. Lastly, it is the only model that can accurately capture the effect of contrastive conjunctions as well as negation and its scope at various tree levels for both positive and negative phrases.

You will no doubt want to see the webpage with the demo.

Along with possibly the data set and the code.

I was surprised by “fine-grained sentiment labels” meaning:

  1. Positive
  2. Somewhat positive
  3. Neutral
  4. Somewhat negative
  5. Negative

But then for many purposes, subject recognition on that level of granularity may be sufficient.

Semantic Vectors

Tuesday, August 16th, 2011

Semantic Vectors

From the webpage:

Semantic Vector indexes, created by applying a Random Projection algorithm to term-document matrices created using Apache Lucene. The package was created as part of a project by the University of Pittsburgh Office of Technology Management, and is now developed and maintained by contributors from the University of Texas, Queensland University of Technology, the Austrian Research Institute for Artificial Intelligence, Google Inc., and other institutions and individuals.

The package creates a WordSpace model, of the kind developed by Stanford University’s Infomap Project and other researchers during the 1990s and early 2000s. Such models are designed to represent words and documents in terms of underlying concepts, and as such can be used for many semantic (concept-aware) matching tasks such as automatic thesaurus generation, knowledge representation, and concept matching.

The Semantic Vectors package uses a Random Projection algorithm, a form of automatic semantic analysis. Other methods supported by the package include Latent Semantic Analysis (LSA) and Reflective Random Indexing. Unlike many other methods, Random Projection does not rely on the use of computationally intensive matrix decomposition algorithms like Singular Value Decomposition (SVD). This makes Random Projection a much more scalable technique in practice. Our application of Random Projection for Natural Language Processing (NLP) is descended from Pentti Kanerva’s work on Sparse Distributed Memory, which in semantic analysis and text mining, this method has also been called Random Indexing. A growing number of researchers have applied Random Projection to NLP tasks, demonstrating:

  • Semantic performance comparable with other forms of Latent Semantic Analysis.
  • Significant computational performance advantages in creating and maintaining models.

So, after reading about random indexing, etc., you can take those techniques out for a spin. It doesn’t get any better than that!