Latent Semantic Analysis

Latent Semantic Analysis (LSA) is an NLP method:

  • mathematical/statistical method for modeling the meaning of words/passages by analysis of text via extracting and inferring relations of expected contextual usage of words in texts
  • idea: words that are used in the same contexts tend to have the same meaning


Problems with Text

Issues with text data:

  • synonymy: many ways to refer to the same object
  • synonymy tends to decrease recall
  • polysemy: many words have more than one distinct meaning (e.g. "chip", "trunk")
  • polysemy tends to decrease precision


Overcoming Synonymy:

  • term extraction, thesauri construction

Overcoming Polysemy:


WEIRD

WEIRD (Koll1979) is the first IR system that dealt with these problems automatically, not with some controlled vocabulary

  • the goal of WIERD: to go from term matching to concept matching
  • can use statistical analysis to empirically find relations among terms
  • so it analyzed term-to-term co-occurrence matrix
  • can use Factor Analysis to identify the right basis for terms s.t. there's little or no loss of information
  • in WEIRD only 7 dimensions were used - based on 7 completely non-overlapping documents found in the collection


The space built by WEIRD acts like an implicit thesaurus

  • synonyms will map to the same concept


LSA

LSA/LSI solves these problems as well

  • it goes further than WIERD: it uses all documents to build a space
  • it does that by applying SVD as a Dimensionality Reduction - which reveals latent structure and "denoises" the data
  • Similarity estimates derived by LSA are not just frequencies or co-occurrences counts: it can infer deeper relations: hence "Latent" and "Semantic"
  • so LSA learns the latent semantic structure of the vocabulary


LSI

Latent Semantic Analysis (LSA) $\approx$ Latent Semantic Indexing (LSI)

  • LSI is the alias of LSA for Information Retrieval
  • indexing and retrieval method that uses SVD to identify patterns in relations between terms and concepts
  • instead of literal match between query and documents (e.g. using cosine in the traditional vector space morels), convert both into the Semantic Space and calculate the cosine there




LSA Steps

3 major steps (by Evangelopoulos2012)


Document preparation


Representation: Vector Space Models

Construct a matrix $D$

  • $D$ is Term-Document Matrix if rows of $D$ - terms, columns of $D$ - documents/passages
  • $D$ is Document-Term Matrix if rows of $D$ - documents/passages, and columns of $D$ - terms
  • each cell - typically a frequency with which a word occurs in a doc
  • also apply weighting: TF or TF-IDF


SVD and Dimensionality Reduction

Let $D$ be an $t \times p$ Term-Passage matrix

  • $t$ rows are terms, $p$ columns are passages, $\text{rank } D = r$
  • then SVD decomposition is $D = T \cdot \Sigma \cdot P^T$
  • $T$ is $t \times r$ Orthogonal Matrix, contains left singular vectors, corresponds to term vectors
  • $\Sigma$ is $r \times r$ a diagonal matrix of singular values
  • $P$ is $r \times p$ Orthogonal Matrix, contains right singular vectors, corresponds to passage vectors
  • and then $T \sqrt\Sigma$ are loadings for terms and $P \sqrt\Sigma$ - for passages


Now reduce the dimensionality:

  • want to combine the surface text information into some deeper abstraction
  • finding the optimal dimensionality for final representation in the Semantic Space is important to properly capture mutual usage of words
  • the "True Semantic Space" should address the Text Problems


So, Apply reduced-rank SVD

  • $D \approx T_k \cdot \Sigma_k \cdot P^T_k$
  • keep only $k$ largest singular values
  • the result: best $k$-dim approximation of the original matrix $D$
  • for NLP $k = 300 \pm 50$ usually works the best
  • but it should be tuned because it heavily depends on the domain


Semantic Space

LSA constructs a semantic space via SVD:

  • $T$ is $t \times r$ Orthogonal Matrix, contains left singular vectors, corresponds to term vectors
  • $\Sigma$ is $r \times r$ a diagonal matrix of singular values
  • $P$ is $r \times p$ Orthogonal Matrix, contains right singular vectors, corresponds to passage vectors
  • and then $T \sqrt\Sigma$ are loadings for terms and $P \sqrt\Sigma$ - for passages


Language-theoretic interpretation:

  • LSA vectors approximate:
  • the meaning of a word as its average effect of the meaning of passages in which they occur
  • the meaning of a passage as meaning of its words


After doing the SVD, we get the reduced space - this is the semantic space

  • the effect of reducing the dimensionality:
  • removed the noise effect of synomymy and polysemy


Comparisons in the Semantic Space

So we approximated $D$ as $D \approx \hat D = T_k \Sigma_k P_k^T$

  • lets omit index $k$: so below by $T$ we will assume $T_k$


Term comparisons:

  • How similar are terms $\mathbf t_i$ and $\mathbf t_j$?
  • In $D$ we would compare rows of $D$. How to compare them in the semantic space?
  • $\hat D \hat D^T$ gives a term-term Gram Matrix
    • $\hat D \hat D^T = T \Sigma \Sigma^T T^T = T \Sigma \, (T \Sigma)^T$
    • thus $\big[\hat D \hat D^T\big]_{ij}$ is the dot product between $i$th and $j$th rows of $T \Sigma$
  • rows of $T \Sigma$ are coordinates for terms in the semantic space


Document comparisons:

  • how similar are documents $\mathbf p_i$ and $\mathbf p_j$ in the semantic space?
  • $\hat D^T \hat D$ gives a document-document Gram Matrix
  • $\hat D^T \hat D = P \Sigma \Sigma^T P^T = P \Sigma \, (P \Sigma)^T$
  • so to compute document $i$ and $j$ you compute the dot product between $i$th and $j$th rows of $P \Sigma$


Generalization to Unseen Documents

What about objects that didn't originally appear in the training set?

  • e.g. a query $\mathbf q$
  • how do we represent $\mathbf q$ in the semantic space?
  • first, let's see how original documents $\mathbf p_i$ are represented in this space


$\hat D = T \Sigma P^T$

  • multiply by $(T \Sigma)^{-1}$ on the left
  • $(T \Sigma)^{-1} \hat D = P^T$
  • $\Sigma^{-1} T^T \hat D = P^T$
  • $P = D^T T \Sigma^{-1}$
  • if $\mathbf d_i$ be some document in the original space (column of $\hat D$) and $\mathbf p_i$ the corresponding representation of $\mathbf d_i$ in the document basis, then
  • $\mathbf p_i = \mathbf d_i^T T \Sigma^{-1}$


This, can represent $\mathbf q$ the same way:

  • $\hat{\mathbf q} = \mathbf q^T T \Sigma^{-1}$
  • where $\hat{\mathbf q}$ is the representation of $\mathbf q$ in the document basis
  • to compare $\hat{\mathbf q}$ all we need to do is to scale it by $\Sigma$ and then compute a dot product


Example

Article Titles Example

Let's consider titles of some articles (from Deerwester90):

  • $c_1$: "Humanmachine interface for ABC computer applications"
  • $c_2$: "A survey of user opinion of computer system response time"
  • $c_3$: "The EPS user interface management system"
  • $c_4$: "Systemand human system engineering testing of EPS"
  • $c_5$: "Relation of user perceived response time to error measurement"
  • $m_1$: "The generation of random, binary, ordered trees"
  • $m_2$: "The intersection graph of paths in trees"
  • $m_3$: "Graph minors IV: Widths of trees and well-quasi-ordering"
  • $m_4$: "Graph minors: A survey"

Matrix:

$D = \left[\begin{array}{c|cccccccc} & c_1 & c_2 & c_3 & c_4 & c_5 & m_1 & m_2 & m_3 & m_4 \\ \hline \text{human} & 1 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 \\ \text{interface} & 1 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 \\ \text{computer} & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\ \text{user} & 0 & 1 & 1 & 0 & 1 & 0 & 0 & 0 & 0 \\ \text{system} & 0 & 1 & 1 & 2 & 0 & 0 & 0 & 0 & 0 \\ \text{response} & 0 & 1 & 0 & 0 & 1 & 0 & 0 & 0 & 0 \\ \text{time} & 0 & 1 & 0 & 0 & 1 & 0 & 0 & 0 & 0 \\ \text{EPS} & 0 & 0 & 1 & 1 & 0 & 0 & 0 & 0 & 0 \\ \text{survey} & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 1 \\ \text{trees} & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 1 & 0 \\ \text{graph} & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 1 \\ \text{minors} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 \\ \end{array}\right]$


Note:

  • row vectors for "human" and "user" are orthogonal: their dot product is zero, but they are supposed to be similar, so it must be positive
  • also, "human" and "minors" are orthogonal, but they are not similar, so it must be negative

Let's apply SVD:

  • $D = W \Sigma P$
  • 2-dim approximation: $D_2 = W_2 \Sigma_2 P_2$

$D_2 = \left[\begin{array}{c|cccccccc} & c_1 & c_2 & c_3 & c_4 & c_5 & m_1 & m_2 & m_3 & m_4 \\ \hline \text{human} & 0.16 & 0.4 & 0.38 & 0.47 & 0.18 & -0.05 & -0.12 & -0.16 & -0.09 \\ \text{interface} & 0.14 & 0.37 & 0.33 & 0.4 & 0.16 & -0.03 & -0.07 & -0.1 & -0.04 \\ \text{computer} & 0.15 & 0.51 & 0.36 & 0.41 & 0.24 & 0.02 & 0.06 & 0.09 & 0.12 \\ \text{user} & 0.26 & 0.84 & 0.61 & 0.7 & 0.39 & 0.03 & 0.08 & 0.12 & 0.19 \\ \text{system} & 0.45 & 1.23 & 1.05 & 1.27 & 0.56 & -0.07 & -0.15 & -0.21 & -0.05 \\ \text{response} & 0.16 & 0.58 & 0.38 & 0.42 & 0.28 & 0.06 & 0.13 & 0.19 & 0.22 \\ \text{time} & 0.16 & 0.58 & 0.38 & 0.42 & 0.28 & 0.06 & 0.13 & 0.19 & 0.22 \\ \text{EPS} & 0.22 & 0.55 & 0.51 & 0.63 & 0.24 & -0.07 & -0.14 & -0.2 & -0.11 \\ \text{survey} & 0.1 & 0.53 & 0.23 & 0.21 & 0.27 & 0.14 & 0.31 & 0.44 & 0.42 \\ \text{trees} &-0.06 & 0.23 & -0.14 & -0.27 & 0.14 & 0.24 & 0.55 & 0.77 & 0.66 \\ \text{graph} &-0.06 & 0.34 & -0.15 & -0.3 & 0.2 & 0.31 & 0.69 & 0.98 & 0.85 \\ \text{minors} &-0.04 & 0.25 & -0.1 & -0.21 & 0.15 & 0.22 & 0.5 & 0.71 & 0.62 \\ \end{array}\right]$


What's the effect of dimensionality reduction here?

  • words appear less or more frequent than originally
  • consider two cells: ("survey", $m_4$) and ("trees", $m_4$)
  • original document: 1 and 0
  • reduced document: 0.42 and 0.66
  • because $m_4$ contains "graph" and "minor", the 0 for "trees" was replaced by 0.42 - they are related terms
  • so it can be seen as estimate of how many times word "trees" would occur in other samples that contain "graph" and "minor"
  • the count for "survey" went down - it's not expected in this context

So in the reconstructed space:

  • dot product between "user" and "human" is positive
  • dot product between "human" and "minors" is negative
  • it tells us way better whether terms are similar or not even when they never co-occur together


Taking 2 principal components is the same as taking only 2 abstract concepts

  • each word in the vocabulary has some amount of these 2 concepts (we see how much by looking at 1st and 2nd column of $W$)


The idea:

  • we don't want to reconstruct the underlying data perfectly, but instead we hope to find the correlation and the abstract concepts


Python code

import numpy as np
import numpy.linalg as la
 
D = [[1, 0, 0, 1, 0, 0, 0, 0, 0],
     [1, 0, 1, 0, 0, 0, 0, 0, 0],
     [1, 1, 0, 0, 0, 0, 0, 0, 0],
     [0, 1, 1, 0, 1, 0, 0, 0, 0],
     [0, 1, 1, 2, 0, 0, 0, 0, 0],
     [0, 1, 0, 0, 1, 0, 0, 0, 0],
     [0, 1, 0, 0, 1, 0, 0, 0, 0],
     [0, 0, 1, 1, 0, 0, 0, 0, 0],
     [0, 1, 0, 0, 0, 0, 0, 0, 1],
     [0, 0, 0, 0, 0, 1, 1, 1, 0],
     [0, 0, 0, 0, 0, 0, 1, 1, 1],
     [0, 0, 0, 0, 0, 0, 0, 1, 1]]
D = np.array(D)
 
rows = ['human', 'interface', 'computer', 'user', 'system', 
        'response', 'time', 'EPS', 'survey', 'trees', 'graph', 'minors']
idx = {n: i for (i, n) in enumerate(rows)}
 
D[idx['human']].dot(D[idx['user']]) # 0
D[idx['human']].dot(D[idx['minors']]) # 0
 
 
T, S, P = la.svd(D) # T=terms, P=passages
 
np.set_printoptions(precision=2, suppress=True)
print T[:, 0:2], S[0:2], P[0:2, :]
 
D_hat = T[:, 0:2].dot(np.diag(S[0:2])).dot(P[0:2, :])
 
D_hat[idx['human']].dot(D_hat[idx['user']]) # 0.955
D_hat[idx['human']].dot(D_hat[idx['minors']]) # -0.251


Can do the same without building $\hat D$:


T = T[:, 0:2]
S = np.diag(S[0:2])
P = P[0:2, :].T

human = T.dot(S)[idx['human']]
user = T.dot(S)[idx['user']]
human.dot(user) # same result: 0.955


Finally, let's calculate cosine between human and user:

human.dot(user) / (la.norm(human) * la.norm(user))
# 0.88784582874340123



Practical Notes

Applications


Limitations

  • makes no use of words order, punctuation
  • if the original terms are already descriptive enough (e.g. for Document Classification), they may be lost during the transformation


When Not Good

  • Sometimes Semantic Spaces alone are not good
  • but we can mix the original vector space and the semantic space together


Mean Centering

LSA and Principal Component Analysis are related via SVD



Extensions of LSA


Links

Sources

  • Koll, Matthew B. "WEIRD: An approach to concept-based information retrieval." 1979.
  • Landauer, Thomas K., Peter W. Foltz, and Darrell Laham. "An introduction to latent semantic analysis." 1998. [1]
  • http://www.scholarpedia.org/article/Latent_semantic_analysis
  • http://edutechwiki.unige.ch/en/Latent_semantic_analysis_and_indexing
  • Evangelopoulos, Nicholas, Xiaoni Zhang, and Victor R. Prybutok. "Latent semantic analysis: five methodological recommendations." (2012). [2] [3]
  • Deerwester, Scott C., et al. "Indexing by latent semantic analysis." 1990. [4]
  • Berry, Michael W., Susan T. Dumais, and Gavin W. O'Brien. "Using linear algebra for intelligent information retrieval." (1995). [5]
  • Korenius, Tuomo, Jorma Laurikkala, and Martti Juhola. "On principal component analysis, cosine and Euclidean measures in information retrieval." 2007. [6]
  • Zhukov, Leonid, and David Gleich. "Topic identification in soft clustering using PCA and ICA". 2004. [7]