CS671: Natural Language Processing

HW3: Paper Review

Vishal Kumar Gupta 10817

@article{silberermodels,
title={Models of Semantic Representation with Visual Attributes},
author={Silberer, Carina and Ferrari, Vittorio and Lapata, Mirella}
year={2013},

annote = { Grounded Language Acquitision is defined as extracting representations for the meaning of natural language tied to physical world. The basic motivation is exploting the visual modality of langauge to learn semantics. Various approaches employed include semantics parsing, cross-mapping of NL and actions, association of language and perpeutal data and learning based on linguistic and perpeutal input. The basic goal is construction of distribution models which are perpetually grounded. Two prime motivation behind the goal are: (1) There exists a direct relation between language interpretation and physical world interactions(Cognitive aspect) and (2) The model can very well learn representations for multimodal data and its applications are diverse(Engineering Aspect).

Visual modalities can be considered as intrinsic representation of language. This can also include auditory, motor and haptic modalities. The main reason for using visual modalities include (1) easy availability of data and 2) images in the database are hierarchial according to wordnet synsets. Previously lower layer features(SIFT) were used. This paper presents the use of high leval features in the form of visual attributes were used which offers many advantages. This include: (1) Using attributes offers generalization over unseen objects and (2) attribute-centric approach for representing images helps build a distributioanl model for lexical meaning also.

For the purpose of dataset images for the concepts in McRae et al.’s (2005)production norms were used. This data set possesed a nominal hierarchial of WordNet. Data was divided into training, development and testing sets. Development set is used to develop and refine attribute annotation scheme. Attribute annotation is conducted on per-concept basis rather than per-image basis. For each of the attribute a L2 regularized SVM(Fan et al., 2008) classifier is learned. Since a single concept is visible in many images and we were entitled to learn a concept w based on its visual attributes a single representation comprising of those attributes was needed. For the purpose the following approach is followed: for each input image a vector is constructed which has prediction scores for the attributes of that image as entries for a given concept. Based on these vectors a normalized probability distribution of attribute w is estimated.

For evaluating the result coorelation coefficients were calculated in order to observe inter-model relation. In most of the cases two modalities are found to perform better than one. The model also performed quite well on unseen data showing generalization as expected. In totalty, the results are summarised as follow:
=> Distributional models perform better with visual attributes.
=> Attribute-centered representation certainly have advantages over conventional approaches based raw features.
}
}

Reference(the paper itself)
link to bibtex file