V. Ubicación organizacional del proceso en la IES
2. Consideraciones preliminares
3.1. Eficacia, Relevancia Y Pertinencia De La Formación Recibida
LDA is a significant topic model on which many researchers based their work to capture other properties of the text. To do so, they added variables to their models to describe the development of topics over time, the relationship among topics, the role of syntax in topic identification and so on. In the following, we briefly present some of recently introduced topic models where the majority of them are based on the fundamentals of latent Dirichlet allocation.
topics over time in a sequentially organized corpus of documents. This approach infers about the latent parameters of the model using the variational method. The parameters of the model follow the multinomial distribution. A state space repre- sentation is used to transmit the multinomial parameters upon the words of each topic. The Correlated Topic Model (CTM) [5] was designed to provide correlation among topics. The key idea thatCTM relies on is that a document discussing about medicine is more likely to be related to disease than astronomy. The assumption of LDAthat topics are drawn from a Dirichlet distribution confinesLDAto provide the correlation between topics. To facilitate topic correlations, topic models assume that topics have correlations via the logistic normal distribution that exhibit a sufficient satisfactory fit on test data.
In 2004 Blei et al. introduced an extension ofLDA-named hierarchical latent Dirich- let allocation [25] - that deals with topics in the manner of hierarchies. On that front, they combine a nested Chinese Restaurant Process (CRP) with a likelihood that relies on a hierarchical variant of latent Dirichlet allocation to derive a prior distribution on hierarchies. In 2010, the supervised topic model [7] was introduced to deal effectively with prediction problems. They designed a topic model to perform prediction regarding the vocabulary. They examine the prediction power of words with respect to the topic class. They compareLDAwith supervised topic model and they find the new model to more effective.
In traditional topic models, such as LDA, most of the syntactic words are removed since we are only interested in meaning and only long-range dependencies are con- cerned. Therefore, topic models focus on identifying semantic words through doc- uments or entire collections. On the contrary, the composite model [26] that was introduced by Griffiths et al. considers the short-range dependencies as well. It blends a Hidden Markov Model (HMM) to capture the parts of speech and a latent Dirichlet allocation to extract words that are deemed semantic. Composite model competes for part-of-speech taggers and it is not used for topic classification itself. In Figure 2.5 it is demonstrated the generating phase of this model where an au-
tomaton is constructed to describe the structure of the language. Figure 2.5 shows the transitions of a three classHMMannotated with the corresponding probabilities. The semantic class shown in the middle consists of three topic sets each one assigned a probability. The other two classes are simple multinomial distribution over words. Document phrases are generated by following the transitions of an automaton like the one in 2.5. Particularly, a word is chosen from the distribution associated with each syntactic class, a topic follows and a word comes next from a distribution asso- ciated with that topic for the semantic class.
Figure 2.5: The generating phase of composite model [26]
Although exchangeable word models are useful for classification or information re- trieval, they are limited for problems that depend on more fine-grained qualities of language. For instance, a topic model is efficient on providing documents relevant to queries but it cannot suggest relevant phrases for question answering. Syntactic Topic Model (STM) [12] is a document model that blends the observed syntactic structure with the latent thematic structure of a document. STMintends to extract groups of words that are utilized the same way in similar documents. STM can be used to incorporate document context into parsing models but is not a full parsing model. It provides a way to learn both simultaneously rather than combining the two heterogeneous methods. Syntactic topic models have been used for statistical natural language generation [17].
In 2009, C. Wang et al. introduce a generative probabilistic model [57] to capture firstly the corpus-wide topic structure and secondly the topic correlation across cor- pora. They test their model on a dataset extracted from six different computer science conferences; they evaluate their model on the abstracts parts of the text. Additionally, researchers have studied the efficiency of topic models on different lev- els of resolution. Bruber et al. [27] consider that each sentence discusses one topic and all the words in a sentence are assigned the sentences topic. The goal of the authors is to perform word sense disambiguation. Wallach [56] extendedLDA to fa- cilitate n-gram statistics by designing a hierarchical Dirichlet bigram language model. They produce more meaningful topics thanLDA since bigrams statistics restrict the dominant role of stopwords.